When Moderation Fosters Persuasion: The Persuasive Power of Deviatory Reviews

When Moderation Fosters Persuasion: The Persuasive Power of Deviatory Reviews Abstract When people seek to persuade others to purchase a particular product or service, they often give an extremely favorable review of it as a means of doing so. Despite the intuitive appeal of this strategy, the current research demonstrates that a moderately positive review is sometimes more persuasive. In particular, when the perceived default evaluation in a given context is extremely positive, moderately positive reviews that deviate from that default can become more persuasive. In contrast, when the perceived default is moderately positive, extremely positive reviews tend to be more persuasive. This deviation effect occurs because reviews that deviate from the perceived default are believed to be more thoughtful, and thus accurate, which enhances their persuasive impact. This effect is demonstrated in eight experiments set in a diverse range of consumer contexts. persuasion, perceived thoughtfulness, defaults, norms, customer reviews Which reviewer would lead consumers to be more likely to try a new restaurant: one who notes that she tried the restaurant, thought it was extremely good, and rates it 5 out of 5 stars; or one who notes that she tried the restaurant, thought it was good, and rates it 4 out of 5 stars? It seems reasonable to predict that the former review would be more persuasive—that is, it would lead consumers to be more likely to try the restaurant. After all, the former review seems to suggest that the restaurant is better than does the latter review and, understandably, consumers tend to prefer really good options to pretty good options (Kuo, Wu, and Deng 2009). Consistent with this general intuition, when people aim to persuade others to purchase a product or service, they often emphasize their extremely favorable endorsement of it as a means of doing so. Consumers who want to convince others to purchase a product they enjoy, for instance, frequently give that product the highest possible rating online. Indeed, ecommerce websites such as Amazon.com are filled with examples of customers urging their fellow shoppers to purchase specific products by rating those products with 5 stars and giving them extremely positive reviews (for just a few examples, see Amazon 2014, 2015, 2016). Similarly, companies seeking to increase sales routinely inform people that their customers view them extremely favorably—for example, that they give them the highest possible ratings online (Groupon 2016; Printing Center USA 2018; Yelp 2015). The notion that extremely positive reviews will be more persuasive—for example, more likely to prompt customer patronage—than moderately positive reviews has considerable face validity. And indeed, we predict that this is often the case. Importantly, however, we theorize that the reverse can also regularly occur. That is, we predict that under specifiable conditions a moderately positive review of a product, service, or company can be more persuasive than an extremely positive review. In particular, we posit that when the default review in a given context is perceived to be extremely positive, moderately positive reviews that deviate from that default might become more persuasive. In essence, we theorize that consumers perceive deviatory reviews—which we define as reviews that deviate from a default—to be more accurate. As a result, such reviews can be more effective in convincing others. If true, this logic implies that a moderately positive review should be more persuasive than an extremely positive review when extremely positive reviews are the default. Conversely, an extremely positive review should be more persuasive than a moderately positive review when moderately positive reviews are the default. We unpack these predictions below. THEORETICAL BACKGROUND Defaults are options that consumers perceive to be the status quo and, thus, that they consider first before considering other options (Huh, Vosgerau, and Morewedge 2014). Defaults are determined by factors both endogenous and exogenous to a decision context. For instance, preselected options are endogenous factors that create defaults (Everett et al. 2015; Johnson and Goldstein 2003; Probst, Shaffer, and Chan 2013). Descriptive norms (which describe how most people generally behave; Deutsch and Gerard 1955; Goldstein, Cialdini, and Griskevicius 2008; Miller 1999) are exogenous factors that create defaults (Miller and Prentice 1996). Importantly, defaults of both endogenous and exogenous origin can powerfully drive behavior. For example, people are more likely to reuse hotel towels when they learn that most others reuse theirs (Goldstein et al. 2008), and people are more likely to contribute to a 401(k) plan when they learn that most others contribute to a 401(k) plan (Bailey, Nofsinger, and O’Neill 2004). Similarly, medical providers are more likely to order laboratory tests for their patients when those tests are preselected (Probst et al. 2013), Mechanical Turk workers are more likely to donate bonus money to charity (vs. keep it for themselves) when the option to donate is preselected (Everett et al. 2015), and people are more likely to become organ donors when organ donation is set as a default on registration forms (Johnson and Goldstein 2003). People are particularly likely to accept, or adhere to, a default when their motivation or ability to think deeply is limited in some way. Put differently, people are especially unlikely to deviate from a default if they have not thoroughly deliberated about their choice. For example, consumers under cognitive load or time pressure—who do not have the cognitive resources or time to deliberate—are less likely to deviate from a default than are consumers who are cognitively unencumbered (Huh et al. 2014). Similarly, consumers who are overloaded with options, and who self-report that the large number of options is hindering their ability to thoroughly deliberate about their choice, are less likely to deviate from a default (Agnew and Szykman 2005). We propose that just as people are more likely to deviate from a default when they have carefully deliberated about their choice, they might believe that others are more likely to deviate from a default when those others have carefully deliberated about their choice. Consistent with this general possibility, past research reveals that people tend to believe that others think, feel, and act as they do (Clement and Krueger 2002; Krueger 2000; Ross, Greene, and House 1977). Simulation processes appear to contribute to these effects, such that people predict and interpret others’ thoughts and behaviors by reflecting on how they themselves think and behave and then employing these metacognitions to simulate others’ thoughts, feelings, experiences, and behaviors (Dimaggio et al. 2008). Following this general logic, we propose that because people are more likely to deviate from a default when they have carefully deliberated about their own evaluations and decisions, they may believe that the same is true of others. In other words, people may perceive that others’ evaluations are more thoughtful when those evaluations deviate from a default. Furthermore, because people tend to believe that increased thought, or deliberation, leads to more accurate assessments (Barden and Petty 2008; Barden and Tormala 2014; Reich and Tormala 2013), they may infer that a person who deviates from a default has a more accurate assessment. And, importantly, substantial research reveals that people are more persuaded by assessments that they perceive to be more accurate (because people themselves seek to hold accurate opinions; Chaiken and Eagly 1989; Filieri and McLeay 2014; Priester and Petty 1995; Stephen and Grewal 2015). Thus, if people perceive that deviating from a default signals that a deviating review is more accurate, people may be more persuaded by deviating (vs. nondeviating) reviews. Of course, this analysis rests on the assumption that consumers think about the accuracy of reviews. Past research is consistent with this idea. Substantial literature suggests that people are often skeptical of the accuracy of reviews, and that as a result they spontaneously evaluate the likelihood that a review is accurate or inaccurate before deciding to rely upon it (Cheung et al. 2009; Larson and Denton 2014; Schlosser 2011; Willemsen, Neijens, and Bronner 2012). Thus, accuracy is an important dimension on which consumers evaluate reviews. Importantly, though, this evaluation process is inherently subjective. We postulate that consumers perceive deviatory reviews to be more accurate, but not that deviatory reviews actually achieve greater objective accuracy. A rich literature suggests that people are more influenced by reviews that they subjectively assess to be more accurate (Connors, Mudambi, and Schuff 2011; Filieri and McLeay 2014; Cheung et al. 2009). In short, we hypothesize that people perceive a reviewer’s evaluation of a product or service to be more accurate, and thus are more persuaded by it, when the reviewer’s evaluation deviates from rather than adheres to a perceived default. One consequence of this effect is that when the default evaluation is believed to be extremely positive, a moderately positive review might be more persuasive. Indeed, evaluative defaults often emerge in the context of reviews, and frequently they are extremely positive. For example, numerous review portals feature the highest possible default star ratings when consumers enter reviews (Active Pest Control 2018; Cress Circle 2018; Durham Volkswagen 2018; Future Marketing Partners 2016; PrestaShop 2015). Consumers who review Durham Volkswagen, for instance, are directed to a web page featuring a preselected 5-star rating. Consumers can submit this preselected rating or toggle to a lower rating before submitting it. Similarly, companies often communicate a descriptive norm to potential reviewers by noting that the majority of their customers give them the highest possible rating (Groupon 2016; Printing Center USA 2018; and Yelp 2015), which may also create an evaluative default (Miller and Prentice 1996). Consumers also form impressions about the normative review expressed in various review contexts. For example, consumers perceive that the normative review is extremely positive on Zillow, YouTube, Uber, and Lyft (City-Data Real Estate Forum 2013; Crotty 2009; Kane 2015; Quora 2013; Siegler 2009; Zillow Advice Thread 2012, 2016). In each of these contexts, we predict that reviews will be more persuasive if they are moderate rather than extreme, because the deviatory status of moderate reviews makes them seem more thoughtful and accurate. We conducted an exploratory test of this possibility by analyzing real consumer review data scraped from an online retailer that sells home goods and accessories. Like the consumer review platforms described above, the vast majority (72.7%) of reviews on this platform feature a rating of 5 out of 5 stars. Conveniently, the platform also allows consumers to indicate if they find a particular review to be helpful. Prior theorizing suggests that helpful votes are a reasonable proxy of persuasion because consumers likely rate reviews as helpful when those reviews are more useful or impactful to their own decisions (Otterbacher 2008). The data included 60,358 reviews. Because helpful votes provide count data, we analyzed them with a Poisson regression (i.e., a generalized logistic regression employed to analyze count data; Faraway 2006; Wood, McInnes, and Norton 2011). First, we conducted a Poisson regression with dummy coding (nondeviatory 5-star reviews were coded as a 1 and deviatory non-5-star reviews were coded as a 0) in which reviews’ deviatory status was entered as a predictor and the number of helpful votes that each review accrued was entered as the dependent variable. This analysis revealed that consumers rated deviatory reviews as more helpful than nondeviatory reviews, b = –.40, z = –32.79, p < .001 (table 1a). Further analysis revealed that this effect persisted when we controlled for each of the scraped covariates. Specifically, we conducted a Poisson generalized linear mixed-model analysis controlling for each review’s year, each review’s word count, and each reviewed product’s price; nesting each review within the reviewed product and the reviewed product's category; and including random effects for product and product category. This analysis again revealed that consumers rated deviatory reviews as more helpful than nondeviatory reviews, b = –.46, z = –6.77, p < .001 (table 1b). Table 1A ANALYSIS OF ALL REVIEWS   b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001    b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001  Table 1A ANALYSIS OF ALL REVIEWS   b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001    b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001  Next, we explored whether this pattern persisted when we specifically compared deviatory 4-star reviews to nondeviatory 5-star reviews. This exploratory analysis suggested that consumers rated 4-star reviews as more helpful than 5-star reviews, both when we did not control for the covariates (b = –.07, z = –4.38, p < .001; table 1c) and when we did (b = –.16, z = –2.31, p < .05; table 1d). Although we hesitate to draw definitive conclusions given the uncontrolled nature of these field data, this pattern provides initial evidence consistent with our theorizing. We conducted more direct tests of our theorizing in eight controlled experiments. RESEARCH OVERVIEW In sum, we theorize that people believe that deviatory (vs. nondeviatory) reviews are more accurate, and that people can be more persuaded by deviatory reviews as a result. We test this hypothesis across eight controlled experiments set in the context of both moderate and extreme defaults. First, studies 1–2 examine these predictions in three different product domains. These studies find that when the default review is extremely positive, people are more persuaded by moderately (vs. extremely) positive reviews. Study 2 further suggests that deviatory reviews are more persuasive because people perceive them to be more accurate. Study 3 replicates this effect in the negative domain. Here, against the backdrop of an extremely negative default, consumers are more persuaded to not purchase a product when they view a moderately rather than extremely negative review. Study 3 also measures perceived thoughtfulness and provides evidence consistent with our prediction that people believe that reviewers who author deviatory (vs. nondeviatory) reviews devoted more thought to their review, and this perception of thoughtfulness leads to greater perceived accuracy. Studies 1–3 document cases in which a moderate review becomes more persuasive when it deviates from an extreme default. Studies 4a and 4b reveal that when the default evaluation is moderate, the opposite occurs—that is, people are more persuaded by deviatory extreme (vs. nondeviatory moderate) reviews. Study 5 employs a moderation design to provide converging evidence for the underlying role of perceived thoughtfulness in driving the deviation effect. Study 6 establishes the deliberative nature of the inferential process underlying the deviation effect. STUDY 1A Study 1a used incentive-compatible decisions to test the hypothesis that people can be more persuaded by deviatory reviews than nondeviatory reviews. Participants read a review for a café on a consumer review website. Typical of the experience consumers undergo when they search reviews online, participants viewed a web page containing a list of the names of nearby cafés, their addresses, and their average ratings. To establish an extreme default for all participants, each of the cafés listed received an average rating of 10 out of 10 stars. After participants viewed this web page, they read a single review for one of the cafés that gave the café either 9 stars or 10 stars. We predicted that participants would be more persuaded to purchase coffee from this café when they viewed a 9-star review (i.e., a review that deviated from the 10-star default) versus a 10-star review (i.e., a review that did not deviate from the 10-star default). Importantly, the stimuli in study 1a were designed to create an externally valid operationalization of the default. Indeed, the normative approach we adopted to establish a default parallels some of the common ways defaults are established in ordinary customer rating contexts. First, every day, millions of consumers visit websites that present lists of products and their average ratings when shoppers search for relevant items online (e.g., Amazon.com, UncommonGoods.com, Yelp.com; Hyken 2017). Second, the vast majority of reviews on numerous websites have the highest possible rating. As a result, there is a widespread belief among consumers that customer reviews in many contexts tend to be extremely positive. For example, consumers believe this is true of Zillow, YouTube, Uber, and Lyft (City-Data Real Estate Forum 2013; Crotty 2009; Kane 2015; Quora 2013; Siegler 2009; Zillow Advice Thread 2012, 2016). Because descriptive norms can create defaults (Miller and Prentice 1996), and these norms are common in everyday consumer experience, we used this approach to establish an extreme default in this initial lab experiment. Method One hundred fourteen undergraduates from an East Coast university participated in a laboratory study for course credit. This study was run as part of a session containing unrelated surveys from different researchers. In this study, all participants read that “The Steam Room” is a café that recently opened near campus (it was in fact fictional and created for the purpose of the study), and they viewed an ostensibly real website featuring customer reviews of The Steam Room as well as several other local cafés. To establish an extreme default, we presented the first page of search results for nearby cafés, and these results revealed that each of the cafés had an average rating of 10 out of 10 stars. After viewing the entire page of results, participants read a single review of The Steam Room. For all participants, this review read: “The coffee is good & the ambiance is cool!” Across conditions, however, we manipulated the extremity of this review by varying whether it was accompanied by a 9-star or a 10-star rating. Following the review, participants indicated how likely they would be to try coffee from The Steam Room on a seven-point scale (1: Not likely at all; 7: Very likely; adapted from Aaker, Vohs, and Mogilner 2010; Jin, He, and Zhang 2014; Kupor, Laurin, and Levav 2015; Reich, Kupor, and Smith 2017). Furthermore, they read that as additional compensation for their participation, they were eligible to receive a 30% discount on up to five cups of coffee from The Steam Room if they prepaid for the coffee now—that is, during the lab session. Participants were asked how many cups of coffee they wanted to prepurchase. The instructions emphasized that this was a real decision, and that at the end of the lab session they would actually prepay for the cups of coffee that they opted to purchase. Participants entered their purchase decision (0 to 5) into an empty field. Although The Steam Room was fictional, we instructed participants that it was real, and that their purchase decision was real, in order to capture actual purchase decisions. To assess whether participants were suspicious of the experimental instructions, after participants entered their choice we asked them whether they had any comments about the survey. Participants entered their responses into an empty field. Following this question, participants were debriefed; most importantly, they learned that The Steam Room was not a real café and that they would not pay for any coffee they opted to purchase. Only two participants (1.8% of the sample) indicated that they were suspicious about whether the café or the purchase decision was real. These participants were retained in the analysis, but excluding them did not alter the significance of the results. Results and Discussion Participants indicated that they were more likely to try coffee from The Steam Room when the reviewer gave it 9 stars (i.e., when the review deviated from the 10-star default; M = 4.58, SD = 1.13) rather than 10 stars (i.e., when the review did not deviate from the default; M = 3.95, SD = 1.33), t(112) = 2.67, p = .009 (Cohen’s d = .51). Participants also purchased more coffee when the reviewer gave 9 stars (M = 3.15, SD = 1.95) rather than 10 stars (M = 2.26, SD = 2.06), t(112) = 2.35, p = .020 (Cohen’s d = .45). In short, the more moderate positive review prompted more favorable behavioral intentions and more purchasing than did the more extreme positive review. Because we established an extremely positive default in this study, this finding provides initial evidence for the notion that deviatory reviews can prompt more purchases than nondeviatory reviews. STUDY 1B The primary aim of study 1b was to replicate the findings of study 1a despite numerous procedural changes. Most importantly, we employed a different operationalization of the default. Numerous websites feature preselected star ratings when consumers enter reviews—for instance, 5 stars may be selected by default, but the consumer can toggle it to a lower rating (Done Right 2016; Dr. Oogle 2016; Future Marketing Partners 2016; PrestaShop 2015; Durham Volkswagen 2018). In study 1b, we leveraged this preselected rating format. We informed participants that a particular review platform automatically populated a default star rating of 10 out of 10 stars when consumers entered reviews. We predicted that participants would be more persuaded by reviews that deviated from, rather than adhered to, that default rating. We made several other minor methodological changes. First, we sought to assess whether a more moderate rating than the one employed in study 1a could still outperform an extreme rating. Therefore, in this study, we manipulated whether a reviewer gave an 8-star or 10-star rating. In addition, we used a real product instead of coffee from a fictional café, and rather than asking participants how many units of that product they would purchase we asked them to choose between the focal product and a 50-cent bonus. As in study 1a, this was designed to be an incentive-compatible decision. Method One hundred thirty-four undergraduates from an East Coast university participated in a laboratory study for course credit. This study was run as part of a session containing unrelated surveys from different researchers. In this study, all participants read about FoodNow, an online grocery shopping service that allows customers to review the groceries that they purchase on the FoodNow website. Participants further read that when customers complete their review, the website automatically enters a default rating of 10 out of 10 stars. Next, participants read that Sunbelt Bakery Apple Cinnamon granola bars (a real product) had recently been added to the FoodNow website, and that the most recently submitted review for these granola bars was submitted 20 seconds ago by a customer named Mark. Participants in the extreme (vs. moderate) review condition read that Mark rated the granola bars 10 (vs. 8) stars. Next, participants indicated how likely they would be to try Sunbelt Bakery Apple Cinnamon granola bars on a seven-point scale (1: Not likely at all; 7: Very likely). Furthermore, they read that as additional compensation for their participation, they were eligible to receive either 50 cents or a Sunbelt Bakery Apple Cinnamon granola bar. The instructions emphasized that this was a real decision and that at the end of the lab session they would actually be entered into a lottery to receive the option that they chose. Participants indicated their choice by selecting a radio button labeled either “50 cents” or “A Sunbelt Bakery Apple Cinnamon granola bar.” Finally, all participants were debriefed. They learned that Sunbelt Bakery granola bars were real, but that the information they received about the FoodNow website and the review was not, and that all participants would therefore receive an additional 50 cents regardless of their choice. Results and Discussion Participants indicated that they were more likely to try the granola bar when the reviewer gave it 8 stars (i.e., when the review deviated from the 10-star default; M = 4.29, SD = 1.51) rather than 10 stars (i.e., when the review did not deviate from the default; M = 3.48, SD = 1.50), t(132) = 3.13, p = .002 (Cohen’s d = .55). Participants also chose the granola bar over the monetary compensation more frequently when the reviewer gave 8 stars (73.8%) rather than 10 stars (55.1%), χ2 = 5.13, p = .023 (Cohen’s d = .40). In sum, despite numerous procedural changes from study 1a, this study provided converging evidence for the notion that deviatory reviews can be more persuasive than nondeviatory reviews. In study 2, we examined whether this occurs because deviatory reviews are perceived to be more accurate. STUDY 2 Study 2 had multiple objectives. First, we tested the hypothesis that deviatory reviews can be more persuasive because people perceive them to be more accurate. In addition, we examined whether the effect would generalize to a different participant population. To this end, we conducted the study online (using Amazon’s Mechanical Turk) rather than with undergraduates in the lab. Finally, to lend further external validity to our studies, we tested the deviation effect in a context in which the default is implicit rather than experimentally established. To accomplish this goal, we moved the study into the context of ridesharing services. Many consumers are aware that the majority of ridesharing reviews contain the highest possible rating (Kane 2015; Quora 2013). Indeed, when we asked 152 Mechanical Turk participants which star rating consumers most often give ridesharing drivers, the majority (60.5%) indicated that the highest possible rating was most frequent (13.8% indicated that a rating lower than the highest possible rating was most frequent and 25.7% indicated that they did not know). We predicted that such widespread awareness of default ratings in this context would be sufficient to produce the deviation effect, even without our establishing the default explicitly in the context of the experiment. Method One hundred ninety-nine Mechanical Turk participants completed an online study for payment. All participants imagined that there was a new ridesharing service, and that they were looking at a consumer review of it on a website with reviews of different ridesharing services. Participants received no explicit information about the distribution of review ratings on this website; they simply read about an individual review of the new ridesharing service. Participants in the extreme review condition read that the reviewer gave the company 10 out of 10 stars, whereas participants in the moderate review condition read that the reviewer gave the company 8 out of 10 stars. Next, all participants indicated the extent to which they perceived that the reviewer’s opinion about the ridesharing service was accurate. Participants responded on a seven-point scale (1: Not at all; 7: Very much; adapted from Brett and Atwater 2001; Davies 1994; Grohmann, Spangenberg, and Sprott 2007). We also assessed participants' behavioral intentions. Results for exploratory items measured in this study and subsequent studies are included in web appendixes A and B. Results and Discussion Participants perceived the review to be more accurate in the moderate (M = 5.09, SD = .96) rather than extreme (M = 4.63, SD = 1.26) review condition, t(197) = 2.88, p = .004 (Cohen’s d = .41). Participants also indicated that they were more likely to use the ridesharing company in the moderate (M = 5.16, SD = 1.11) rather than extreme (M = 4.69, SD = 1.35) review condition, t(197) = 2.65, p = .009 (Cohen’s d = .38). A mediation analysis with bootstrapping was consistent with the notion that perceived accuracy mediated the behavioral intention effect (95% CI: .1056 to .6003; figure 1). FIGURE 1 View largeDownload slide MEDIATION OF THE BEHAVIORAL INTENTIONS EFFECT THROUGH PERCEIVED ACCURACY IN STUDY 2 NOTE.—The path coefficients are unstandardized betas. The value in parentheses indicates the effect of condition on the dependent variable after we control for the mediator. *p < .05, **p < .01, ***p < .001. FIGURE 1 View largeDownload slide MEDIATION OF THE BEHAVIORAL INTENTIONS EFFECT THROUGH PERCEIVED ACCURACY IN STUDY 2 NOTE.—The path coefficients are unstandardized betas. The value in parentheses indicates the effect of condition on the dependent variable after we control for the mediator. *p < .05, **p < .01, ***p < .001. In short, the deviation effect—whereby a moderate endorsement can be more persuasive than an extreme endorsement when extreme is the default—appears to be mediated at least in part by perceptions that the moderate endorsement is more accurate. In addition to providing initial process evidence, study 2 speaks to the robustness of the effect. We replicated the core finding of studies 1a and 1b in a different participant population and in a different consumer context in which people implicitly know the default without researchers referencing or displaying it in the experimental context. STUDY 3 Studies 1 and 2 explored the consequences of deviating from an extremely positive default to a moderately positive evaluation. Study 3 further tested the robustness of the deviation effect by examining whether it extends to the negative domain—for example, a deviation from an extremely negative default to a moderately negative evaluation. Participants in study 3 viewed a 1-star or 2-star review on a website that had a 1-star default. We surmised that participants would be less likely to purchase the reviewed product after viewing a 2-star (deviatory) review rather than a 1-star (nondeviatory) review. In other words, we predicted that participants would be more persuaded that the product was of poor quality after reading a moderately (vs. extremely) negative review of it. In addition to testing the deviation effect in a negative context, study 3 offered an initial test of our hypothesis that people perceive deviatory reviews to be more accurate because they believe reviewers who author deviatory (vs. nondeviatory) reviews devoted more thought to their review. As described earlier, we theorize that it is this perception of increased thoughtfulness that leads people to perceive these reviewers’ evaluations to be more accurate and persuasive. Against the backdrop of an extremely negative default, then, we hypothesized that a moderately (vs. extremely) negative review would seem more thoughtful and thus accurate, which would increase its negative impact. To assess this process, we measured perceived thoughtfulness and accuracy in study 3. Importantly, we measured these variables after behavioral intentions. It could be argued that in study 2 we made the hypothesized process (through accuracy) especially salient by measuring it immediately before behavioral intentions. Studies 1a and 1b did not measure process at all, so it is unlikely that this measure is required for the deviation effect to obtain. Nevertheless, to ensure that the deviation effect is not contingent upon measuring perceptions of accuracy immediately before assessing consumers’ behavior, we reversed the order of these measures in study 3. Finally, study 3 further tested the generalizability of the deviation effect by examining it in a different participant pool. Whereas the previous studies documented the deviation effect in samples of university students (studies 1a and 1b) and Mechanical Turk workers (study 2), study 3 examined whether it extends to participants recruited through SurveyMonkey’s online survey platform. Method One hundred ninety-four participants were recruited from SurveyMonkey to take part in an online study for a donation to a charity. All participants read about a restaurant review website that automatically enters a 1-star rating that reviewers can change when they review a restaurant. Participants in the extreme (vs. moderate) review condition read that they see a restaurant review on this website, and the reviewer rated the restaurant 1 (vs. 2) stars. We assessed participants’ interest in trying the restaurant using the measure employed in study 1 (adapted to refer to the restaurant). On the next survey screen, participants evaluated the review’s accuracy via the measure described in study 2 (adapted to refer to the restaurant), as well as how much thought they perceived that the reviewer put into his assessment of the restaurant. Participants responded on a seven-point scale (1: Not much at all; 7: A lot; adapted from Kupor et al. 2014; Schrift and Amar 2015; Zhang and Epley 2012). Results and Discussion As predicted, participants were less interested in trying the restaurant when the review was moderately (M = 2.48; SD = 1.42) rather than extremely (M = 3.03; SD = 1.55; t(192) = 2.58, p = .011; Cohen’s d = .37) negative. Also, participants perceived the moderate (vs. extreme) review to be more thoughtful (MModerate = 4.43, SDModerate = 1.55; MExtreme = 3.08, SDExtreme = 1.70; t(192) = 5.76, p < .001; Cohen’s d = .83) and accurate (MModerate = 4.46, SDModerate =1.26; MExtreme = 3.26, SDExtreme = 1.51; t(192) = 6.03, p < .001; Cohen’s d = .87). We predicted that participants would be less likely to try the restaurant when they viewed the moderately negative review because its deviation from the extremely negative default signaled that the reviewer was more thoughtful in his assessment, which in turn fostered the perception that this assessment was more accurate. A serial mediation with bootstrapping (Hayes 2013) was consistent with this account (95% CI: –.4517, –.0036; figure 2). In short, we replicated the deviation effect and obtained evidence consistent with its proposed mechanism despite numerous changes to the experimental paradigm—specifically, testing it in the context of an extremely negative default, using a different participant population, and reversing the order of the process and outcome measures. FIGURE 2 View largeDownload slide MEDIATION MODELS IN STUDY 3 (FIRST PANEL), STUDY 4A (SECOND PANEL), STUDY 4B (THIRD PANEL), AND STUDY 5 (FOURTH PANEL) NOTE.—The path coefficients are unstandardized betas. Values in parentheses indicate the effect of the interaction on the dependent variable after we control for the mediators. *p < .05, **p < .01, ***p < .001, ^p < .10. FIGURE 2 View largeDownload slide MEDIATION MODELS IN STUDY 3 (FIRST PANEL), STUDY 4A (SECOND PANEL), STUDY 4B (THIRD PANEL), AND STUDY 5 (FOURTH PANEL) NOTE.—The path coefficients are unstandardized betas. Values in parentheses indicate the effect of the interaction on the dependent variable after we control for the mediators. *p < .05, **p < .01, ***p < .001, ^p < .10. STUDY 4A The first four studies provided initial evidence for the deviation effect: people believe reviews that deviate from a default evaluation are more thoughtful and accurate, and thus are more persuaded by them. However, in these studies the evidence was restricted to cases in which a moderate review deviated from an extreme default. In study 4a, we varied whether a reviewer expressed a moderately or extremely positive review in addition to whether the default was moderately or extremely positive. We hypothesize that people perceive a moderately positive review to be more thoughtful and accurate when the default is extremely positive, but that they perceive an extremely positive review to be more thoughtful and accurate when the default is moderately positive. In other words, we predict that people perceive deviatory reviews to be more thoughtful and accurate—and are more persuaded by them—regardless of whether the deviatory review is more moderate or extreme than the default. Study 4a tested this prediction in the context of restaurant reviews and a 5-star rating scale. Method Four hundred Mechanical Turk participants completed an online study for payment. Participants were randomly assigned to conditions in a 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) between-participants design. Participants in the extreme (vs. moderate) default condition read that when reviewers enter a review on a particular review website, the website automatically populates a default rating of 5 (vs. 4) stars that reviewers can change. Next, participants in the extreme (vs. moderate) review condition read that they see a review of a restaurant on this website, and the reviewer rated it 5 (vs. 4) stars. After participants read this information, they indicated their perceptions of the reviewer’s thoughtfulness and accuracy, as well as their own likelihood of trying the restaurant, using the measures employed in study 3. In this study and study 4b, we measured perceived thoughtfulness and accuracy before behavioral intentions. Results Perceived Thoughtfulness We began by submitting perceived thoughtfulness to a 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) ANOVA. This analysis uncovered a main effect of default, F(1, 396) = 5.64, p = .018, but not of the reviewer’s evaluation, F(1, 396) = 1.66, p = .199. Most importantly, there was a significant interaction, F(1, 396) = 90.41, p < .001 (table 2). When there was an extreme default, participants inferred that the reviewer devoted more thought to his evaluation when his review was moderate rather than extreme, F(1, 396) = 59.20, p < .001 (Cohen’s d = 1.08). In contrast, when there was a moderate default, participants inferred the reviewer had thought more about his assessment when his review was extreme rather than moderate, F(1, 396) = 33.27, p < .001 (Cohen’s d = .82). TABLE 1B ANALYSIS OF ALL REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      TABLE 1B ANALYSIS OF ALL REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      Perceived Accuracy Analysis of the perceived accuracy data revealed a main effect of the reviewer’s evaluation, F(1, 396) = 7.29, p = .007, but not the default, F(1, 396) = 2.05, p = .153. Most relevant to our theorizing, we found the predicted interaction, F(1, 396) = 87.61, p < .001 (table 2). When there was an extreme default, participants thought that the reviewer’s evaluation was more accurate when his review was moderate rather than extreme, F(1, 396) = 73.88, p < .001 (Cohen’s d = 1.22). In contrast, when the default was moderate, participants thought that the reviewer’s opinion was more accurate when his review was extreme rather than moderate, F(1, 396) = 21.84, p < .001 (Cohen’s d = .66). Behavioral Intentions The behavioral intentions data revealed no main effect of default, F(1, 396) = 2.09, p = .149, or the reviewer’s evaluation, F(1, 396) = .06, p = .805, but we again observed the predicted interaction, F(1, 396) = 58.75, p < .001 (table 2). When there was an extreme default, participants were more likely to try the restaurant when the reviewer’s evaluation was moderate rather than extreme, F(1, 396) = 31.79, p < .001 (Cohen’s d = .81). In contrast, when there was a moderate default, participants were more likely to try the restaurant when they viewed an extreme evaluation rather than a moderate one, F(1, 396) = 27.09, p < .001 (Cohen’s d = .72). Mediation Replicating study 3, a serial mediated moderation model with bootstrapping provided evidence consistent with our prediction that the interaction effect (when we controlled for the main effects) on behavioral intentions occurred because the deviation boosted perceived thoughtfulness, which in turn boosted perceived accuracy (95% CI: –2.7662, –1.7445; figure 2). STUDY 4B Study 4b sought to replicate the findings from study 4a in a different context (house cleaner reviews) and using a different review rating scale (10 rather than 5 stars). Method Four hundred one Mechanical Turk participants completed an online study for payment. As in study 4a, participants were randomly assigned to conditions in a 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) between-participants design. All participants were asked to imagine that there was a new house cleaning service called “Elite Cleaners” and that they were reading reviews of it on a review website. Participants in the extreme (vs. moderate) default condition further read that when reviewers submit a review to this website, it automatically enters a default star rating of 10 (vs. 8) stars that reviewers can change. Next, participants in the extreme (vs. moderate) review condition read that they see a review of Elite Cleaners on this website, and the reviewer rated it 10 (vs. 8) stars. All participants then indicated their perception of the reviewer’s thoughtfulness, the review’s accuracy, and their willingness to use Elite Cleaners using the measures employed in the previous studies (adapted to refer to Elite Cleaners). Results Perceived Thoughtfulness A 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) ANOVA on the perceived thoughtfulness data revealed no main effect of default, F(1, 397) = 2.45, p = .118, or the reviewer’s evaluation, F(1, 397) = .16, p = .688. However, the predicted interaction emerged, F(1, 397) = 176.57, p < .001 (table 2). When there was an extreme default, participants perceived the review to be more thoughtful when it was moderate rather than extreme, F(1, 397) = 92.97, p < .001 (Cohen’s d = 1.44). In contrast, when there was a moderate default, participants perceived the review to be more thoughtful when it was extreme rather than moderate, F(1, 397) = 83.69, p < .001 (Cohen’s d = 1.25). Perceived Accuracy Analysis of the perceived accuracy data revealed a main effect of default, F(1, 397) = 4.30, p = .039, but not the reviewer’s evaluation, F(1, 397) = .70, p = .402. Most importantly, the predicted interaction emerged, F(1, 397) = 150.05, p < .001 (table 2). When there was an extreme default, participants thought that the reviewer’s evaluation was more accurate when his review was moderate rather than extreme, F(1, 397) = 85.38, p < .001 (Cohen’s d = 1.38). In contrast, when there was a moderate default, participants thought that the reviewer’s evaluation was more accurate when it was extreme rather than moderate, F(1, 397) = 65.31, p < .001 (Cohen’s d = 1.10). Behavioral Intentions The behavioral intentions data revealed a main effect of the reviewer’s evaluation, F(1, 397) = 6.07, p = .014, but not of the default, F(1, 397) = .17, p = .678. More crucially, we found a significant interaction, F(1, 397) = 62.65, p < .001 (table 2). When there was an extreme default, participants were more likely to use Elite Cleaners when the reviewer’s evaluation was moderate rather than extreme, F(1, 397) = 14.74, p < .001 (Cohen’s d = .55). In contrast, when there was a moderate default, participants were more likely to use Elite Cleaners when the reviewer’s evaluation was extreme rather than moderate, F(1, 397) = 54.29, p < .001 (Cohen’s d = 1.03). Mediation Following the same procedure as in study 4a, a serial mediated moderation model with bootstrapping provided evidence consistent with our prediction that the deviatory reviews elicited more favorable behavioral intentions because they boosted perceived thoughtfulness, which in turn boosted perceived accuracy (95% CI: –2.7486, –1.8723; figure 2). STUDY 5 Studies 4a and 4b suggested that people infer that deviatory reviews are more accurate because deviation signals that the reviewer devoted careful thought to the review. In study 5, we employed a moderation design to provide converging evidence for the role of perceived thoughtfulness in this deviation effect. In particular, if the deviation effect occurs because deviatory reviews signal greater thoughtfulness, other evidence that a reviewer thought carefully about his or her recommendation should moderate the effect. In other words, if consumers perceive nondeviatory (vs. deviatory) reviews to be less accurate because consumers believe they are the result of less careful thought, then awareness that a nondeviatory review was the result of considerable thought should eliminate the persuasive advantage of deviatory reviews. Participants in study 5 were randomly assigned to view either an extremely lengthy review containing an elaborated discussion of the reviewed product, or a shorter review with less extensive discussion. We predicted that the considerable length of the former review would lead participants to conclude—regardless of the review’s deviatory status—that the reviewer devoted substantial thought to his review. As a result, we hypothesized that participants would find this review to be relatively persuasive regardless of its deviatory status. In contrast, when the review was relatively brief, we predicted that it would lack a clear signal regarding the author’s thoughtfulness, in which case perceived thoughtfulness (and thus accuracy and persuasiveness) would be derived from the review’s deviatory status. Several additional aspects of study 5 are worth noting. First, to enhance the external validity of the findings we leveraged real review content. Specifically, participants read real consumer reviews (sourced from Amazon.com), and we simply varied whether each review’s star rating deviated from a default. Second, to ensure that our findings were not dependent on the idiosyncratic content of any particular review, we administered three different long reviews and three different shorter reviews. Thus, participants were randomly assigned to read one review in a 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) × 3 (review content: version 1 vs. version 2 vs. version 3) between-participants design. We made an a priori decision to collapse across the stimulus sampling (i.e., review content) if it did not interact with our design of core theoretical interest—that is, the 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) design. Third, to provide more evidence that the process driving the deviation effect is not dependent on the order in which we measure thoughtfulness, accuracy, and behavioral intentions, we again measured behavioral intentions prior to thoughtfulness and accuracy as we did in study 3. Fourth, to further establish the robustness of the effect, we used a different product context: reviews for pens. Finally, we further extended the generalizability of the findings by conducting study 5 with a different online participant panel: Survey Sampling International (SSI). Method One thousand two hundred twenty-three participants, recruited from SSI, were randomly assigned to conditions in a 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) × 3 (review content: version 1 vs. version 2 vs. version 3) between-participants design. In this study, participants read a review for a pen that ostensibly had been posted on a consumer review website. Following the same procedure as in study 1a, all participants viewed an extreme default. Specifically, participants viewed an initial page of search results for pens, and all of the reviews displayed gave 5 out of 5 stars. Participants then read one of these reviews, which was either one of three long reviews (containing an elaborated discussion of the reviewed pen), or one of three brief reviews. All were real reviews extracted from Amazon, but we altered two features. First, to hold brand name constant we replaced all brand mentions with the (ostensibly real) brand name “Active Pen.” Second, we manipulated the review’s star rating. In the extreme (vs. moderate) review conditions, the pen’s rating was 5 (vs. 4) stars. After participants read the review, they indicated their likelihood of purchasing the pen as in studies 4a and 4b (this measure was adapted to refer to the pen). On the next survey screen, participants indicated their perceptions of the review’s accuracy and thoughtfulness, again assessed via the measures employed in studies 4a and 4b (adapted to refer to the pen). Results As previously noted, our core theoretical interest in study 5 was in the 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) design. As a robustness check—to ensure that any effects observed were not due to idiosyncratic features of a particular review we happened to sample—participants viewed one of three short reviews or three long reviews, and we manipulated whether that review was associated with a 4- or 5-star rating. Thus, we first conducted 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) × 3 (review content: version 1 vs. version 2 vs. version 3) ANOVAs on each dependent measure to test whether our stimulus sampling interacted with the 2 × 2 design of theoretical interest. As noted, we made an a priori decision that if there was no three-way interaction we would collapse across the stimulus samples to test our core predictions. Behavioral Intentions The 2 × 2 × 3 ANOVA on behavioral intentions revealed no three-way interaction, F(1, 1212) = 2.60, p = .107, but the predicted two-way interaction between reviewer evaluation and review length did emerge, F(1, 1212) = 3.94, p = .047. Therefore, we collapsed across review content condition. A subsequent 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) ANOVA on the collapsed data revealed main effects for review length, F(1, 1219) = 119.91, p < .001, and reviewer evaluation, F(1, 1219) = 8.01, p = .005, qualified by the predicted interaction, F(1, 1219) = 11.53, p = .001. As shown in table 3, participants reported that they were more likely to buy the reviewed pen when they viewed the moderate rating under short, F(1, 1219) = 19.32, p < .001 (Cohen’s d = .35), but not long, F(1, 1219) = .16, p = .688, review conditions. TABLE 1C ANALYSIS OF 4-STAR AND 5-STAR REVIEWS   b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001    b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001  TABLE 1C ANALYSIS OF 4-STAR AND 5-STAR REVIEWS   b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001    b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001  TABLE 1D ANALYSIS OF 4-STAR AND 5-STAR REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      TABLE 1D ANALYSIS OF 4-STAR AND 5-STAR REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      TABLE 2 PERCEIVED THOUGHTFULNESS, PERCEIVED ACCURACY, AND BEHAVIORAL INTENTIONS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 4A AND STUDY 4B   Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)    Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)  TABLE 2 PERCEIVED THOUGHTFULNESS, PERCEIVED ACCURACY, AND BEHAVIORAL INTENTIONS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 4A AND STUDY 4B   Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)    Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)  TABLE 3 BEHAVIORAL INTENTIONS, PERCEIVED ACCURACY, AND PERCEIVED THOUGHTFULNESS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 5   Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)    Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)  TABLE 3 BEHAVIORAL INTENTIONS, PERCEIVED ACCURACY, AND PERCEIVED THOUGHTFULNESS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 5   Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)    Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)  Perceived Accuracy The 2 × 2 × 3 ANOVA on the perceived accuracy data revealed no three-way interaction, F(1, 1212) = .10, p = .754, but the predicted 2 × 2 interaction emerged, F(1, 1212) = 23.46, p < .001. Therefore, we collapsed across review content condition. A 2 × 2 ANOVA on the collapsed data revealed main effects for review length, F(1, 1219) = 56.81, p < .001, and reviewer evaluation, F(1, 1219) = 6.49, p = .011, qualified by the predicted interaction, F(1, 1219) = 11.56, p < .001 (table 3). As hypothesized, participants viewed the moderate review as more accurate than the extreme review under short, F(1, 1219) = 17.65, p < .001 (Cohen’s d = .32), but not long, F(1, 1219) = .36, p = .546, review conditions. Perceived Thoughtfulness The 2 × 2 × 3 ANOVA on the perceived thoughtfulness data revealed no three-way interaction, F(1, 1212) = .35, p = .552, but the predicted 2 × 2 interaction emerged, F(1, 1212) = 21.83, p < .001. After we collapsed across review content condition, a 2 × 2 ANOVA revealed a main effect for review length, F(1, 1219) = 919.81, p < .001, indicating that the thoughtfulness manipulation was successful. The analysis indicated no main effect of the reviewer’s evaluation, F(1, 1219) = 1.81, p = .179, but the predicted interaction emerged, F(1, 1219) = 4.69, p = .031 (table 3). Participants perceived the moderate review to be more thoughtful than the extreme review under short, F(1, 1219) = 6.13, p = .013 (Cohen’s d = .18), but not long, F(1, 1219) = .34, p = .561, review conditions. Mediation As illustrated in figure 2, a moderated mediation analysis with bootstrapping provided evidence consistent with our prediction that the interaction between review length and reviewer evaluation affected behavioral intentions through perceived thoughtfulness and accuracy (95% CI: .1040, .4088). Discussion In showing that the deviation effect is both mediated and moderated by perceived thoughtfulness, study 5 offers convergent evidence for the mechanism driving this effect. Also important, the moderation pattern observed in study 5 suggests that the deviation effect has limits. In particular, other salient evidence that a reviewer devoted substantial thought to a review eliminates the influence of that review’s deviatory status on perceived thoughtfulness, and thus on perceived accuracy and persuasion. Because this moderation pattern provides crucial insight into the process driving the deviation effect, we conducted a conceptual replication in a different domain (restaurants) with a different thoughtfulness manipulation. Specifically, the replication study presented participants with a review’s rating but not its content, and manipulated thoughtfulness by providing some (but not other) participants with explicit information that a reviewer devoted substantial thought to his review. This study revealed the same moderation and mediation patterns as in study 5 (web appendix C). STUDY 6 As outlined earlier, we propose an attributional mechanism for the deviation effect, whereby people draw inferences about reviewer thoughtfulness and accuracy based on whether a review deviates from or adheres to a default. We theorize that this process is most likely to manifest under conditions that allow for deliberative cognitive processing, because it requires that people reflect on others’ actions (i.e., their moderate or extreme endorsements) and make situational corrections (i.e., for perceived defaults) before forming inferences about them (i.e., their thoughtfulness and accuracy). This sort of attributional reasoning requires ample processing resources (Gilbert, Pelham, and Krull 1988). Past research further reveals that simulating and forming inferences about other people’s thought processes can require careful deliberation (Campbell and Kirmani 2000; Goldman and Jordan 2013). Thus, multiple streams of research suggest that the process we have outlined should be more likely to emerge under conditions that permit rather than restrict thorough cognitive processing. What happens when processing is constrained? Research suggests that people often rely on a “more is better,” or “numerosity,” heuristic when their processing is limited in some way (Khan and Kupor 2016; Pelham, Sumarta, and Myaskovsky 1994; Petty and Cacioppo 1984; Tormala, Petty, and Briñol 2002). This implies that people may simply find higher ratings more persuasive when their processing is constrained. Thus, we predict that if people are unable to deliberate, they will be more persuaded by an extreme review regardless of its deviatory status. To examine this possibility, we presented participants in study 6 with a 4- or 5-star review in the context of a 5-star default, and we manipulated cognitive load. We predicted that participants whose cognitive resources were unencumbered would more favorably evaluate the reviewed product when they viewed the deviatory 4-star review. Importantly, though, we predicted that this pattern would reverse when participants were under cognitive load. Under load, we expected participants to more favorably evaluate the reviewed product when they viewed the 5-star review. If obtained, this result would suggest that the deviation effect is driven by a thoughtful inferential process requiring ample cognitive resources. Method Four hundred sixty Mechanical Turk participants completed an online study for payment. At the beginning of the survey, all participants read that one of the goals of the survey was to research multitasking, and that they might be prompted to switch between tasks during the study. Participants further learned that the study aimed to investigate people’s interests in various topics, so they would listen to a recording and report how interested they were in it. Following these opening instructions, participants were randomly assigned to conditions in a 2 (reviewer evaluation: moderate vs. extreme) × 2 (cognitive load: no vs. yes) between-participants design. Participants in the moderate review condition read a review for a pen in which the reviewer wrote: “Active Pens are good—4 out of 5 stars!” In the extreme review condition, the reviewer wrote: “Active Pens are great—5 out of 5 stars!” To establish a 5-star default, all participants read that this review had been entered into a review platform that automatically preselects a 5-star rating. As noted, participants were also randomly assigned to either a cognitive load condition or a no cognitive load condition. In the load condition, participants viewed the information about the pen while completing a distraction task (adapted from Lin, Lee, and Robertson 2011; Lee, Lee, and Boyle 2007; Sakuraba 2012). Specifically, before reading any information about the reviewed pen or the review platform, participants under load were instructed that while they viewed written information on the next screen, they would listen to an unrelated audio recording about Pluto. Participants in this condition then viewed the focal information about the pen while simultaneously listening to the audio recording. To bolster the cover story, after these participants listened to the recording, they indicated their interest in the topic discussed in the recording on a seven-point scale (1: Not interesting at all; 7: Very interesting). In order to control for any potential impact of the recording’s content, we had participants in the no load condition listen to and rate their interest in the same recording prior to viewing the information about the pen. Thus, all participants listened to the same recording. This recording was used to constrain the processing resources of participants in the load condition but not in the no load condition. Finally, all participants indicated their likelihood of purchasing an Active Pen using the measure employed in study 5. Results and Discussion A 2 (reviewer evaluation: moderate vs. extreme) × 2 (cognitive load: no vs. yes) ANOVA revealed no main effect of load, F(1, 456) = .04, p = .847, or reviewer evaluation, F(1, 456) = .02, p = .889, but the predicted interaction emerged, F(1, 456) = 10.16, p = .002. As expected given the extreme default, participants in the no load condition had more favorable behavioral intentions when they viewed the moderate (M = 4.44; SD = 1.29) rather than extreme (M = 3.99; SD = 1.52) review, F(1, 456) = 5.29, p = .022; Cohen’s d = .32). This effect reversed among participants in the load condition: these participants had more favorable behavioral intentions when they viewed the extreme (M = 4.45; SD = 1.52) rather than moderate (M = 4.04; SD = 1.47) review, F(1, 456) = 4.87, p = .028; Cohen’s d = .27). In short, cognitive load reversed the deviation effect, leading people under load to be more persuaded by an extreme, nondeviating review than by a moderate, deviating review. These results provide support for the notion that the deviation effect is driven by thoughtful attributional reasoning. They also point to an additional constraint on the effect. When consumers’ processing resources are constrained in some way, which may occur when consumers are distracted or under time pressure while searching for products, they are more likely to be persuaded by extreme reviews even when those reviews adhere to a preselected default. Indeed, consumers may simply rely on a “more stars are better” heuristic under these conditions. In contrast, when consumers can think more carefully, the default is more likely to influence their assessments and the deviation effect is thus more likely to emerge. In documenting this boundary on the deviation effect, this study might provide initial insight into the conditions under which spontaneous accuracy inferences drive (vs. do not drive) product evaluations, and thus illuminate the situations in which the deviation effect emerges versus reverses. In particular, study 6 is consistent with the possibility that the accuracy inferred from a deviatory rating primarily shapes behavioral intentions when consumers’ cognitive resources are unencumbered. It could be that extremely positive ratings imply greater quality (regardless of the default) when consumers do not spontaneously consider the accuracy of those ratings, and extremely positive ratings may thus prove more persuasive when consumers’ cognitive resources are restricted in some way. We encourage future research to explore this possibility more deeply. GENERAL DISCUSSION This research provides evidence for a deviation effect in persuasion. We found that when the default evaluation in a review context is extreme, moderate reviews can actually be more persuasive than extreme ones. However, when the default review is moderate, extreme reviews are more persuasive. We obtained reliable evidence for this effect across eight studies despite numerous changes to the product category, study materials, and sample population. For instance, we found the deviation effect in the context of reviews for a café, a granola bar, a ridesharing service, a restaurant, a cleaning service, and pens. We also found it using different operationalizations of deviations, including deviations from defaults established via explicit descriptive norms, an implicit descriptive norm, and preselected star ratings. Also important, we observed the deviation effect in diverse samples, including university students and participants from numerous online panels: Mechanical Turk, SurveyMonkey, and SSI. Finally, we observed these effects on measures of behavioral intentions and actual choice. Importantly, our studies offer both mediation and moderation evidence consistent with the proposed account for this effect—that is, that people perceive deviatory reviews to be more accurate because they believe those reviews are the result of greater thought. It is worth noting that recent work (Pieters 2017) has outlined a series of conditions that researchers should strive to meet before drawing conclusions about mediation effects. These include directionality, reliability, unconfoundedness, distinctiveness, and power. As outlined below, we made an effort to follow these recommendations in the current work. First, to establish directionality, we not only measured the proposed mediator (studies 2–5), but also manipulated it (study 5 and the supplemental study in web appendix C). We also leveraged prior literature (Kupor et al. 2014; Priester and Petty 1995, 2003; Schrift and Amar 2015) suggesting that perceptions of a source’s thoughtfulness and accuracy can drive persuasive outcomes even when those variables are experimentally manipulated, making the reverse causal direction implausible. To ensure that our measures were reliable, we borrowed heavily from past research (cited in our methods sections) showing that these measures are both sensitive to experimental manipulations and predictive of related outcomes. To reduce confoundedness between measures, in addition to directly manipulating the mediator (as noted), we created spatiotemporal distance between the mediators and outcomes in several of our studies (studies 3 and 5) by placing those measures on separate survey screens. Furthermore, as we explain in the forthcoming section, “Alternative Explanations,” we rule out numerous competing accounts on conceptual grounds. Regarding distinctiveness of our mediator and outcome variables, we note that (1) past research has documented numerous contexts in which these constructs diverge (Kupor et al. 2014; Schrift and Amar 2015), and (2) the raw correlations between our mediators and outcome variables ranged from .19 to .82 across studies. Finally, in an effort to test our hypotheses with reasonable power, we followed the Simmons, Nelson, and Simonsohn (2013) recommendation of collecting at least 50 participants per condition in studies 1a and 1b, and collected approximately 100 participants per condition in all subsequent studies. Follow-up analyses revealed that several of the studies testing mediation (studies 4a and 4b and web appendix C) had at least 80% power to identify direct and indirect effects. Taken together, these features suggest that the current studies provide meaningful mechanistic insight. In terms of theoretical contribution, this research is the first to suggest that the perceived default in a review context can shape people’s willingness to be influenced by an extreme or moderate review. Moreover, while past research has revealed much about the effects of preselected options on consumers’ judgments (Everett et al. 2015; Johnson and Goldstein 2003; Probst et al. 2013), it has not examined how observers perceive (and are thus influenced by) the judgments of others who adhere to or deviate from these defaults. Thus, the current studies advance our understanding of the diverse effects of defaults in consumer contexts. This research is also the first to show that a moderate review can be more persuasive than an extreme review. This finding is especially interesting in that it conflicts with people’s lay intuitions about what should be persuasive. Indeed, a follow-up study revealed that people do not intuit the persuasive power of moderate reviews in the context of extreme defaults. In this study, we asked participants to imagine review websites on which the default rating was 5 stars. Participants predicted that their own reviews on these websites would be more persuasive if they rated a product 5 rather than 4 stars, and thus that they would be more likely to give 5 stars if they sought to persuade others to purchase the product (see full details in web appendix D). This study suggests that consumers do not intuit that deviatory reviews can be more persuasive. Thus, in their attempts to increase their own persuasive influence, people might inadvertently decrease it by avoiding moderate endorsements precisely when such endorsements would be more persuasive. In demonstrating the effect of deviations on perceptions of thoughtfulness and accuracy, the current studies may also uncover a novel determinant of source credibility. In general, sources believed to be more thoughtful and accurate are seen as more credible (Berlo, Lemert, and Mertz 1969; Chesney 2006; Flanagin and Metzger 2000; Hovland, Janis, and Kelley 1953; Kaufman, Stasson, and Hart 1999; McCroskey 1966), and credibility is a well-known factor in persuasion (Briñol and Petty 2009; Petty, Cacioppo, and Schumann 1983). Consistent with this general possibility, we obtained mixed evidence for the notion that deviations impact perceptions of trustworthiness—a well-established component of credibility—in the studies in which we measured it (see web appendixes A and B). Although the current research was not designed to delineate the role of deviations in shaping source credibility, it may hint at a novel means of establishing credibility: authoring a deviatory review. Particularly intriguing is the notion that making one’s review more moderate might in some cases make one seem more credible. Future research investigating how deviations might shape perceptions of credibility in its different forms (i.e., expertise and trustworthiness; Conchie and Burns 2009; McGinnies and Ward 1980; Ong 2012) would be worthwhile. The current research also contributes to the word-of-mouth (WOM) literature. Recent work on WOM has examined how a review’s impact can be shaped by reviewer characteristics (e.g., reviewer expertise and similarity to the observer; Forman, Ghose, and Wiesenfeld 2008; Reichelt, Sievert, and Jacob 2014; Rosario et al. 2016), product characteristics (e.g., product lifespan, product risk, and product category; Kupor, Tormala, and Norton 2014; Murray 1991; Rosario et al. 2016; Wangenheim and Bayón 2004), and characteristics of the review itself (e.g., review length and valence; Chevalier and Mayzlin 2006; Doh and Hwang 2009). In addition, this literature has begun to examine how the review platform can moderate the impact of reviews. For example, platforms that impose more structure in their display—by including review summaries, for instance—have been found to exert greater impact on consumer behavior (Rosario et al. 2016). Our research contributes to this growing literature in revealing that another feature of review platforms—that is, the perceived default rating—can moderate the impact of the reviews presented. Alternative Explanations Although we provided evidence consistent with the proposed mechanism behind the deviation effect, it is important to consider alternative processes that might have played an added role. Consideration of Negatives First, could the current results reflect a blemishing or two-sided messaging effect? The blemishing effect refers to the idea that delivering a small dose of weak negative information after strong positive information can enhance persuasion (Ein-Gar, Shiv, and Tormala 2012). Likewise, the two-sided messaging effect suggests that people are sometimes more persuaded by recommenders who have considered both a product’s positives and negatives (e.g., rather than just its positives; Eisend 2006; Rucker, Petty, and Briñol 2008). Could it be that moderate reviews outperformed extreme reviews in the current studies because they seemed to incorporate negative information? Although it is theoretically possible that moderate endorsements could foster the perception that negative information has been considered, this would not offer a viable account for the current results. For instance, studies 4a and 4b found that a moderately positive evaluation is more persuasive in the context of an extremely positive default, but that the opposite is true in the presence of a moderate default. This result is inconsistent with both a blemishing and a two-sided message account. Neither would predict that manipulating an external default (which does not alter whether the review itself contains negative information) would moderate the deviation effect. After all, if moderate reviews signal that negative information has been considered, they should do so regardless of the default. Studies 3 and 6 provide further evidence inconsistent with the possibility that the blemishing effect underlies the deviation effect. First, the blemishing effect exclusively occurs when a small dose of negative information is added to otherwise positive information, but study 3 finds that the deviation effect persists when a reviewer deviates from an extremely negative default to a moderately negative evaluation. Study 6 is also inconsistent with a blemishing effect—the blemishing effect most reliably emerges under conditions of low elaboration (Ein-Gar et al. 2012), whereas precisely the opposite is true for the deviation effect (study 6). Furthermore, in additional empirical work (see web appendix E), we found that the deviation effect extends beyond deviations from rating extremity defaults. Deviations from other types of defaults foster persuasion as well. For example, companies often communicate a default by highlighting on their websites that the majority of consumers give them specific positive feedback in reviews (Groupon 2016; Printing Center USA 2018; Yelp 2015). When companies communicate such a default (e.g., “Most people say that they love getting their house cleaned by us…If you did, please write ‘I loved it and recommend it’”), authoring an equally favorable review that does not employ the default language (e.g., “It was fantastic and I recommend it”) appears to be more persuasive. That deviations can foster persuasion even when they provide no hint of negative considerations further suggests that the deviation effect is unlikely to stem from perceptions of blemishing or two-sided messaging. Expectation Violations As another possibility, is the deviation effect partially driven by the fact that deviations are unexpected? After all, we propose an attributional inference process, and unexpected events are known to prompt attributional reasoning (Clary and Tesser 1983; Reich and Tormala 2013). To investigate this possibility, we assigned Mechanical Turk participants to either the deviation or no deviation condition from study 4a, and then asked them, “To what extent is this reviewer’s rating unexpected?” and “To what extent is this reviewer’s rating surprising?” Participants responded on separate scales (1: Not at all; 7: Very much). Participants perceived the reviews to be equally expected (MDeviation = 3.34, SDDeviation = 1.89; MNo Deviation = 3.18, SDNo Deviation = 1.81; t(207) = .63, p = .532) and unsurprising (MDeviation = 3.15, SDDeviation = 1.87; MNo Deviation = 3.23, SDNo Deviation = 1.91; t(207) = .30, p = .764). These results do not preclude the possibility that some types of deviations are perceived as unexpected, but the fact that the deviation effect emerged in contexts in which deviations are not unexpected suggests that an expectancy violation account is insufficient to explain our findings. Attention Related to this point, is the deviation effect partially driven by increased attention to deviatory reviews? Although this is a reasonable proposition, we suspect it is unlikely to account for our full pattern of data. First, if deviations operate by attracting attention, it seems likely that the deviation effect would be amplified among consumers devoting relatively low levels of attention and processing to a review at baseline—for example, among those under cognitive load. Indeed, past research reveals that persuasive interventions that operate by boosting processing tend to have greater impact when consumers’ baseline processing levels are low (Kupor and Tormala 2015; Priester and Petty 1995; Tormala and Clarkson 2008). In contrast, our studies reveal that the deviation effect not only fails to increase but in fact reverses under low processing conditions (study 6). This pattern is inconsistent with a pure attentional mechanism. Moreover, an attention account would not predict the results of study 5, in which review length moderated the effect of deviatory reviews. If deviations increase persuasion by increasing attention, they should do so regardless of whether they feature brief or extended discussions of the reviewed product. In short, despite offering an intuitively appealing account at first blush, it is unlikely that all of our findings can be explained by an attentional mechanism. Unseen Reviews Another alternative that is worth considering is whether the deviation effect could be due to an inference that if a moderate review contains a positive evaluation of the reviewed product, then unseen reviews that are more extremely positive may contain even more glowing evaluations. Perhaps this inference prompted speculation about what those reviews might say, which yielded a more favorable response. However, studies 4–5 are incompatible with this alternative: this account would not predict that an extreme deviatory review would be more persuasive in the context of a moderately positive default (studies 4a and 4b), nor that the deviation effect would disappear when nondeviatory reviews resulted from perceived substantial thought (study 5). Thus, the moderation patterns observed in studies 4–5 are inconsistent with this alterative account. Experimental Artifact Also relevant, our studies demonstrate that perceived thoughtfulness and accuracy need not be made artificially salient within the experimental context for the deviation effect to manifest. Although some of our studies measured these constructs prior to behavioral intentions, which could have increased their salience, not all did so. In studies 1a, 1b, 3, 5, and 6, we observed the deviation effect even when participants did not complete thoughtfulness and accuracy measures prior to reporting their behavioral intentions. Thus, consistent with prior literature (Cheung et al. 2009; Willemsen et al. 2012), participants appeared to draw spontaneous inferences about others’ thoughtfulness and accuracy when viewing others’ reviews. Of course, as with most psychological effects, we suspect the deviation effect will be stronger when its psychological drivers are particularly salient. Likewise, when another dimension is more salient, the effect might shift. For instance, because people do not anticipate the deviation effect (web appendix D), it is possible that the effect attenuates when people anticipate that they will need to justify their decisions to others. This is because people who anticipate they will need to justify their decisions to others often internalize the criteria that they believe others employ, and use those criteria to make their decisions (Huber and Seiser 2001; Simonson 1989). Thus, the deviation effect may attenuate in contexts in which people anticipate external evaluation (e.g., when superiors will evaluate their decisions; Simonson 1989). We encourage future research to investigate this possibility. Extreme Means Biased? Another alternative that is important to consider is whether the deviation effect occurs because consumers are aware that moderate reviews are less likely to be skewed by a priori biases (e.g., “brag-and-moan” biases; Hu, Zhang, and Pavlou 2009), and people seek to protect themselves from such biases by discounting extreme reviews. This is an intriguing possibility, but it is incompatible with studies 4–5. Studies 4a and 4b demonstrated that a review’s deviatory status (rather than its extremity per se) drove its persuasiveness. Study 5 showed that participants were equally persuaded by extreme and moderate reviews when those reviews were especially lengthy. Nonetheless, to the extent that extreme endorsements are more biased (Hu et al. 2009), or at least less measured, identifying means of steering consumers away from their influence could have implications for consumer welfare. We encourage future research to investigate that possibility. Boundaries and Extensions There are likely to be constraints on the deviation effect in addition to the two that we documented (i.e., review length and cognitive load). For example, the magnitude of deviation may moderate its effect (i.e., the deviation effect likely attenuates as reviews become increasingly moderate). The current studies found that 4-star deviatory reviews can be more persuasive than 5-star nondeviatory reviews (studies 4a, 5, and 6), and that 8-star (studies 1b, 2, and 4b) and 9-star (study 1a) deviatory reviews can be more persuasive than 10-star nondeviatory reviews. Is a 7-star deviatory review also more persuasive than a 10-star nondeviatory review? We submit that it could be if it is perceived to be sufficiently favorable. In an initial test of this possibility (web appendix F), we compared 7-star, 8-star, and 9-star deviatory reviews to a 10-star nondeviatory review. This study revealed that even a 7-star deviatory review was more persuasive than a 10-star nondeviatory review. Nonetheless, there is likely an extremity threshold, below which a review is no longer sufficiently favorable to outperform an extremely favorable nondeviatory review. This threshold might vary across contexts as a function of numerous factors (e.g., the perceived strength of the default), and we encourage future exploration of this issue. Related to this point, future research could profit from leveraging information integration theory to precisely estimate consumers’ judgments in the context of defaults, and to model how these judgments may shift as reviews become increasingly moderate. Information integration theory formalizes the manner in which different parameters (e.g., a review’s extremity and perceived accuracy) are integrated into judgment by assigning each parameter a unique weight (w; reflecting the parameter’s impact on judgment) and specifying an algebraic composition rule representing their integration (Anderson 1981; Peysakhovich and Karmarkar 2016). Although the primary focus of our research is to examine the psychological processes through which deviating endorsements are persuasive, our findings underscore the potential value of developing a formal model of information integration in the current context. Indeed, our studies demonstrate the manner in which people consider not only the extremity of consumer reviews when they process them, but also the accuracy of those reviews. These are dynamic parameters that could be modeled. Precisely how much weight each parameter receives, and how that weight interacts with degree of deviation and other contextual factors, is beyond the scope of the current research. Nevertheless, this is an interesting and potentially fruitful path for future work. One contextual factor that could be important to examine in future research relates to the manner in which reviews are aggregated online. For example, some review platforms rank products as a function of their average rating. In such cases, a product that has a default rating of 5 out of 5 stars would have a higher ranking if none of its ratings deviated from 5 stars compared to if some of its ratings deviated to a lower rating. On websites in which deviating reviews lower a product’s ranking, some consumers might be less likely to consider that product as a mere function of its lower ranking. If true—that is, if consumers were not even exposed to deviating reviews—the deviation effect would be unlikely to manifest. We suspect there are numerous situational and individual factors that moderate consumers’ likelihood of considering a lower-ranked option, and thus moderate whether the deviation effect emerges on these ranked review platforms. For example, highly involved consumers who are interested in learning about a wide range of alternatives should be more likely to consider lower-ranked options. Likewise, consumers who are cognitively unencumbered (e.g., who are not distracted or under time pressure) might be more inclined to consider lower-ranked options. Conversely, consumers who are uninterested or whose processing resources are constrained (e.g., who are distracted or under time pressure) might be less likely to do so. Thus, consumers’ depth of processing not only may moderate the persuasive impact of deviating reviews that consumers view (as in study 6), but also may determine whether consumers view a deviating review in the first place (which, of course, would be required to open someone up to the deviation effect). We submit that once consumers have evaluated multiple products—not just the top-ranked one(s)—they might ultimately choose a lower-ranked option if it has a deviating favorable endorsement, because the deviating endorsement appears more thoughtful and accurate. We encourage future research to examine this possibility. Also relevant to future research, the inferences that consumers draw from deviatory evaluations may have numerous downstream consequences beyond increasing persuasion. For example, inferences of thoughtfulness and accuracy can affect a variety of outcomes, including liking, certainty, and trustworthiness (Barden and Petty 2008; Conchie and Burns 2009; Kupor et al. 2014; Kupor, Flynn, and Norton 2017). Consistent with the possibility that deviations may have a wider range of consequences than those documented here, our analysis of supplemental measures suggested that deviations may affect perceived trustworthiness and knowledge (see web appendixes A and B). These measures were purely exploratory, but the fact that deviations affected responses to them suggests that they might be worth future study. We encourage research to investigate the expanse of inferences that people draw from deviations, and the consequences of these inferences. Of importance, we investigated our theorizing in the context of reviews not only because defaults are common in these contexts, but also because reviews are a primary determinant of consumers’ purchase decisions (Allsop, Bassett, and Hoskins 2007). Examining our theorizing in this context thus provides insight into a key driver of consumer decision making. It is worth noting, however, that the psychology underlying the effects we observed might apply to deviations from defaults in other settings as well. For instance, people buying a new car often face the decision of whether to alter the car’s default features, people purchasing restaurant meals often have the option to change the default side dish, and social media users can opt (or not) to change their default privacy settings. Based on the current insights, consumers who observe others deviating from such defaults might infer that these decisions are more thoughtful and accurate, and thus be more influenced by them. Future research could investigate this possibility. Strategic Defaults Finally, our research has strategic implications for marketers. For example, our findings suggest that marketers may increase sales by publicizing positive reviews that have clearly deviated from a default evaluation. Similarly, when a product receives a moderately positive review in the context of an extremely positive default, firms might enhance that review’s positive impact (e.g., on sales) by highlighting that the default is extremely positive. In other words, making the default salient could increase perceptions of a moderately positive review’s accuracy, and heighten sales as a result. In addition, if marketers are certain (through market testing, historical review data, or other means) that reviewers will rate their products highly, then implementing a default rating that differs from reviewers’ normative ratings could boost sales when consumers see that reviewers deviated from that default but still gave the product a favorable rating. In other words, consumers might be more persuaded to purchase a favorably reviewed product when those favorable reviews are given in the context of a different default, thus indicating that reviewers put thought into their favorable ratings. In a similar vein, it is possible that consumers not only perceive others as thinking more carefully when others deviate from a default, but also perceive themselves as thinking more carefully when they themselves deviate from a default. Therefore, the strategic use of defaults could boost sales not only among consumers who read reviews, but also among the reviewers themselves. People are more certain of their attitudes when they perceive that they have thought more about them (Barden and Petty 2008). Attitude certainty, in turn, increases attitudes’ persistence over time, resistance to attack, and impact on behavior (Rucker et al. 2014). Thus, if marketers’ strategic use of defaults boosts reviewers’ perceptions of their own thoughtfulness—and thus their attitude certainty—defaults may generate substantially more favorable future attitudes and behavior. DATA COLLECTION INFORMATION The first author supervised the collection of data for the first study by research assistants at Boston University in 2017–2018. The first author managed the collection of data for studies 2, 4a, 4b, and 6 using Mechanical Turk. The first author managed the collection of data for study 3 using SurveyMonkey, and for study 5 using SSI. The data were analyzed by the first author. The authors thank Sofia Nikolakaki and Ewart Thomas for their help with scraping and analyzing the secondary data. Supplementary materials are included in the web appendix accompanying the online version of this article. References Aaker Jennifer, Vohs Kathleen D., Mogilner Cassie ( 2010), “Nonprofits Are Seen as Warm and For-Profits as Competent: Firm Stereotypes Matter,” Journal of Consumer Research , 37 ( 2), 224– 37. Google Scholar CrossRef Search ADS   Active Pest Control ( 2018), Review Us, https://activepestcontrol.com/review-us/. Agnew Julie R., Szykman Lisa R. ( 2005), “Asset Allocation and Information Overload: The Influence of Information Display, Asset Choice, and Investor Experience,” Journal of Behavioral Finance , 6 ( 2), 57– 70. Google Scholar CrossRef Search ADS   Allsop Dee T., Bassett Bryce R., Hoskins James A. ( 2007), “Word-of-Mouth Research: Principles and Applications,” Journal of Advertising Research , 47 ( 4), 398– 411. Google Scholar CrossRef Search ADS   Amazon ( 2014), “Medicasp Therapeutic Shampoo,” https://www.amazon.com/Medicasp-Therapeutic-Shampoo/dp/B00BCYZOIG. Amazon ( 2015), “Buy this book!” (customer review by Sunder), https://www.amazon.com/gp/customer-reviews/R985UQZQ5FH2C/ref=cm_cr_arp_d_rvw_ttl?ie=UTF8&ASIN=B00506VMJ2. Amazon ( 2016), “Nice ribbon in a variety of colors,” https://www.amazon.com/gp/product/B0097K2JF8/ref=s9u_simh_gw_i2?ie=UTF8&fpl=fresh&pd_rd_i=B0097K2JF8&pd_rd_r=d8072b17-381f-11e8-bb6d-df8617f41b33&pd_rd_w=5Ll4c&pd_rd_wg=8wFGg&pf_rd_m=ATVPDKIKX0DER&pf_rd_s=&pf_rd_r=6ARY8T8CW7JZXZZGEH62&pf_rd_t=36701&pf_rd_p=0411ffec-c026-40ae-aac5-2cd3d48aeeac&pf_rd_i=desktop. Anderson Norman H. ( 1981), Foundation of Information Integration Theory , New York: Academic Press. ASUS ( 2015), “Status Bar Transparency Setting” (forum post from user Naresh), https://www.asus.com/zentalk/thread-2941-1-1.html. Bailey Jeffrey J., Nofsinger John R., O’Neill Michele ( 2004), “401(k) Retirement Plan Contribution Decision Factors: The Role of Social Norms,” Journal of Business and Management , 9 ( 4), 327– 44. Barden Jamie, Petty Richard E. ( 2008), “The Mere Perception of Elaboration Creates Attitude Certainty: Exploring the Thoughtfulness Heuristic,” Journal of Personality and Social Psychology , 95 ( 3), 489– 509. Google Scholar CrossRef Search ADS   Barden Jamie, Tormala Zakary L. ( 2014), “Elaboration and Attitude Strength: The New Meta‐cognitive Perspective,” Social and Personality Psychology Compass , 8 ( 1), 17– 29. Google Scholar CrossRef Search ADS   Bargh John A., Chartrand Tanya L. ( 1999), “The Unbearable Automaticity of Being,” American Psychologist , 54 ( 7), 462– 79. Google Scholar CrossRef Search ADS   Berlo David K., Lemert James, Mertz Robert ( 1969), “Dimensions for Evaluating the Acceptability of Message Sources,” Public Opinion Quarterly , 33 ( 4), 563– 76. Google Scholar CrossRef Search ADS   Brett Joan F., Atwater Leanne E. ( 2001), “360° Feedback: Accuracy, Reactions, and Perceptions of Usefulness,” Journal of Applied Psychology , 86 ( 5), 930– 42. Google Scholar CrossRef Search ADS   Briñol Pablo, Petty Richard ( 2009), “ Persuasion: Insights from the self-validation hypothesis,” Advances in Experimental Social Psychology , 41, 69– 118. Google Scholar CrossRef Search ADS   Campbell Margaret C., Kirmani Amna ( 2000), “ Consumers' use of persuasion knowledge: The effects of accessibility and cognitive capacity on perceptions of an influence agent,” Journal of Consumer Research , 27 ( 1), 69– 83. Google Scholar CrossRef Search ADS   Cress ( 2018), http://www.cressfuneralservice.com/review/. Chaiken Shelly, Eagly Alice H. ( 1989), “Heuristic and Systematic Information Processing Within and Beyond the Persuasion Context,” in Unintended Thought , ed. Uleman James S., Bargh John A., New York: Guilford, 212– 47. Chesney Thomas ( 2006), “An Empirical Examination of Wikipedia’s Credibility,” First Monday , 11 ( 11), http://firstmonday.org/issues/issue11_11/chesney/index.html. Cheung Man Yee, Luo Chuan, Sia Choon Ling, Chen Huaping ( 2009), “ Credibility of electronic word-of-mouth: Informational and normative determinants of on-line consumer recommendations,” International Journal of Electronic Commerce , 13 ( 4), 9– 38. Google Scholar CrossRef Search ADS   Chevalier Judith A., Mayzlin Dina ( 2006), “The Effect of Word of Mouth on Sales: Online Book Reviews,” Journal of Marketing Research , 43 ( 3), 345– 54. Google Scholar CrossRef Search ADS   City-Data Real Estate Forum ( 2013), http://www.city-data.com/forum/real-estate/1807095-hiring-realtor-based-zillow-reviews.html. Clary Gil, Tesser Abraham ( 1983), “Reactions to Unexpected Events,” Personality and Social Psychology Bulletin , 9 ( 4), 609– 20. Google Scholar CrossRef Search ADS   Clement Russell W., Krueger Joachim ( 2002), “Social Categorization Moderates Social Projection,” Journal of Experimental Social Psychology , 38 ( 3), 219– 31. Google Scholar CrossRef Search ADS   Conchie Stacey M., Burns Calvin ( 2009), “Improving Occupational Safety: Using a Trusted Information Source to Communicate about Risk,” Journal of Risk Research , 12 ( 1), 13– 25. Google Scholar CrossRef Search ADS   Connors Laura, M. Mudambi Susan, Schuff David ( 2011), “Is it the review or the reviewer? A multi-method approach to determine the antecedents of online review helpfulness.” System Sciences (HICSS), 2011 44th Hawaii International Conference on IEEE. Crotty David ( 2009), “How Meaningful Are User Ratings?” https://scholarlykitchen.sspnet.org/2009/11/16/how-meaningful-are-user-ratings-this-article-4-5-stars/. Davies Martin ( 1994), “Private Self-Consciousness and the Perceived Accuracy of True and False Personality Feedback,” Personality and Individual Differences , 17 ( 5), 697– 701. Google Scholar CrossRef Search ADS   Deutsch Morton, Gerard Harold B. ( 1955), “A Study of Normative and Informational Social Influences upon Individual Judgment,” Journal of Abnormal and Social Psychology , 51 ( 3), 629– 36. Google Scholar CrossRef Search ADS   Dimaggio Giancarlo, Lysaker Paul H., Carcione Antonino, Nicolò Giuseppe, Semerar Antonio ( 2008), “Know Yourself and You Shall Know the Other…to a Certain Extent: Multiple Paths of Influence of Self-Reflection on Mindreading,” Consciousness and Cognition , 17 ( 3), 778– 89. Google Scholar CrossRef Search ADS   Doh Sun-Jae, Hwang Jang-Sun ( 2009), “How Consumers Evaluate eWOM (Electronic Word-of-Mouth) Messages,” Cyberpsychology & Behavior: The Impact of the Internet, Multimedia and Virtual Reality on Behavior and Society , 12 ( 2), 193– 7. Dr. Oogle ( 2016), https://www.doctor-oogle.com/. Durham Volkswagen ( 2018), https://www.durhamvw.com/review-us/ Ein-Gar Danit, Shiv Baba, Tormala Zakary L. ( 2012), “When Blemishing Leads to Blossoming: The Positive Effect of Negative Information,” Journal of Consumer Research , 38 ( 5), 846– 59. Google Scholar CrossRef Search ADS   Eisend Martin ( 2006), “Two-Sided Advertising: A Meta-analysis,” International Journal of Research in Marketing , 23 ( 2), 187– 98. Google Scholar CrossRef Search ADS   Everett Jim A. C., Caviola Lucius, Kahane Guy, Savulescu Julian, Faber Nadira S. ( 2015), “Doing Good by Doing Nothing? The Role of Social Norms in Explaining Default Effects in Altruistic Contexts,” European Journal of Social Psychology , 45 ( 2), 230– 41. Google Scholar CrossRef Search ADS   Faraway Julian ( 2006), Extending the Linear Model with R: Generalized Linear, Mixed Effects and Nonparametric Regression Models , Boca Raton, FL: Chapman & Hall/CRC Texts in Statistical Science. Filieri Raffaele, McLeay Fraser ( 2014), “E-WOM and Accommodation: An Analysis of the Factors That Influence Travelers’ Adoption of Information from Online Reviews,” Journal of Travel Research , 53 ( 1), 44– 57. Google Scholar CrossRef Search ADS   Flanagin Andrew J., Metzger Miriam J. ( 2000), “Perceptions of Internet Information Credibility,” Journalism & Mass Communication Quarterly , 77 ( 3), 515– 40. Google Scholar CrossRef Search ADS   Forman Chris, Ghose Anindya, Wiesenfeld Batia ( 2008), “Examining the Relationship between Reviews and Sales: The Role of Reviewer Identity Disclosure in Electronic Markets,” Information Systems Research , 19 ( 3), 291– 313. Google Scholar CrossRef Search ADS   Goldman Alvin I., Jordan Lucy C. ( 2013), “Mindreading by Simulation: The Roles of Imagination and Mirroring,” in Understanding Other Minds: Perspectives from Developmental Social Neuroscience , ed. Baron-Cohen Simon, Lombardo Michael, Tager-Flusberg Helen, Oxford, UK: Oxford University Press, 448– 66. Google Scholar CrossRef Search ADS   Goldstein Noah J., Cialdini Robert B., Griskevicius Vladas ( 2008), “A Room with a Viewpoint: Using Social Norms to Motivate Environmental Conservation in Hotels,” Journal of Consumer Research , 35 ( 3), 472– 82. Google Scholar CrossRef Search ADS   Grohmann Bianca, Spangenberg Eric R., Sprott David E. ( 2007), “The Influence of Tactile Input on the Evaluation of Retail Product Offerings,” Journal of Retailing , 83 ( 2), 237– 45. Google Scholar CrossRef Search ADS   Groupon ( 2016), “$31 for $70 Worth of Services—The Laboratory Production Studios,” https://www.groupon.com/deals/the-laboratory-production-studios. Hayes Andrew F. ( 2013), Introduction to Mediation, Moderation, and Conditional Process Analysis , New York: Guilford. Hovland Carl, Janis Irving, Kelley Harold ( 1953), Communication and Persuasion , New Haven, CT: Yale University Press, 19– 48. Hu Nan, Zhang Jie, Pavlou Paul A. ( 2009), “Overcoming the J-Shaped Distribution of Product Reviews,” Communications of the Acm , 52 ( 10), 144– 7. Google Scholar CrossRef Search ADS   Huber Oswald, Seiser Gabriele ( 2001), “Accounting and Convincing: The Effect of Two Types of Justification on the Decision Process,” Journal of Behavioral Decision Making , 14 ( 1), 69– 85. Google Scholar CrossRef Search ADS   Huh Young Eun, Vosgerau Joachim, Morewedge Carey K. ( 2014), “Social Defaults: Observed Choices Become Choice Defaults,” Journal of Consumer Research , 41 ( 3), 746– 60. Google Scholar CrossRef Search ADS   Hyken Shep ( 2017), “Sixty-Four Percent of U.S. Households Have Amazon Prime,” Forbes , https://www.forbes.com/sites/shephyken/2017/06/17/sixty-four-percent-of-u-s-households-have-amazon-prime/#23a811844586/. Jin Liyin, He Yanqun, Zhang Ying ( 2014), “How Power States Influence Consumers’ Perceptions of Price Unfairness,” Journal of Consumer Research , 40 ( 5), 818– 33. Google Scholar CrossRef Search ADS   Johnson Eric J., Goldstein Daniel ( 2003), “Do Defaults Save Lives?” Science , 302 ( 5649), 1338– 9. Google Scholar CrossRef Search ADS   Kane Kat ( 2015), “The Big Hidden Problem with Uber? Insincere 5-Star Ratings,” Wired, http://www.wired.com/2015/03/bogus-uber-reviews/. Kaufman Douglas, Stasson Mark, Hart Jason ( 1999), “Are the Tabloids Always Wrong or Is That Just What We Think? Need for Cognition and Perceptions of Articles in Print Media,” Journal of Applied Social Psychology , 29 ( 9), 1984– 2000. Google Scholar CrossRef Search ADS   Khan Uzma, Kupor Daniella M. ( 2016), “Risk (Mis)Perception: When Greater Risk Reduces Risk Valuation,” Journal of Consumer Research , 43 ( 5), 769– 86. Krueger Joachim ( 2000), “The Projective Perception of the Social World: A Building Block of Social Comparison Processes,” in Handbook of Social Comparison: Theory and Research , ed. Suls Jerry, Wheeler Ladd, New York: Plenum/Kluwer, 323– 51. Google Scholar CrossRef Search ADS   Kuo Ying-Feng, Wu Chi-Ming, Deng Wei-Jaw ( 2009), “The Relationships among Service Quality, Perceived Value, Customer Satisfaction, and Post-Purchase Intention in Mobile Value-Added Services,” Computers in Human Behavior , 25 ( 4), 887– 96. Google Scholar CrossRef Search ADS   Kupor Daniella, Flynn Frank, Norton Michael ( 2017), “Half a Gift Is Not Half-Hearted: A Giver-Receiver Asymmetry in the Thoughtfulness of Partial Gifts,” Personality & Social Psychology Bulletin , 43 ( 12), 1686– 95. Google Scholar CrossRef Search ADS   Kupor Daniella M., Laurin Kristin, Levav Jonathan ( 2015), “Anticipating Divine Protection? Reminders of God Can Increase Nonmoral Risk Taking,” Psychological Science , 26 ( 4), 374– 84. Google Scholar CrossRef Search ADS   Kupor Daniella M., Tormala Zakary L. ( 2015), “Persuasion, Interrupted: The Effect of Momentary Interruptions on Message Processing and Persuasion,” Journal of Consumer Research , 42 ( 2), 300– 15. Kupor Daniella M., Tormala Zakary, Norton Michael I. ( 2014), “The Allure of Unknown Outcomes: Exploring the Role of Uncertainty in the Preference for Potential,” Journal of Experimental Social Psychology , 55 ( 6), 210– 6. Google Scholar CrossRef Search ADS   Kupor Daniella M., Tormala Zakary L., Norton Michael I., Rucker Derek D. ( 2014), “Thought Calibration: How Thinking Just the Right Amount Increases One’s Influence and Appeal,” Social Psychological and Personality Science , 5 ( 3), 263– 70. Google Scholar CrossRef Search ADS   Larson Lindsay RL, Denton L. Trey ( 2014), “ eWOM Watchdogs: Ego-Threatening Product Domains and the Policing of Positive Online Reviews,” Psychology & Marketing , 31 ( 9), 801– 11. Google Scholar CrossRef Search ADS   Lee Yi-Ching, Lee John D., Boyle Linda Ng ( 2007), “Visual Attention in Driving: The Effects of Cognitive Load and Visual Disruption,” Human Factors , 49 ( 4), 721– 33. Google Scholar CrossRef Search ADS   Lin Lin, Lee Jennifer, Robertson Tip ( 2011), “Reading While Watching Video: The Effect of Video Content on Reading Comprehension and Media Multitasking Ability,” Journal of Educational Computing Research , 45 ( 2), 183– 201. Google Scholar CrossRef Search ADS   McCroskey James ( 1966), “Scales for the Measurement of Ethos,” Speech Monographs , 33 ( 1), 65– 72. Google Scholar CrossRef Search ADS   McGinnies Elliott, Ward Charles D. ( 1980), “Better Liked than Right: Trustworthiness and Expertise as Factors in Credibility,” Personality and Social Psychology Bulletin , 6 ( 3), 467– 72. Google Scholar CrossRef Search ADS   MerchantCircle.com ( 2016), “M D Script,” http://www.merchantcircle.com/m-d-script-fall-city-wa. Miller Dale T. ( 1999), “The Norm of Self-Interest,” American Psychologist , 54 ( 12), 1053– 60. Google Scholar CrossRef Search ADS   Miller Dale T., Prentice Deborah A. ( 1996), “The Construction of Social Norms and Standards,” in Social Psychology: Handbook of Basic Principles , ed. Higgins E. Tory, Kruglanski Arie W., New York: Guilford, 799– 829. Murray Keith B. ( 1991), “A Test of Services Marketing Theory: Consumer Information Acquisition Activities,” Journal of Marketing , 55 ( 1), 10– 25. Google Scholar CrossRef Search ADS   Otterbacher Jahna ( 2008), “ Managing Information in Online Product Review Communities: Two Approaches,” ECIS  706– 717. Ong Beng Soo ( 2012), “The Perceived Influence of User Reviews in the Hospitality Industry,” Journal of Hospitality Marketing & Management , 21 ( 5), 463– 85. Google Scholar CrossRef Search ADS   Pelham Brett W., Sumarta Tin T., Myaskovsky Laura ( 1994), “The Easy Path from Many to Much: The Numerosity Heuristic,” Cognitive Psychology , 26 ( 2), 103– 33. Google Scholar CrossRef Search ADS   Petty Richard E., Cacioppo John T. ( 1984), “The Effects of Involvement on Responses to Argument Quantity and Quality: Central and Peripheral Routes to Persuasion,” Journal of Personality and Social Psychology , 46 ( 1), 69– 81. Google Scholar CrossRef Search ADS   Petty Richard E., Cacioppo John T., Schumann David ( 1983), “Central and Peripheral Routes to Advertising Effectiveness: The Moderating Role of Involvement,” Journal of Consumer Research , 10 ( 2), 135– 46. Google Scholar CrossRef Search ADS   Peysakhovich Alexander, Karmarkar Uma R. ( 2016), “Asymmetric Effects of Favorable and Unfavorable Information on Decision Making under Ambiguity,” Management Science , 62 ( 8), 2163– 78. Google Scholar CrossRef Search ADS   PrestaShop ( 2015), “Default Star Rating to 5 Instead of 3,” https://www.prestashop.com/forums/topic/449039-default-star-rating-to-5-instead-of-3/. Priester Joseph R., Petty Richard E. ( 1995), “Source Attributions and Persuasion: Perceived Honesty as a Determinant of Message Scrutiny,” Personality and Social Psychology Bulletin , 21 ( 6), 637– 54. Google Scholar CrossRef Search ADS   Priester Joseph R., Petty Richard E. ( 2003), “The Influence of Spokesperson Trustworthiness on Message Elaboration, Attitude Strength, and Advertising Effectiveness,” Journal of Consumer Psychology , 13 ( 4), 408– 21. Google Scholar CrossRef Search ADS   Pieters Rik ( 2017), “ Meaningful mediation analysis: Plausible causal inference and informative communication,” Journal of Consumer Research , 44( 3), 692– 716. Google Scholar CrossRef Search ADS   Printing Center USA ( 2018), https://www.printingcenterusa.com/ Probst C. Adam, Shaffer Victoria A., Chan Y. Raymond ( 2013), “The Effect of Defaults in an Electronic Health Record on Laboratory Test Ordering Practices for Pediatric Patients,” Health Psychology , 32 ( 9), 995– 1002. Google Scholar CrossRef Search ADS   Quora ( 2013), “Lyft (Company): How Critical Are You as a Passenger When Rating a Ridesharing Driver?” https://www.quora.com/Lyft-company-How-critical-are-you-as-a-passenger-when-rating-a-ridesharing-driver. Reich Taly, Kupor Daniella M., Smith Rosanna K. ( 2017), “Made by Mistake: When Mistakes Increase Product Preference,” Journal of Consumer Research , 44 ( 5), 1085– 1103. Reich Taly, Tormala Zakary L. ( 2013), “When Contradictions Foster Persuasion: An Attributional Perspective,” Journal of Experimental Social Psychology , 49 ( 3), 426– 39. Google Scholar CrossRef Search ADS   Reichelt Jonas, Sievert Jens, Jacob Frank ( 2014), “How Credibility Affects eWOM Reading: The Influences of Expertise, Trustworthiness, and Similarity on Utilitarian and Social Functions,” Journal of Marketing Communications , 20 ( 1–2), 65– 81. Google Scholar CrossRef Search ADS   Rosario Ana Babić, Sotgiu Francesca, De Valck Kristine, Bijmolt Tammo H. A. ( 2016), “The Effect of Electronic Word of Mouth on Sales: A Meta-Analytic Review of Platform, Product, and Metric Factors,” Journal of Marketing Research , 53 ( 3), 297– 318. Google Scholar CrossRef Search ADS   Ross Lee, Greene David, House Pamela ( 1977), “The ‘False Consensus Effect’: An Egocentric Bias in Social Perception and Attribution Processes,” Journal of Experimental Social Psychology , 13 ( 3), 279– 301. Google Scholar CrossRef Search ADS   Rucker Derek D., Petty Richard E., Briñol Pablo ( 2008), “What’s in a Frame Anyway? A Meta-Cognitive Analysis of the Impact of One versus Two Sided Message Framing on Attitude Certainty,” Journal of Consumer Psychology , 18 ( 2), 137– 49. Google Scholar CrossRef Search ADS   Rucker Derek D., Tormala Zakary L., Petty Richard E., Briñol Pablo ( 2014), “Consumer Conviction and Commitment: An Appraisal-Based Framework for Attitude Certainty,” Journal of Consumer Psychology , 24 ( 1), 119– 36. Google Scholar CrossRef Search ADS   Sakuraba Hisako ( 2012), “Do Simultaneously Presented Visual and Auditory Stimuli Attract Our Attention? The Effect of Divided Attention on Memory,” master’s thesis, Department of Psychological Science, University of Central Missouri, Warrensburg, MO 64093. Schlosser Ann E. ( 2011), “ Can including pros and cons increase the helpfulness and persuasiveness of online reviews? The interactive effects of ratings and arguments,” Journal of Consumer Psychology  21( 3), 226- 239. Google Scholar CrossRef Search ADS   Schrift Rom Y., Amar Moty ( 2015), “Pain and Preferences: Observed Decisional Conflict and the Convergence of Preferences,” Journal of Consumer Research , 42 ( 4), 515– 34. Siegler M. G. ( 2009), “YouTube Comes to a 5-Star Realization: Its Ratings Are Useless,” TechCrunch, https://techcrunch.com/2009/09/22/youtube-comes-to-a-5-star-realization-its-ratings-are-useless/. Simonson Itamar ( 1989), “Choice Based on Reasons: The Case of Attraction and Compromise Effects,” Journal of Consumer Research , 16 ( 2), 158– 74. Google Scholar CrossRef Search ADS   Stephen Andrew T., Grewal Lauren ( 2015), “In Mobile We Trust: How Mobile Reviews Can Overcome Consumer Distrust of User-Generated Reviews,” in NA-Advances in Consumer Research , Vol. 43, ed. Diehl Kristin, Yoon Carolyn, Duluth, MN: Association for Consumer Research, 117– 21. Tormala Zakary L., Clarkson Joshua ( 2008), “Source Trustworthiness and Information Processing in Multiple Message Situations: A Contextual Analysis,” Social Cognition , 26 ( 3), 357– 67. Google Scholar CrossRef Search ADS   Tormala Zakary L., Petty Richard E., Briñol Pablo ( 2002), “Ease of Retrieval Effects in Persuasion: A Self-Validation Analysis,” Personality and Social Psychology Bulletin , 28 ( 12), 1700– 12. Google Scholar CrossRef Search ADS   Wangenheim Florian, Bayón Tomás ( 2004), “The Effect of Word of Mouth on Service Switching,” European Journal of Marketing , 38 ( 9/10), 1173– 85. Google Scholar CrossRef Search ADS   Willemsen Lotte M., Neijens Peter C., Bronner Fred ( 2012), “ The ironic effect of source identification on the perceived credibility of online product reviewers,” Journal of Computer-Mediated Communication , 18( 1), 16– 31. Google Scholar CrossRef Search ADS   Wood Stacy, McInnes Melayne Morgan, Norton David A. ( 2011), “The Bad Thing about Good Games: The Relationship between Close Sporting Events and Game-Day Traffic Fatalities,” Journal of Consumer Research , 38 ( 4), 611– 21. Google Scholar CrossRef Search ADS   Yelp ( 2015), “PeakFit Strength and Conditioning,” https://www.yelp.com/biz/peakfit-strength-and-conditioning-san-dimas-2. Zhang Yan, Epley Nicholas ( 2012), “Exaggerated, Mispredicted, and Misplaced: When ‘It’s the Thought That Counts’ in Gift Exchanges,” Journal of Experimental Psychology: General , 141 ( 4), 667– 81. Google Scholar CrossRef Search ADS   Zillow Advice Thread ( 2016), https://webcache.googleusercontent.com/search?q=cache:NF-gzkL-6jAJ:https://www.zillow.com/advice/PA/pro-to-pro-agents/question-discussion-guide/+&cd=1&hl=en&ct=clnk&gl=us. © The Author(s) 2018. Published by Oxford University Press on behalf of Journal of Consumer Research, Inc. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Consumer Research Oxford University Press

When Moderation Fosters Persuasion: The Persuasive Power of Deviatory Reviews

Loading next page...
 
/lp/ou_press/when-moderation-fosters-persuasion-the-persuasive-power-of-deviatory-8RIojublud
Publisher
University of Chicago Press
Copyright
© The Author(s) 2018. Published by Oxford University Press on behalf of Journal of Consumer Research, Inc. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com
ISSN
0093-5301
eISSN
1537-5277
D.O.I.
10.1093/jcr/ucy021
Publisher site
See Article on Publisher Site

Abstract

Abstract When people seek to persuade others to purchase a particular product or service, they often give an extremely favorable review of it as a means of doing so. Despite the intuitive appeal of this strategy, the current research demonstrates that a moderately positive review is sometimes more persuasive. In particular, when the perceived default evaluation in a given context is extremely positive, moderately positive reviews that deviate from that default can become more persuasive. In contrast, when the perceived default is moderately positive, extremely positive reviews tend to be more persuasive. This deviation effect occurs because reviews that deviate from the perceived default are believed to be more thoughtful, and thus accurate, which enhances their persuasive impact. This effect is demonstrated in eight experiments set in a diverse range of consumer contexts. persuasion, perceived thoughtfulness, defaults, norms, customer reviews Which reviewer would lead consumers to be more likely to try a new restaurant: one who notes that she tried the restaurant, thought it was extremely good, and rates it 5 out of 5 stars; or one who notes that she tried the restaurant, thought it was good, and rates it 4 out of 5 stars? It seems reasonable to predict that the former review would be more persuasive—that is, it would lead consumers to be more likely to try the restaurant. After all, the former review seems to suggest that the restaurant is better than does the latter review and, understandably, consumers tend to prefer really good options to pretty good options (Kuo, Wu, and Deng 2009). Consistent with this general intuition, when people aim to persuade others to purchase a product or service, they often emphasize their extremely favorable endorsement of it as a means of doing so. Consumers who want to convince others to purchase a product they enjoy, for instance, frequently give that product the highest possible rating online. Indeed, ecommerce websites such as Amazon.com are filled with examples of customers urging their fellow shoppers to purchase specific products by rating those products with 5 stars and giving them extremely positive reviews (for just a few examples, see Amazon 2014, 2015, 2016). Similarly, companies seeking to increase sales routinely inform people that their customers view them extremely favorably—for example, that they give them the highest possible ratings online (Groupon 2016; Printing Center USA 2018; Yelp 2015). The notion that extremely positive reviews will be more persuasive—for example, more likely to prompt customer patronage—than moderately positive reviews has considerable face validity. And indeed, we predict that this is often the case. Importantly, however, we theorize that the reverse can also regularly occur. That is, we predict that under specifiable conditions a moderately positive review of a product, service, or company can be more persuasive than an extremely positive review. In particular, we posit that when the default review in a given context is perceived to be extremely positive, moderately positive reviews that deviate from that default might become more persuasive. In essence, we theorize that consumers perceive deviatory reviews—which we define as reviews that deviate from a default—to be more accurate. As a result, such reviews can be more effective in convincing others. If true, this logic implies that a moderately positive review should be more persuasive than an extremely positive review when extremely positive reviews are the default. Conversely, an extremely positive review should be more persuasive than a moderately positive review when moderately positive reviews are the default. We unpack these predictions below. THEORETICAL BACKGROUND Defaults are options that consumers perceive to be the status quo and, thus, that they consider first before considering other options (Huh, Vosgerau, and Morewedge 2014). Defaults are determined by factors both endogenous and exogenous to a decision context. For instance, preselected options are endogenous factors that create defaults (Everett et al. 2015; Johnson and Goldstein 2003; Probst, Shaffer, and Chan 2013). Descriptive norms (which describe how most people generally behave; Deutsch and Gerard 1955; Goldstein, Cialdini, and Griskevicius 2008; Miller 1999) are exogenous factors that create defaults (Miller and Prentice 1996). Importantly, defaults of both endogenous and exogenous origin can powerfully drive behavior. For example, people are more likely to reuse hotel towels when they learn that most others reuse theirs (Goldstein et al. 2008), and people are more likely to contribute to a 401(k) plan when they learn that most others contribute to a 401(k) plan (Bailey, Nofsinger, and O’Neill 2004). Similarly, medical providers are more likely to order laboratory tests for their patients when those tests are preselected (Probst et al. 2013), Mechanical Turk workers are more likely to donate bonus money to charity (vs. keep it for themselves) when the option to donate is preselected (Everett et al. 2015), and people are more likely to become organ donors when organ donation is set as a default on registration forms (Johnson and Goldstein 2003). People are particularly likely to accept, or adhere to, a default when their motivation or ability to think deeply is limited in some way. Put differently, people are especially unlikely to deviate from a default if they have not thoroughly deliberated about their choice. For example, consumers under cognitive load or time pressure—who do not have the cognitive resources or time to deliberate—are less likely to deviate from a default than are consumers who are cognitively unencumbered (Huh et al. 2014). Similarly, consumers who are overloaded with options, and who self-report that the large number of options is hindering their ability to thoroughly deliberate about their choice, are less likely to deviate from a default (Agnew and Szykman 2005). We propose that just as people are more likely to deviate from a default when they have carefully deliberated about their choice, they might believe that others are more likely to deviate from a default when those others have carefully deliberated about their choice. Consistent with this general possibility, past research reveals that people tend to believe that others think, feel, and act as they do (Clement and Krueger 2002; Krueger 2000; Ross, Greene, and House 1977). Simulation processes appear to contribute to these effects, such that people predict and interpret others’ thoughts and behaviors by reflecting on how they themselves think and behave and then employing these metacognitions to simulate others’ thoughts, feelings, experiences, and behaviors (Dimaggio et al. 2008). Following this general logic, we propose that because people are more likely to deviate from a default when they have carefully deliberated about their own evaluations and decisions, they may believe that the same is true of others. In other words, people may perceive that others’ evaluations are more thoughtful when those evaluations deviate from a default. Furthermore, because people tend to believe that increased thought, or deliberation, leads to more accurate assessments (Barden and Petty 2008; Barden and Tormala 2014; Reich and Tormala 2013), they may infer that a person who deviates from a default has a more accurate assessment. And, importantly, substantial research reveals that people are more persuaded by assessments that they perceive to be more accurate (because people themselves seek to hold accurate opinions; Chaiken and Eagly 1989; Filieri and McLeay 2014; Priester and Petty 1995; Stephen and Grewal 2015). Thus, if people perceive that deviating from a default signals that a deviating review is more accurate, people may be more persuaded by deviating (vs. nondeviating) reviews. Of course, this analysis rests on the assumption that consumers think about the accuracy of reviews. Past research is consistent with this idea. Substantial literature suggests that people are often skeptical of the accuracy of reviews, and that as a result they spontaneously evaluate the likelihood that a review is accurate or inaccurate before deciding to rely upon it (Cheung et al. 2009; Larson and Denton 2014; Schlosser 2011; Willemsen, Neijens, and Bronner 2012). Thus, accuracy is an important dimension on which consumers evaluate reviews. Importantly, though, this evaluation process is inherently subjective. We postulate that consumers perceive deviatory reviews to be more accurate, but not that deviatory reviews actually achieve greater objective accuracy. A rich literature suggests that people are more influenced by reviews that they subjectively assess to be more accurate (Connors, Mudambi, and Schuff 2011; Filieri and McLeay 2014; Cheung et al. 2009). In short, we hypothesize that people perceive a reviewer’s evaluation of a product or service to be more accurate, and thus are more persuaded by it, when the reviewer’s evaluation deviates from rather than adheres to a perceived default. One consequence of this effect is that when the default evaluation is believed to be extremely positive, a moderately positive review might be more persuasive. Indeed, evaluative defaults often emerge in the context of reviews, and frequently they are extremely positive. For example, numerous review portals feature the highest possible default star ratings when consumers enter reviews (Active Pest Control 2018; Cress Circle 2018; Durham Volkswagen 2018; Future Marketing Partners 2016; PrestaShop 2015). Consumers who review Durham Volkswagen, for instance, are directed to a web page featuring a preselected 5-star rating. Consumers can submit this preselected rating or toggle to a lower rating before submitting it. Similarly, companies often communicate a descriptive norm to potential reviewers by noting that the majority of their customers give them the highest possible rating (Groupon 2016; Printing Center USA 2018; and Yelp 2015), which may also create an evaluative default (Miller and Prentice 1996). Consumers also form impressions about the normative review expressed in various review contexts. For example, consumers perceive that the normative review is extremely positive on Zillow, YouTube, Uber, and Lyft (City-Data Real Estate Forum 2013; Crotty 2009; Kane 2015; Quora 2013; Siegler 2009; Zillow Advice Thread 2012, 2016). In each of these contexts, we predict that reviews will be more persuasive if they are moderate rather than extreme, because the deviatory status of moderate reviews makes them seem more thoughtful and accurate. We conducted an exploratory test of this possibility by analyzing real consumer review data scraped from an online retailer that sells home goods and accessories. Like the consumer review platforms described above, the vast majority (72.7%) of reviews on this platform feature a rating of 5 out of 5 stars. Conveniently, the platform also allows consumers to indicate if they find a particular review to be helpful. Prior theorizing suggests that helpful votes are a reasonable proxy of persuasion because consumers likely rate reviews as helpful when those reviews are more useful or impactful to their own decisions (Otterbacher 2008). The data included 60,358 reviews. Because helpful votes provide count data, we analyzed them with a Poisson regression (i.e., a generalized logistic regression employed to analyze count data; Faraway 2006; Wood, McInnes, and Norton 2011). First, we conducted a Poisson regression with dummy coding (nondeviatory 5-star reviews were coded as a 1 and deviatory non-5-star reviews were coded as a 0) in which reviews’ deviatory status was entered as a predictor and the number of helpful votes that each review accrued was entered as the dependent variable. This analysis revealed that consumers rated deviatory reviews as more helpful than nondeviatory reviews, b = –.40, z = –32.79, p < .001 (table 1a). Further analysis revealed that this effect persisted when we controlled for each of the scraped covariates. Specifically, we conducted a Poisson generalized linear mixed-model analysis controlling for each review’s year, each review’s word count, and each reviewed product’s price; nesting each review within the reviewed product and the reviewed product's category; and including random effects for product and product category. This analysis again revealed that consumers rated deviatory reviews as more helpful than nondeviatory reviews, b = –.46, z = –6.77, p < .001 (table 1b). Table 1A ANALYSIS OF ALL REVIEWS   b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001    b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001  Table 1A ANALYSIS OF ALL REVIEWS   b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001    b  SE  z  p-value  Intercept  –0.46  0.01  –47.52  p < .001  Review star rating  –0.40  0.01  –32.79  p < .001  Next, we explored whether this pattern persisted when we specifically compared deviatory 4-star reviews to nondeviatory 5-star reviews. This exploratory analysis suggested that consumers rated 4-star reviews as more helpful than 5-star reviews, both when we did not control for the covariates (b = –.07, z = –4.38, p < .001; table 1c) and when we did (b = –.16, z = –2.31, p < .05; table 1d). Although we hesitate to draw definitive conclusions given the uncontrolled nature of these field data, this pattern provides initial evidence consistent with our theorizing. We conducted more direct tests of our theorizing in eight controlled experiments. RESEARCH OVERVIEW In sum, we theorize that people believe that deviatory (vs. nondeviatory) reviews are more accurate, and that people can be more persuaded by deviatory reviews as a result. We test this hypothesis across eight controlled experiments set in the context of both moderate and extreme defaults. First, studies 1–2 examine these predictions in three different product domains. These studies find that when the default review is extremely positive, people are more persuaded by moderately (vs. extremely) positive reviews. Study 2 further suggests that deviatory reviews are more persuasive because people perceive them to be more accurate. Study 3 replicates this effect in the negative domain. Here, against the backdrop of an extremely negative default, consumers are more persuaded to not purchase a product when they view a moderately rather than extremely negative review. Study 3 also measures perceived thoughtfulness and provides evidence consistent with our prediction that people believe that reviewers who author deviatory (vs. nondeviatory) reviews devoted more thought to their review, and this perception of thoughtfulness leads to greater perceived accuracy. Studies 1–3 document cases in which a moderate review becomes more persuasive when it deviates from an extreme default. Studies 4a and 4b reveal that when the default evaluation is moderate, the opposite occurs—that is, people are more persuaded by deviatory extreme (vs. nondeviatory moderate) reviews. Study 5 employs a moderation design to provide converging evidence for the underlying role of perceived thoughtfulness in driving the deviation effect. Study 6 establishes the deliberative nature of the inferential process underlying the deviation effect. STUDY 1A Study 1a used incentive-compatible decisions to test the hypothesis that people can be more persuaded by deviatory reviews than nondeviatory reviews. Participants read a review for a café on a consumer review website. Typical of the experience consumers undergo when they search reviews online, participants viewed a web page containing a list of the names of nearby cafés, their addresses, and their average ratings. To establish an extreme default for all participants, each of the cafés listed received an average rating of 10 out of 10 stars. After participants viewed this web page, they read a single review for one of the cafés that gave the café either 9 stars or 10 stars. We predicted that participants would be more persuaded to purchase coffee from this café when they viewed a 9-star review (i.e., a review that deviated from the 10-star default) versus a 10-star review (i.e., a review that did not deviate from the 10-star default). Importantly, the stimuli in study 1a were designed to create an externally valid operationalization of the default. Indeed, the normative approach we adopted to establish a default parallels some of the common ways defaults are established in ordinary customer rating contexts. First, every day, millions of consumers visit websites that present lists of products and their average ratings when shoppers search for relevant items online (e.g., Amazon.com, UncommonGoods.com, Yelp.com; Hyken 2017). Second, the vast majority of reviews on numerous websites have the highest possible rating. As a result, there is a widespread belief among consumers that customer reviews in many contexts tend to be extremely positive. For example, consumers believe this is true of Zillow, YouTube, Uber, and Lyft (City-Data Real Estate Forum 2013; Crotty 2009; Kane 2015; Quora 2013; Siegler 2009; Zillow Advice Thread 2012, 2016). Because descriptive norms can create defaults (Miller and Prentice 1996), and these norms are common in everyday consumer experience, we used this approach to establish an extreme default in this initial lab experiment. Method One hundred fourteen undergraduates from an East Coast university participated in a laboratory study for course credit. This study was run as part of a session containing unrelated surveys from different researchers. In this study, all participants read that “The Steam Room” is a café that recently opened near campus (it was in fact fictional and created for the purpose of the study), and they viewed an ostensibly real website featuring customer reviews of The Steam Room as well as several other local cafés. To establish an extreme default, we presented the first page of search results for nearby cafés, and these results revealed that each of the cafés had an average rating of 10 out of 10 stars. After viewing the entire page of results, participants read a single review of The Steam Room. For all participants, this review read: “The coffee is good & the ambiance is cool!” Across conditions, however, we manipulated the extremity of this review by varying whether it was accompanied by a 9-star or a 10-star rating. Following the review, participants indicated how likely they would be to try coffee from The Steam Room on a seven-point scale (1: Not likely at all; 7: Very likely; adapted from Aaker, Vohs, and Mogilner 2010; Jin, He, and Zhang 2014; Kupor, Laurin, and Levav 2015; Reich, Kupor, and Smith 2017). Furthermore, they read that as additional compensation for their participation, they were eligible to receive a 30% discount on up to five cups of coffee from The Steam Room if they prepaid for the coffee now—that is, during the lab session. Participants were asked how many cups of coffee they wanted to prepurchase. The instructions emphasized that this was a real decision, and that at the end of the lab session they would actually prepay for the cups of coffee that they opted to purchase. Participants entered their purchase decision (0 to 5) into an empty field. Although The Steam Room was fictional, we instructed participants that it was real, and that their purchase decision was real, in order to capture actual purchase decisions. To assess whether participants were suspicious of the experimental instructions, after participants entered their choice we asked them whether they had any comments about the survey. Participants entered their responses into an empty field. Following this question, participants were debriefed; most importantly, they learned that The Steam Room was not a real café and that they would not pay for any coffee they opted to purchase. Only two participants (1.8% of the sample) indicated that they were suspicious about whether the café or the purchase decision was real. These participants were retained in the analysis, but excluding them did not alter the significance of the results. Results and Discussion Participants indicated that they were more likely to try coffee from The Steam Room when the reviewer gave it 9 stars (i.e., when the review deviated from the 10-star default; M = 4.58, SD = 1.13) rather than 10 stars (i.e., when the review did not deviate from the default; M = 3.95, SD = 1.33), t(112) = 2.67, p = .009 (Cohen’s d = .51). Participants also purchased more coffee when the reviewer gave 9 stars (M = 3.15, SD = 1.95) rather than 10 stars (M = 2.26, SD = 2.06), t(112) = 2.35, p = .020 (Cohen’s d = .45). In short, the more moderate positive review prompted more favorable behavioral intentions and more purchasing than did the more extreme positive review. Because we established an extremely positive default in this study, this finding provides initial evidence for the notion that deviatory reviews can prompt more purchases than nondeviatory reviews. STUDY 1B The primary aim of study 1b was to replicate the findings of study 1a despite numerous procedural changes. Most importantly, we employed a different operationalization of the default. Numerous websites feature preselected star ratings when consumers enter reviews—for instance, 5 stars may be selected by default, but the consumer can toggle it to a lower rating (Done Right 2016; Dr. Oogle 2016; Future Marketing Partners 2016; PrestaShop 2015; Durham Volkswagen 2018). In study 1b, we leveraged this preselected rating format. We informed participants that a particular review platform automatically populated a default star rating of 10 out of 10 stars when consumers entered reviews. We predicted that participants would be more persuaded by reviews that deviated from, rather than adhered to, that default rating. We made several other minor methodological changes. First, we sought to assess whether a more moderate rating than the one employed in study 1a could still outperform an extreme rating. Therefore, in this study, we manipulated whether a reviewer gave an 8-star or 10-star rating. In addition, we used a real product instead of coffee from a fictional café, and rather than asking participants how many units of that product they would purchase we asked them to choose between the focal product and a 50-cent bonus. As in study 1a, this was designed to be an incentive-compatible decision. Method One hundred thirty-four undergraduates from an East Coast university participated in a laboratory study for course credit. This study was run as part of a session containing unrelated surveys from different researchers. In this study, all participants read about FoodNow, an online grocery shopping service that allows customers to review the groceries that they purchase on the FoodNow website. Participants further read that when customers complete their review, the website automatically enters a default rating of 10 out of 10 stars. Next, participants read that Sunbelt Bakery Apple Cinnamon granola bars (a real product) had recently been added to the FoodNow website, and that the most recently submitted review for these granola bars was submitted 20 seconds ago by a customer named Mark. Participants in the extreme (vs. moderate) review condition read that Mark rated the granola bars 10 (vs. 8) stars. Next, participants indicated how likely they would be to try Sunbelt Bakery Apple Cinnamon granola bars on a seven-point scale (1: Not likely at all; 7: Very likely). Furthermore, they read that as additional compensation for their participation, they were eligible to receive either 50 cents or a Sunbelt Bakery Apple Cinnamon granola bar. The instructions emphasized that this was a real decision and that at the end of the lab session they would actually be entered into a lottery to receive the option that they chose. Participants indicated their choice by selecting a radio button labeled either “50 cents” or “A Sunbelt Bakery Apple Cinnamon granola bar.” Finally, all participants were debriefed. They learned that Sunbelt Bakery granola bars were real, but that the information they received about the FoodNow website and the review was not, and that all participants would therefore receive an additional 50 cents regardless of their choice. Results and Discussion Participants indicated that they were more likely to try the granola bar when the reviewer gave it 8 stars (i.e., when the review deviated from the 10-star default; M = 4.29, SD = 1.51) rather than 10 stars (i.e., when the review did not deviate from the default; M = 3.48, SD = 1.50), t(132) = 3.13, p = .002 (Cohen’s d = .55). Participants also chose the granola bar over the monetary compensation more frequently when the reviewer gave 8 stars (73.8%) rather than 10 stars (55.1%), χ2 = 5.13, p = .023 (Cohen’s d = .40). In sum, despite numerous procedural changes from study 1a, this study provided converging evidence for the notion that deviatory reviews can be more persuasive than nondeviatory reviews. In study 2, we examined whether this occurs because deviatory reviews are perceived to be more accurate. STUDY 2 Study 2 had multiple objectives. First, we tested the hypothesis that deviatory reviews can be more persuasive because people perceive them to be more accurate. In addition, we examined whether the effect would generalize to a different participant population. To this end, we conducted the study online (using Amazon’s Mechanical Turk) rather than with undergraduates in the lab. Finally, to lend further external validity to our studies, we tested the deviation effect in a context in which the default is implicit rather than experimentally established. To accomplish this goal, we moved the study into the context of ridesharing services. Many consumers are aware that the majority of ridesharing reviews contain the highest possible rating (Kane 2015; Quora 2013). Indeed, when we asked 152 Mechanical Turk participants which star rating consumers most often give ridesharing drivers, the majority (60.5%) indicated that the highest possible rating was most frequent (13.8% indicated that a rating lower than the highest possible rating was most frequent and 25.7% indicated that they did not know). We predicted that such widespread awareness of default ratings in this context would be sufficient to produce the deviation effect, even without our establishing the default explicitly in the context of the experiment. Method One hundred ninety-nine Mechanical Turk participants completed an online study for payment. All participants imagined that there was a new ridesharing service, and that they were looking at a consumer review of it on a website with reviews of different ridesharing services. Participants received no explicit information about the distribution of review ratings on this website; they simply read about an individual review of the new ridesharing service. Participants in the extreme review condition read that the reviewer gave the company 10 out of 10 stars, whereas participants in the moderate review condition read that the reviewer gave the company 8 out of 10 stars. Next, all participants indicated the extent to which they perceived that the reviewer’s opinion about the ridesharing service was accurate. Participants responded on a seven-point scale (1: Not at all; 7: Very much; adapted from Brett and Atwater 2001; Davies 1994; Grohmann, Spangenberg, and Sprott 2007). We also assessed participants' behavioral intentions. Results for exploratory items measured in this study and subsequent studies are included in web appendixes A and B. Results and Discussion Participants perceived the review to be more accurate in the moderate (M = 5.09, SD = .96) rather than extreme (M = 4.63, SD = 1.26) review condition, t(197) = 2.88, p = .004 (Cohen’s d = .41). Participants also indicated that they were more likely to use the ridesharing company in the moderate (M = 5.16, SD = 1.11) rather than extreme (M = 4.69, SD = 1.35) review condition, t(197) = 2.65, p = .009 (Cohen’s d = .38). A mediation analysis with bootstrapping was consistent with the notion that perceived accuracy mediated the behavioral intention effect (95% CI: .1056 to .6003; figure 1). FIGURE 1 View largeDownload slide MEDIATION OF THE BEHAVIORAL INTENTIONS EFFECT THROUGH PERCEIVED ACCURACY IN STUDY 2 NOTE.—The path coefficients are unstandardized betas. The value in parentheses indicates the effect of condition on the dependent variable after we control for the mediator. *p < .05, **p < .01, ***p < .001. FIGURE 1 View largeDownload slide MEDIATION OF THE BEHAVIORAL INTENTIONS EFFECT THROUGH PERCEIVED ACCURACY IN STUDY 2 NOTE.—The path coefficients are unstandardized betas. The value in parentheses indicates the effect of condition on the dependent variable after we control for the mediator. *p < .05, **p < .01, ***p < .001. In short, the deviation effect—whereby a moderate endorsement can be more persuasive than an extreme endorsement when extreme is the default—appears to be mediated at least in part by perceptions that the moderate endorsement is more accurate. In addition to providing initial process evidence, study 2 speaks to the robustness of the effect. We replicated the core finding of studies 1a and 1b in a different participant population and in a different consumer context in which people implicitly know the default without researchers referencing or displaying it in the experimental context. STUDY 3 Studies 1 and 2 explored the consequences of deviating from an extremely positive default to a moderately positive evaluation. Study 3 further tested the robustness of the deviation effect by examining whether it extends to the negative domain—for example, a deviation from an extremely negative default to a moderately negative evaluation. Participants in study 3 viewed a 1-star or 2-star review on a website that had a 1-star default. We surmised that participants would be less likely to purchase the reviewed product after viewing a 2-star (deviatory) review rather than a 1-star (nondeviatory) review. In other words, we predicted that participants would be more persuaded that the product was of poor quality after reading a moderately (vs. extremely) negative review of it. In addition to testing the deviation effect in a negative context, study 3 offered an initial test of our hypothesis that people perceive deviatory reviews to be more accurate because they believe reviewers who author deviatory (vs. nondeviatory) reviews devoted more thought to their review. As described earlier, we theorize that it is this perception of increased thoughtfulness that leads people to perceive these reviewers’ evaluations to be more accurate and persuasive. Against the backdrop of an extremely negative default, then, we hypothesized that a moderately (vs. extremely) negative review would seem more thoughtful and thus accurate, which would increase its negative impact. To assess this process, we measured perceived thoughtfulness and accuracy in study 3. Importantly, we measured these variables after behavioral intentions. It could be argued that in study 2 we made the hypothesized process (through accuracy) especially salient by measuring it immediately before behavioral intentions. Studies 1a and 1b did not measure process at all, so it is unlikely that this measure is required for the deviation effect to obtain. Nevertheless, to ensure that the deviation effect is not contingent upon measuring perceptions of accuracy immediately before assessing consumers’ behavior, we reversed the order of these measures in study 3. Finally, study 3 further tested the generalizability of the deviation effect by examining it in a different participant pool. Whereas the previous studies documented the deviation effect in samples of university students (studies 1a and 1b) and Mechanical Turk workers (study 2), study 3 examined whether it extends to participants recruited through SurveyMonkey’s online survey platform. Method One hundred ninety-four participants were recruited from SurveyMonkey to take part in an online study for a donation to a charity. All participants read about a restaurant review website that automatically enters a 1-star rating that reviewers can change when they review a restaurant. Participants in the extreme (vs. moderate) review condition read that they see a restaurant review on this website, and the reviewer rated the restaurant 1 (vs. 2) stars. We assessed participants’ interest in trying the restaurant using the measure employed in study 1 (adapted to refer to the restaurant). On the next survey screen, participants evaluated the review’s accuracy via the measure described in study 2 (adapted to refer to the restaurant), as well as how much thought they perceived that the reviewer put into his assessment of the restaurant. Participants responded on a seven-point scale (1: Not much at all; 7: A lot; adapted from Kupor et al. 2014; Schrift and Amar 2015; Zhang and Epley 2012). Results and Discussion As predicted, participants were less interested in trying the restaurant when the review was moderately (M = 2.48; SD = 1.42) rather than extremely (M = 3.03; SD = 1.55; t(192) = 2.58, p = .011; Cohen’s d = .37) negative. Also, participants perceived the moderate (vs. extreme) review to be more thoughtful (MModerate = 4.43, SDModerate = 1.55; MExtreme = 3.08, SDExtreme = 1.70; t(192) = 5.76, p < .001; Cohen’s d = .83) and accurate (MModerate = 4.46, SDModerate =1.26; MExtreme = 3.26, SDExtreme = 1.51; t(192) = 6.03, p < .001; Cohen’s d = .87). We predicted that participants would be less likely to try the restaurant when they viewed the moderately negative review because its deviation from the extremely negative default signaled that the reviewer was more thoughtful in his assessment, which in turn fostered the perception that this assessment was more accurate. A serial mediation with bootstrapping (Hayes 2013) was consistent with this account (95% CI: –.4517, –.0036; figure 2). In short, we replicated the deviation effect and obtained evidence consistent with its proposed mechanism despite numerous changes to the experimental paradigm—specifically, testing it in the context of an extremely negative default, using a different participant population, and reversing the order of the process and outcome measures. FIGURE 2 View largeDownload slide MEDIATION MODELS IN STUDY 3 (FIRST PANEL), STUDY 4A (SECOND PANEL), STUDY 4B (THIRD PANEL), AND STUDY 5 (FOURTH PANEL) NOTE.—The path coefficients are unstandardized betas. Values in parentheses indicate the effect of the interaction on the dependent variable after we control for the mediators. *p < .05, **p < .01, ***p < .001, ^p < .10. FIGURE 2 View largeDownload slide MEDIATION MODELS IN STUDY 3 (FIRST PANEL), STUDY 4A (SECOND PANEL), STUDY 4B (THIRD PANEL), AND STUDY 5 (FOURTH PANEL) NOTE.—The path coefficients are unstandardized betas. Values in parentheses indicate the effect of the interaction on the dependent variable after we control for the mediators. *p < .05, **p < .01, ***p < .001, ^p < .10. STUDY 4A The first four studies provided initial evidence for the deviation effect: people believe reviews that deviate from a default evaluation are more thoughtful and accurate, and thus are more persuaded by them. However, in these studies the evidence was restricted to cases in which a moderate review deviated from an extreme default. In study 4a, we varied whether a reviewer expressed a moderately or extremely positive review in addition to whether the default was moderately or extremely positive. We hypothesize that people perceive a moderately positive review to be more thoughtful and accurate when the default is extremely positive, but that they perceive an extremely positive review to be more thoughtful and accurate when the default is moderately positive. In other words, we predict that people perceive deviatory reviews to be more thoughtful and accurate—and are more persuaded by them—regardless of whether the deviatory review is more moderate or extreme than the default. Study 4a tested this prediction in the context of restaurant reviews and a 5-star rating scale. Method Four hundred Mechanical Turk participants completed an online study for payment. Participants were randomly assigned to conditions in a 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) between-participants design. Participants in the extreme (vs. moderate) default condition read that when reviewers enter a review on a particular review website, the website automatically populates a default rating of 5 (vs. 4) stars that reviewers can change. Next, participants in the extreme (vs. moderate) review condition read that they see a review of a restaurant on this website, and the reviewer rated it 5 (vs. 4) stars. After participants read this information, they indicated their perceptions of the reviewer’s thoughtfulness and accuracy, as well as their own likelihood of trying the restaurant, using the measures employed in study 3. In this study and study 4b, we measured perceived thoughtfulness and accuracy before behavioral intentions. Results Perceived Thoughtfulness We began by submitting perceived thoughtfulness to a 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) ANOVA. This analysis uncovered a main effect of default, F(1, 396) = 5.64, p = .018, but not of the reviewer’s evaluation, F(1, 396) = 1.66, p = .199. Most importantly, there was a significant interaction, F(1, 396) = 90.41, p < .001 (table 2). When there was an extreme default, participants inferred that the reviewer devoted more thought to his evaluation when his review was moderate rather than extreme, F(1, 396) = 59.20, p < .001 (Cohen’s d = 1.08). In contrast, when there was a moderate default, participants inferred the reviewer had thought more about his assessment when his review was extreme rather than moderate, F(1, 396) = 33.27, p < .001 (Cohen’s d = .82). TABLE 1B ANALYSIS OF ALL REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      TABLE 1B ANALYSIS OF ALL REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      Fixed effects  b  SE  z  p-value  Intercept  203.00  12.30  16.54  p < .001  Price  0.00  0.00  –3.14  p < .01  Review word count  0.01  0.00  67.56  p < .001  Review year  –0.10  0.01  –16.61  p < .001  Review star rating  –0.46  0.07  –6.77  p < .001  Random effects  Variance  SD      Product category  0.13  0.36      Product  0.95  0.97      Perceived Accuracy Analysis of the perceived accuracy data revealed a main effect of the reviewer’s evaluation, F(1, 396) = 7.29, p = .007, but not the default, F(1, 396) = 2.05, p = .153. Most relevant to our theorizing, we found the predicted interaction, F(1, 396) = 87.61, p < .001 (table 2). When there was an extreme default, participants thought that the reviewer’s evaluation was more accurate when his review was moderate rather than extreme, F(1, 396) = 73.88, p < .001 (Cohen’s d = 1.22). In contrast, when the default was moderate, participants thought that the reviewer’s opinion was more accurate when his review was extreme rather than moderate, F(1, 396) = 21.84, p < .001 (Cohen’s d = .66). Behavioral Intentions The behavioral intentions data revealed no main effect of default, F(1, 396) = 2.09, p = .149, or the reviewer’s evaluation, F(1, 396) = .06, p = .805, but we again observed the predicted interaction, F(1, 396) = 58.75, p < .001 (table 2). When there was an extreme default, participants were more likely to try the restaurant when the reviewer’s evaluation was moderate rather than extreme, F(1, 396) = 31.79, p < .001 (Cohen’s d = .81). In contrast, when there was a moderate default, participants were more likely to try the restaurant when they viewed an extreme evaluation rather than a moderate one, F(1, 396) = 27.09, p < .001 (Cohen’s d = .72). Mediation Replicating study 3, a serial mediated moderation model with bootstrapping provided evidence consistent with our prediction that the interaction effect (when we controlled for the main effects) on behavioral intentions occurred because the deviation boosted perceived thoughtfulness, which in turn boosted perceived accuracy (95% CI: –2.7662, –1.7445; figure 2). STUDY 4B Study 4b sought to replicate the findings from study 4a in a different context (house cleaner reviews) and using a different review rating scale (10 rather than 5 stars). Method Four hundred one Mechanical Turk participants completed an online study for payment. As in study 4a, participants were randomly assigned to conditions in a 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) between-participants design. All participants were asked to imagine that there was a new house cleaning service called “Elite Cleaners” and that they were reading reviews of it on a review website. Participants in the extreme (vs. moderate) default condition further read that when reviewers submit a review to this website, it automatically enters a default star rating of 10 (vs. 8) stars that reviewers can change. Next, participants in the extreme (vs. moderate) review condition read that they see a review of Elite Cleaners on this website, and the reviewer rated it 10 (vs. 8) stars. All participants then indicated their perception of the reviewer’s thoughtfulness, the review’s accuracy, and their willingness to use Elite Cleaners using the measures employed in the previous studies (adapted to refer to Elite Cleaners). Results Perceived Thoughtfulness A 2 (default: moderate vs. extreme) × 2 (reviewer evaluation: moderate vs. extreme) ANOVA on the perceived thoughtfulness data revealed no main effect of default, F(1, 397) = 2.45, p = .118, or the reviewer’s evaluation, F(1, 397) = .16, p = .688. However, the predicted interaction emerged, F(1, 397) = 176.57, p < .001 (table 2). When there was an extreme default, participants perceived the review to be more thoughtful when it was moderate rather than extreme, F(1, 397) = 92.97, p < .001 (Cohen’s d = 1.44). In contrast, when there was a moderate default, participants perceived the review to be more thoughtful when it was extreme rather than moderate, F(1, 397) = 83.69, p < .001 (Cohen’s d = 1.25). Perceived Accuracy Analysis of the perceived accuracy data revealed a main effect of default, F(1, 397) = 4.30, p = .039, but not the reviewer’s evaluation, F(1, 397) = .70, p = .402. Most importantly, the predicted interaction emerged, F(1, 397) = 150.05, p < .001 (table 2). When there was an extreme default, participants thought that the reviewer’s evaluation was more accurate when his review was moderate rather than extreme, F(1, 397) = 85.38, p < .001 (Cohen’s d = 1.38). In contrast, when there was a moderate default, participants thought that the reviewer’s evaluation was more accurate when it was extreme rather than moderate, F(1, 397) = 65.31, p < .001 (Cohen’s d = 1.10). Behavioral Intentions The behavioral intentions data revealed a main effect of the reviewer’s evaluation, F(1, 397) = 6.07, p = .014, but not of the default, F(1, 397) = .17, p = .678. More crucially, we found a significant interaction, F(1, 397) = 62.65, p < .001 (table 2). When there was an extreme default, participants were more likely to use Elite Cleaners when the reviewer’s evaluation was moderate rather than extreme, F(1, 397) = 14.74, p < .001 (Cohen’s d = .55). In contrast, when there was a moderate default, participants were more likely to use Elite Cleaners when the reviewer’s evaluation was extreme rather than moderate, F(1, 397) = 54.29, p < .001 (Cohen’s d = 1.03). Mediation Following the same procedure as in study 4a, a serial mediated moderation model with bootstrapping provided evidence consistent with our prediction that the deviatory reviews elicited more favorable behavioral intentions because they boosted perceived thoughtfulness, which in turn boosted perceived accuracy (95% CI: –2.7486, –1.8723; figure 2). STUDY 5 Studies 4a and 4b suggested that people infer that deviatory reviews are more accurate because deviation signals that the reviewer devoted careful thought to the review. In study 5, we employed a moderation design to provide converging evidence for the role of perceived thoughtfulness in this deviation effect. In particular, if the deviation effect occurs because deviatory reviews signal greater thoughtfulness, other evidence that a reviewer thought carefully about his or her recommendation should moderate the effect. In other words, if consumers perceive nondeviatory (vs. deviatory) reviews to be less accurate because consumers believe they are the result of less careful thought, then awareness that a nondeviatory review was the result of considerable thought should eliminate the persuasive advantage of deviatory reviews. Participants in study 5 were randomly assigned to view either an extremely lengthy review containing an elaborated discussion of the reviewed product, or a shorter review with less extensive discussion. We predicted that the considerable length of the former review would lead participants to conclude—regardless of the review’s deviatory status—that the reviewer devoted substantial thought to his review. As a result, we hypothesized that participants would find this review to be relatively persuasive regardless of its deviatory status. In contrast, when the review was relatively brief, we predicted that it would lack a clear signal regarding the author’s thoughtfulness, in which case perceived thoughtfulness (and thus accuracy and persuasiveness) would be derived from the review’s deviatory status. Several additional aspects of study 5 are worth noting. First, to enhance the external validity of the findings we leveraged real review content. Specifically, participants read real consumer reviews (sourced from Amazon.com), and we simply varied whether each review’s star rating deviated from a default. Second, to ensure that our findings were not dependent on the idiosyncratic content of any particular review, we administered three different long reviews and three different shorter reviews. Thus, participants were randomly assigned to read one review in a 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) × 3 (review content: version 1 vs. version 2 vs. version 3) between-participants design. We made an a priori decision to collapse across the stimulus sampling (i.e., review content) if it did not interact with our design of core theoretical interest—that is, the 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) design. Third, to provide more evidence that the process driving the deviation effect is not dependent on the order in which we measure thoughtfulness, accuracy, and behavioral intentions, we again measured behavioral intentions prior to thoughtfulness and accuracy as we did in study 3. Fourth, to further establish the robustness of the effect, we used a different product context: reviews for pens. Finally, we further extended the generalizability of the findings by conducting study 5 with a different online participant panel: Survey Sampling International (SSI). Method One thousand two hundred twenty-three participants, recruited from SSI, were randomly assigned to conditions in a 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) × 3 (review content: version 1 vs. version 2 vs. version 3) between-participants design. In this study, participants read a review for a pen that ostensibly had been posted on a consumer review website. Following the same procedure as in study 1a, all participants viewed an extreme default. Specifically, participants viewed an initial page of search results for pens, and all of the reviews displayed gave 5 out of 5 stars. Participants then read one of these reviews, which was either one of three long reviews (containing an elaborated discussion of the reviewed pen), or one of three brief reviews. All were real reviews extracted from Amazon, but we altered two features. First, to hold brand name constant we replaced all brand mentions with the (ostensibly real) brand name “Active Pen.” Second, we manipulated the review’s star rating. In the extreme (vs. moderate) review conditions, the pen’s rating was 5 (vs. 4) stars. After participants read the review, they indicated their likelihood of purchasing the pen as in studies 4a and 4b (this measure was adapted to refer to the pen). On the next survey screen, participants indicated their perceptions of the review’s accuracy and thoughtfulness, again assessed via the measures employed in studies 4a and 4b (adapted to refer to the pen). Results As previously noted, our core theoretical interest in study 5 was in the 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) design. As a robustness check—to ensure that any effects observed were not due to idiosyncratic features of a particular review we happened to sample—participants viewed one of three short reviews or three long reviews, and we manipulated whether that review was associated with a 4- or 5-star rating. Thus, we first conducted 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) × 3 (review content: version 1 vs. version 2 vs. version 3) ANOVAs on each dependent measure to test whether our stimulus sampling interacted with the 2 × 2 design of theoretical interest. As noted, we made an a priori decision that if there was no three-way interaction we would collapse across the stimulus samples to test our core predictions. Behavioral Intentions The 2 × 2 × 3 ANOVA on behavioral intentions revealed no three-way interaction, F(1, 1212) = 2.60, p = .107, but the predicted two-way interaction between reviewer evaluation and review length did emerge, F(1, 1212) = 3.94, p = .047. Therefore, we collapsed across review content condition. A subsequent 2 (reviewer evaluation: moderate vs. extreme) × 2 (review length: short vs. long) ANOVA on the collapsed data revealed main effects for review length, F(1, 1219) = 119.91, p < .001, and reviewer evaluation, F(1, 1219) = 8.01, p = .005, qualified by the predicted interaction, F(1, 1219) = 11.53, p = .001. As shown in table 3, participants reported that they were more likely to buy the reviewed pen when they viewed the moderate rating under short, F(1, 1219) = 19.32, p < .001 (Cohen’s d = .35), but not long, F(1, 1219) = .16, p = .688, review conditions. TABLE 1C ANALYSIS OF 4-STAR AND 5-STAR REVIEWS   b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001    b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001  TABLE 1C ANALYSIS OF 4-STAR AND 5-STAR REVIEWS   b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001    b  SE  z  p-value  Intercept  –0.80  0.01  –58.04  p < .001  Review star rating  –0.07  0.02  –4.38  p < .001  TABLE 1D ANALYSIS OF 4-STAR AND 5-STAR REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      TABLE 1D ANALYSIS OF 4-STAR AND 5-STAR REVIEWS, CONTROLLING FOR ALL COVARIATES Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      Fixed effects  b  SE  z  p-value  Intercept  240.00  14.19  16.88  p < .001  Price  0.00  0.00  –2.00  p < .05  Review word count  0.01  0.00  59.12  p < .001  Review year  –0.12  0.01  –16.97  p < .001  Review star rating  –0.16  0.07  −2.31  p < .05  Random effects  Variance  SD      Product category  0.13  0.37      Product  0.95  0.98      TABLE 2 PERCEIVED THOUGHTFULNESS, PERCEIVED ACCURACY, AND BEHAVIORAL INTENTIONS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 4A AND STUDY 4B   Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)    Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)  TABLE 2 PERCEIVED THOUGHTFULNESS, PERCEIVED ACCURACY, AND BEHAVIORAL INTENTIONS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 4A AND STUDY 4B   Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)    Study 4A   Study 4B     Extreme default   Moderate default   Extreme default   Moderate default   DV  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Extreme review  Moderate review  Perceived thoughtfulness                   M  3.40  5.19  4.58  3.22  3.26  5.41  5.10  3.08   (SD)  (1.90)  (1.38)  (1.72)  (1.59)  (1.79)  (1.13)  (1.51)  (1.73)  Perceived accuracy                   M  3.38  5.21  4.58  3.57  3.59  5.48  5.06  3.41   (SD)  (1.69)  (1.28)  (1.55)  (1.53)  (1.58)  (1.11)  (1.60)  (1.40)  Behavioral intentions                   M  3.93  5.16  4.89  3.75  3.88  4.67  4.96  3.47   (SD)  (1.73)  (1.26)  (1.65)  (1.53)  (1.50)  (1.36)  (1.36)  (1.53)  TABLE 3 BEHAVIORAL INTENTIONS, PERCEIVED ACCURACY, AND PERCEIVED THOUGHTFULNESS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 5   Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)    Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)  TABLE 3 BEHAVIORAL INTENTIONS, PERCEIVED ACCURACY, AND PERCEIVED THOUGHTFULNESS AS A FUNCTION OF DEFAULT CONDITION AND REVIEW CONDITION IN STUDY 5   Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)    Short reviews   Long reviews   DV  Moderate review  Extreme review  Moderate review  Extreme review  Behavioral intentions           M  4.30  3.75  4.96  5.01   (SD)  (1.57)  (1.58)  (1.50)  (1.50)  Perceived accuracy           M  4.92  4.46  5.25  5.31   (SD)  (1.27)  (1.59)  (1.26)  (1.33)  Perceived thoughtfulness           M  3.40  3.08  5.93  6.01   (SD)  (1.76)  (1.79)  (1.38)  (1.32)  Perceived Accuracy The 2 × 2 × 3 ANOVA on the perceived accuracy data revealed no three-way interaction, F(1, 1212) = .10, p = .754, but the predicted 2 × 2 interaction emerged, F(1, 1212) = 23.46, p < .001. Therefore, we collapsed across review content condition. A 2 × 2 ANOVA on the collapsed data revealed main effects for review length, F(1, 1219) = 56.81, p < .001, and reviewer evaluation, F(1, 1219) = 6.49, p = .011, qualified by the predicted interaction, F(1, 1219) = 11.56, p < .001 (table 3). As hypothesized, participants viewed the moderate review as more accurate than the extreme review under short, F(1, 1219) = 17.65, p < .001 (Cohen’s d = .32), but not long, F(1, 1219) = .36, p = .546, review conditions. Perceived Thoughtfulness The 2 × 2 × 3 ANOVA on the perceived thoughtfulness data revealed no three-way interaction, F(1, 1212) = .35, p = .552, but the predicted 2 × 2 interaction emerged, F(1, 1212) = 21.83, p < .001. After we collapsed across review content condition, a 2 × 2 ANOVA revealed a main effect for review length, F(1, 1219) = 919.81, p < .001, indicating that the thoughtfulness manipulation was successful. The analysis indicated no main effect of the reviewer’s evaluation, F(1, 1219) = 1.81, p = .179, but the predicted interaction emerged, F(1, 1219) = 4.69, p = .031 (table 3). Participants perceived the moderate review to be more thoughtful than the extreme review under short, F(1, 1219) = 6.13, p = .013 (Cohen’s d = .18), but not long, F(1, 1219) = .34, p = .561, review conditions. Mediation As illustrated in figure 2, a moderated mediation analysis with bootstrapping provided evidence consistent with our prediction that the interaction between review length and reviewer evaluation affected behavioral intentions through perceived thoughtfulness and accuracy (95% CI: .1040, .4088). Discussion In showing that the deviation effect is both mediated and moderated by perceived thoughtfulness, study 5 offers convergent evidence for the mechanism driving this effect. Also important, the moderation pattern observed in study 5 suggests that the deviation effect has limits. In particular, other salient evidence that a reviewer devoted substantial thought to a review eliminates the influence of that review’s deviatory status on perceived thoughtfulness, and thus on perceived accuracy and persuasion. Because this moderation pattern provides crucial insight into the process driving the deviation effect, we conducted a conceptual replication in a different domain (restaurants) with a different thoughtfulness manipulation. Specifically, the replication study presented participants with a review’s rating but not its content, and manipulated thoughtfulness by providing some (but not other) participants with explicit information that a reviewer devoted substantial thought to his review. This study revealed the same moderation and mediation patterns as in study 5 (web appendix C). STUDY 6 As outlined earlier, we propose an attributional mechanism for the deviation effect, whereby people draw inferences about reviewer thoughtfulness and accuracy based on whether a review deviates from or adheres to a default. We theorize that this process is most likely to manifest under conditions that allow for deliberative cognitive processing, because it requires that people reflect on others’ actions (i.e., their moderate or extreme endorsements) and make situational corrections (i.e., for perceived defaults) before forming inferences about them (i.e., their thoughtfulness and accuracy). This sort of attributional reasoning requires ample processing resources (Gilbert, Pelham, and Krull 1988). Past research further reveals that simulating and forming inferences about other people’s thought processes can require careful deliberation (Campbell and Kirmani 2000; Goldman and Jordan 2013). Thus, multiple streams of research suggest that the process we have outlined should be more likely to emerge under conditions that permit rather than restrict thorough cognitive processing. What happens when processing is constrained? Research suggests that people often rely on a “more is better,” or “numerosity,” heuristic when their processing is limited in some way (Khan and Kupor 2016; Pelham, Sumarta, and Myaskovsky 1994; Petty and Cacioppo 1984; Tormala, Petty, and Briñol 2002). This implies that people may simply find higher ratings more persuasive when their processing is constrained. Thus, we predict that if people are unable to deliberate, they will be more persuaded by an extreme review regardless of its deviatory status. To examine this possibility, we presented participants in study 6 with a 4- or 5-star review in the context of a 5-star default, and we manipulated cognitive load. We predicted that participants whose cognitive resources were unencumbered would more favorably evaluate the reviewed product when they viewed the deviatory 4-star review. Importantly, though, we predicted that this pattern would reverse when participants were under cognitive load. Under load, we expected participants to more favorably evaluate the reviewed product when they viewed the 5-star review. If obtained, this result would suggest that the deviation effect is driven by a thoughtful inferential process requiring ample cognitive resources. Method Four hundred sixty Mechanical Turk participants completed an online study for payment. At the beginning of the survey, all participants read that one of the goals of the survey was to research multitasking, and that they might be prompted to switch between tasks during the study. Participants further learned that the study aimed to investigate people’s interests in various topics, so they would listen to a recording and report how interested they were in it. Following these opening instructions, participants were randomly assigned to conditions in a 2 (reviewer evaluation: moderate vs. extreme) × 2 (cognitive load: no vs. yes) between-participants design. Participants in the moderate review condition read a review for a pen in which the reviewer wrote: “Active Pens are good—4 out of 5 stars!” In the extreme review condition, the reviewer wrote: “Active Pens are great—5 out of 5 stars!” To establish a 5-star default, all participants read that this review had been entered into a review platform that automatically preselects a 5-star rating. As noted, participants were also randomly assigned to either a cognitive load condition or a no cognitive load condition. In the load condition, participants viewed the information about the pen while completing a distraction task (adapted from Lin, Lee, and Robertson 2011; Lee, Lee, and Boyle 2007; Sakuraba 2012). Specifically, before reading any information about the reviewed pen or the review platform, participants under load were instructed that while they viewed written information on the next screen, they would listen to an unrelated audio recording about Pluto. Participants in this condition then viewed the focal information about the pen while simultaneously listening to the audio recording. To bolster the cover story, after these participants listened to the recording, they indicated their interest in the topic discussed in the recording on a seven-point scale (1: Not interesting at all; 7: Very interesting). In order to control for any potential impact of the recording’s content, we had participants in the no load condition listen to and rate their interest in the same recording prior to viewing the information about the pen. Thus, all participants listened to the same recording. This recording was used to constrain the processing resources of participants in the load condition but not in the no load condition. Finally, all participants indicated their likelihood of purchasing an Active Pen using the measure employed in study 5. Results and Discussion A 2 (reviewer evaluation: moderate vs. extreme) × 2 (cognitive load: no vs. yes) ANOVA revealed no main effect of load, F(1, 456) = .04, p = .847, or reviewer evaluation, F(1, 456) = .02, p = .889, but the predicted interaction emerged, F(1, 456) = 10.16, p = .002. As expected given the extreme default, participants in the no load condition had more favorable behavioral intentions when they viewed the moderate (M = 4.44; SD = 1.29) rather than extreme (M = 3.99; SD = 1.52) review, F(1, 456) = 5.29, p = .022; Cohen’s d = .32). This effect reversed among participants in the load condition: these participants had more favorable behavioral intentions when they viewed the extreme (M = 4.45; SD = 1.52) rather than moderate (M = 4.04; SD = 1.47) review, F(1, 456) = 4.87, p = .028; Cohen’s d = .27). In short, cognitive load reversed the deviation effect, leading people under load to be more persuaded by an extreme, nondeviating review than by a moderate, deviating review. These results provide support for the notion that the deviation effect is driven by thoughtful attributional reasoning. They also point to an additional constraint on the effect. When consumers’ processing resources are constrained in some way, which may occur when consumers are distracted or under time pressure while searching for products, they are more likely to be persuaded by extreme reviews even when those reviews adhere to a preselected default. Indeed, consumers may simply rely on a “more stars are better” heuristic under these conditions. In contrast, when consumers can think more carefully, the default is more likely to influence their assessments and the deviation effect is thus more likely to emerge. In documenting this boundary on the deviation effect, this study might provide initial insight into the conditions under which spontaneous accuracy inferences drive (vs. do not drive) product evaluations, and thus illuminate the situations in which the deviation effect emerges versus reverses. In particular, study 6 is consistent with the possibility that the accuracy inferred from a deviatory rating primarily shapes behavioral intentions when consumers’ cognitive resources are unencumbered. It could be that extremely positive ratings imply greater quality (regardless of the default) when consumers do not spontaneously consider the accuracy of those ratings, and extremely positive ratings may thus prove more persuasive when consumers’ cognitive resources are restricted in some way. We encourage future research to explore this possibility more deeply. GENERAL DISCUSSION This research provides evidence for a deviation effect in persuasion. We found that when the default evaluation in a review context is extreme, moderate reviews can actually be more persuasive than extreme ones. However, when the default review is moderate, extreme reviews are more persuasive. We obtained reliable evidence for this effect across eight studies despite numerous changes to the product category, study materials, and sample population. For instance, we found the deviation effect in the context of reviews for a café, a granola bar, a ridesharing service, a restaurant, a cleaning service, and pens. We also found it using different operationalizations of deviations, including deviations from defaults established via explicit descriptive norms, an implicit descriptive norm, and preselected star ratings. Also important, we observed the deviation effect in diverse samples, including university students and participants from numerous online panels: Mechanical Turk, SurveyMonkey, and SSI. Finally, we observed these effects on measures of behavioral intentions and actual choice. Importantly, our studies offer both mediation and moderation evidence consistent with the proposed account for this effect—that is, that people perceive deviatory reviews to be more accurate because they believe those reviews are the result of greater thought. It is worth noting that recent work (Pieters 2017) has outlined a series of conditions that researchers should strive to meet before drawing conclusions about mediation effects. These include directionality, reliability, unconfoundedness, distinctiveness, and power. As outlined below, we made an effort to follow these recommendations in the current work. First, to establish directionality, we not only measured the proposed mediator (studies 2–5), but also manipulated it (study 5 and the supplemental study in web appendix C). We also leveraged prior literature (Kupor et al. 2014; Priester and Petty 1995, 2003; Schrift and Amar 2015) suggesting that perceptions of a source’s thoughtfulness and accuracy can drive persuasive outcomes even when those variables are experimentally manipulated, making the reverse causal direction implausible. To ensure that our measures were reliable, we borrowed heavily from past research (cited in our methods sections) showing that these measures are both sensitive to experimental manipulations and predictive of related outcomes. To reduce confoundedness between measures, in addition to directly manipulating the mediator (as noted), we created spatiotemporal distance between the mediators and outcomes in several of our studies (studies 3 and 5) by placing those measures on separate survey screens. Furthermore, as we explain in the forthcoming section, “Alternative Explanations,” we rule out numerous competing accounts on conceptual grounds. Regarding distinctiveness of our mediator and outcome variables, we note that (1) past research has documented numerous contexts in which these constructs diverge (Kupor et al. 2014; Schrift and Amar 2015), and (2) the raw correlations between our mediators and outcome variables ranged from .19 to .82 across studies. Finally, in an effort to test our hypotheses with reasonable power, we followed the Simmons, Nelson, and Simonsohn (2013) recommendation of collecting at least 50 participants per condition in studies 1a and 1b, and collected approximately 100 participants per condition in all subsequent studies. Follow-up analyses revealed that several of the studies testing mediation (studies 4a and 4b and web appendix C) had at least 80% power to identify direct and indirect effects. Taken together, these features suggest that the current studies provide meaningful mechanistic insight. In terms of theoretical contribution, this research is the first to suggest that the perceived default in a review context can shape people’s willingness to be influenced by an extreme or moderate review. Moreover, while past research has revealed much about the effects of preselected options on consumers’ judgments (Everett et al. 2015; Johnson and Goldstein 2003; Probst et al. 2013), it has not examined how observers perceive (and are thus influenced by) the judgments of others who adhere to or deviate from these defaults. Thus, the current studies advance our understanding of the diverse effects of defaults in consumer contexts. This research is also the first to show that a moderate review can be more persuasive than an extreme review. This finding is especially interesting in that it conflicts with people’s lay intuitions about what should be persuasive. Indeed, a follow-up study revealed that people do not intuit the persuasive power of moderate reviews in the context of extreme defaults. In this study, we asked participants to imagine review websites on which the default rating was 5 stars. Participants predicted that their own reviews on these websites would be more persuasive if they rated a product 5 rather than 4 stars, and thus that they would be more likely to give 5 stars if they sought to persuade others to purchase the product (see full details in web appendix D). This study suggests that consumers do not intuit that deviatory reviews can be more persuasive. Thus, in their attempts to increase their own persuasive influence, people might inadvertently decrease it by avoiding moderate endorsements precisely when such endorsements would be more persuasive. In demonstrating the effect of deviations on perceptions of thoughtfulness and accuracy, the current studies may also uncover a novel determinant of source credibility. In general, sources believed to be more thoughtful and accurate are seen as more credible (Berlo, Lemert, and Mertz 1969; Chesney 2006; Flanagin and Metzger 2000; Hovland, Janis, and Kelley 1953; Kaufman, Stasson, and Hart 1999; McCroskey 1966), and credibility is a well-known factor in persuasion (Briñol and Petty 2009; Petty, Cacioppo, and Schumann 1983). Consistent with this general possibility, we obtained mixed evidence for the notion that deviations impact perceptions of trustworthiness—a well-established component of credibility—in the studies in which we measured it (see web appendixes A and B). Although the current research was not designed to delineate the role of deviations in shaping source credibility, it may hint at a novel means of establishing credibility: authoring a deviatory review. Particularly intriguing is the notion that making one’s review more moderate might in some cases make one seem more credible. Future research investigating how deviations might shape perceptions of credibility in its different forms (i.e., expertise and trustworthiness; Conchie and Burns 2009; McGinnies and Ward 1980; Ong 2012) would be worthwhile. The current research also contributes to the word-of-mouth (WOM) literature. Recent work on WOM has examined how a review’s impact can be shaped by reviewer characteristics (e.g., reviewer expertise and similarity to the observer; Forman, Ghose, and Wiesenfeld 2008; Reichelt, Sievert, and Jacob 2014; Rosario et al. 2016), product characteristics (e.g., product lifespan, product risk, and product category; Kupor, Tormala, and Norton 2014; Murray 1991; Rosario et al. 2016; Wangenheim and Bayón 2004), and characteristics of the review itself (e.g., review length and valence; Chevalier and Mayzlin 2006; Doh and Hwang 2009). In addition, this literature has begun to examine how the review platform can moderate the impact of reviews. For example, platforms that impose more structure in their display—by including review summaries, for instance—have been found to exert greater impact on consumer behavior (Rosario et al. 2016). Our research contributes to this growing literature in revealing that another feature of review platforms—that is, the perceived default rating—can moderate the impact of the reviews presented. Alternative Explanations Although we provided evidence consistent with the proposed mechanism behind the deviation effect, it is important to consider alternative processes that might have played an added role. Consideration of Negatives First, could the current results reflect a blemishing or two-sided messaging effect? The blemishing effect refers to the idea that delivering a small dose of weak negative information after strong positive information can enhance persuasion (Ein-Gar, Shiv, and Tormala 2012). Likewise, the two-sided messaging effect suggests that people are sometimes more persuaded by recommenders who have considered both a product’s positives and negatives (e.g., rather than just its positives; Eisend 2006; Rucker, Petty, and Briñol 2008). Could it be that moderate reviews outperformed extreme reviews in the current studies because they seemed to incorporate negative information? Although it is theoretically possible that moderate endorsements could foster the perception that negative information has been considered, this would not offer a viable account for the current results. For instance, studies 4a and 4b found that a moderately positive evaluation is more persuasive in the context of an extremely positive default, but that the opposite is true in the presence of a moderate default. This result is inconsistent with both a blemishing and a two-sided message account. Neither would predict that manipulating an external default (which does not alter whether the review itself contains negative information) would moderate the deviation effect. After all, if moderate reviews signal that negative information has been considered, they should do so regardless of the default. Studies 3 and 6 provide further evidence inconsistent with the possibility that the blemishing effect underlies the deviation effect. First, the blemishing effect exclusively occurs when a small dose of negative information is added to otherwise positive information, but study 3 finds that the deviation effect persists when a reviewer deviates from an extremely negative default to a moderately negative evaluation. Study 6 is also inconsistent with a blemishing effect—the blemishing effect most reliably emerges under conditions of low elaboration (Ein-Gar et al. 2012), whereas precisely the opposite is true for the deviation effect (study 6). Furthermore, in additional empirical work (see web appendix E), we found that the deviation effect extends beyond deviations from rating extremity defaults. Deviations from other types of defaults foster persuasion as well. For example, companies often communicate a default by highlighting on their websites that the majority of consumers give them specific positive feedback in reviews (Groupon 2016; Printing Center USA 2018; Yelp 2015). When companies communicate such a default (e.g., “Most people say that they love getting their house cleaned by us…If you did, please write ‘I loved it and recommend it’”), authoring an equally favorable review that does not employ the default language (e.g., “It was fantastic and I recommend it”) appears to be more persuasive. That deviations can foster persuasion even when they provide no hint of negative considerations further suggests that the deviation effect is unlikely to stem from perceptions of blemishing or two-sided messaging. Expectation Violations As another possibility, is the deviation effect partially driven by the fact that deviations are unexpected? After all, we propose an attributional inference process, and unexpected events are known to prompt attributional reasoning (Clary and Tesser 1983; Reich and Tormala 2013). To investigate this possibility, we assigned Mechanical Turk participants to either the deviation or no deviation condition from study 4a, and then asked them, “To what extent is this reviewer’s rating unexpected?” and “To what extent is this reviewer’s rating surprising?” Participants responded on separate scales (1: Not at all; 7: Very much). Participants perceived the reviews to be equally expected (MDeviation = 3.34, SDDeviation = 1.89; MNo Deviation = 3.18, SDNo Deviation = 1.81; t(207) = .63, p = .532) and unsurprising (MDeviation = 3.15, SDDeviation = 1.87; MNo Deviation = 3.23, SDNo Deviation = 1.91; t(207) = .30, p = .764). These results do not preclude the possibility that some types of deviations are perceived as unexpected, but the fact that the deviation effect emerged in contexts in which deviations are not unexpected suggests that an expectancy violation account is insufficient to explain our findings. Attention Related to this point, is the deviation effect partially driven by increased attention to deviatory reviews? Although this is a reasonable proposition, we suspect it is unlikely to account for our full pattern of data. First, if deviations operate by attracting attention, it seems likely that the deviation effect would be amplified among consumers devoting relatively low levels of attention and processing to a review at baseline—for example, among those under cognitive load. Indeed, past research reveals that persuasive interventions that operate by boosting processing tend to have greater impact when consumers’ baseline processing levels are low (Kupor and Tormala 2015; Priester and Petty 1995; Tormala and Clarkson 2008). In contrast, our studies reveal that the deviation effect not only fails to increase but in fact reverses under low processing conditions (study 6). This pattern is inconsistent with a pure attentional mechanism. Moreover, an attention account would not predict the results of study 5, in which review length moderated the effect of deviatory reviews. If deviations increase persuasion by increasing attention, they should do so regardless of whether they feature brief or extended discussions of the reviewed product. In short, despite offering an intuitively appealing account at first blush, it is unlikely that all of our findings can be explained by an attentional mechanism. Unseen Reviews Another alternative that is worth considering is whether the deviation effect could be due to an inference that if a moderate review contains a positive evaluation of the reviewed product, then unseen reviews that are more extremely positive may contain even more glowing evaluations. Perhaps this inference prompted speculation about what those reviews might say, which yielded a more favorable response. However, studies 4–5 are incompatible with this alternative: this account would not predict that an extreme deviatory review would be more persuasive in the context of a moderately positive default (studies 4a and 4b), nor that the deviation effect would disappear when nondeviatory reviews resulted from perceived substantial thought (study 5). Thus, the moderation patterns observed in studies 4–5 are inconsistent with this alterative account. Experimental Artifact Also relevant, our studies demonstrate that perceived thoughtfulness and accuracy need not be made artificially salient within the experimental context for the deviation effect to manifest. Although some of our studies measured these constructs prior to behavioral intentions, which could have increased their salience, not all did so. In studies 1a, 1b, 3, 5, and 6, we observed the deviation effect even when participants did not complete thoughtfulness and accuracy measures prior to reporting their behavioral intentions. Thus, consistent with prior literature (Cheung et al. 2009; Willemsen et al. 2012), participants appeared to draw spontaneous inferences about others’ thoughtfulness and accuracy when viewing others’ reviews. Of course, as with most psychological effects, we suspect the deviation effect will be stronger when its psychological drivers are particularly salient. Likewise, when another dimension is more salient, the effect might shift. For instance, because people do not anticipate the deviation effect (web appendix D), it is possible that the effect attenuates when people anticipate that they will need to justify their decisions to others. This is because people who anticipate they will need to justify their decisions to others often internalize the criteria that they believe others employ, and use those criteria to make their decisions (Huber and Seiser 2001; Simonson 1989). Thus, the deviation effect may attenuate in contexts in which people anticipate external evaluation (e.g., when superiors will evaluate their decisions; Simonson 1989). We encourage future research to investigate this possibility. Extreme Means Biased? Another alternative that is important to consider is whether the deviation effect occurs because consumers are aware that moderate reviews are less likely to be skewed by a priori biases (e.g., “brag-and-moan” biases; Hu, Zhang, and Pavlou 2009), and people seek to protect themselves from such biases by discounting extreme reviews. This is an intriguing possibility, but it is incompatible with studies 4–5. Studies 4a and 4b demonstrated that a review’s deviatory status (rather than its extremity per se) drove its persuasiveness. Study 5 showed that participants were equally persuaded by extreme and moderate reviews when those reviews were especially lengthy. Nonetheless, to the extent that extreme endorsements are more biased (Hu et al. 2009), or at least less measured, identifying means of steering consumers away from their influence could have implications for consumer welfare. We encourage future research to investigate that possibility. Boundaries and Extensions There are likely to be constraints on the deviation effect in addition to the two that we documented (i.e., review length and cognitive load). For example, the magnitude of deviation may moderate its effect (i.e., the deviation effect likely attenuates as reviews become increasingly moderate). The current studies found that 4-star deviatory reviews can be more persuasive than 5-star nondeviatory reviews (studies 4a, 5, and 6), and that 8-star (studies 1b, 2, and 4b) and 9-star (study 1a) deviatory reviews can be more persuasive than 10-star nondeviatory reviews. Is a 7-star deviatory review also more persuasive than a 10-star nondeviatory review? We submit that it could be if it is perceived to be sufficiently favorable. In an initial test of this possibility (web appendix F), we compared 7-star, 8-star, and 9-star deviatory reviews to a 10-star nondeviatory review. This study revealed that even a 7-star deviatory review was more persuasive than a 10-star nondeviatory review. Nonetheless, there is likely an extremity threshold, below which a review is no longer sufficiently favorable to outperform an extremely favorable nondeviatory review. This threshold might vary across contexts as a function of numerous factors (e.g., the perceived strength of the default), and we encourage future exploration of this issue. Related to this point, future research could profit from leveraging information integration theory to precisely estimate consumers’ judgments in the context of defaults, and to model how these judgments may shift as reviews become increasingly moderate. Information integration theory formalizes the manner in which different parameters (e.g., a review’s extremity and perceived accuracy) are integrated into judgment by assigning each parameter a unique weight (w; reflecting the parameter’s impact on judgment) and specifying an algebraic composition rule representing their integration (Anderson 1981; Peysakhovich and Karmarkar 2016). Although the primary focus of our research is to examine the psychological processes through which deviating endorsements are persuasive, our findings underscore the potential value of developing a formal model of information integration in the current context. Indeed, our studies demonstrate the manner in which people consider not only the extremity of consumer reviews when they process them, but also the accuracy of those reviews. These are dynamic parameters that could be modeled. Precisely how much weight each parameter receives, and how that weight interacts with degree of deviation and other contextual factors, is beyond the scope of the current research. Nevertheless, this is an interesting and potentially fruitful path for future work. One contextual factor that could be important to examine in future research relates to the manner in which reviews are aggregated online. For example, some review platforms rank products as a function of their average rating. In such cases, a product that has a default rating of 5 out of 5 stars would have a higher ranking if none of its ratings deviated from 5 stars compared to if some of its ratings deviated to a lower rating. On websites in which deviating reviews lower a product’s ranking, some consumers might be less likely to consider that product as a mere function of its lower ranking. If true—that is, if consumers were not even exposed to deviating reviews—the deviation effect would be unlikely to manifest. We suspect there are numerous situational and individual factors that moderate consumers’ likelihood of considering a lower-ranked option, and thus moderate whether the deviation effect emerges on these ranked review platforms. For example, highly involved consumers who are interested in learning about a wide range of alternatives should be more likely to consider lower-ranked options. Likewise, consumers who are cognitively unencumbered (e.g., who are not distracted or under time pressure) might be more inclined to consider lower-ranked options. Conversely, consumers who are uninterested or whose processing resources are constrained (e.g., who are distracted or under time pressure) might be less likely to do so. Thus, consumers’ depth of processing not only may moderate the persuasive impact of deviating reviews that consumers view (as in study 6), but also may determine whether consumers view a deviating review in the first place (which, of course, would be required to open someone up to the deviation effect). We submit that once consumers have evaluated multiple products—not just the top-ranked one(s)—they might ultimately choose a lower-ranked option if it has a deviating favorable endorsement, because the deviating endorsement appears more thoughtful and accurate. We encourage future research to examine this possibility. Also relevant to future research, the inferences that consumers draw from deviatory evaluations may have numerous downstream consequences beyond increasing persuasion. For example, inferences of thoughtfulness and accuracy can affect a variety of outcomes, including liking, certainty, and trustworthiness (Barden and Petty 2008; Conchie and Burns 2009; Kupor et al. 2014; Kupor, Flynn, and Norton 2017). Consistent with the possibility that deviations may have a wider range of consequences than those documented here, our analysis of supplemental measures suggested that deviations may affect perceived trustworthiness and knowledge (see web appendixes A and B). These measures were purely exploratory, but the fact that deviations affected responses to them suggests that they might be worth future study. We encourage research to investigate the expanse of inferences that people draw from deviations, and the consequences of these inferences. Of importance, we investigated our theorizing in the context of reviews not only because defaults are common in these contexts, but also because reviews are a primary determinant of consumers’ purchase decisions (Allsop, Bassett, and Hoskins 2007). Examining our theorizing in this context thus provides insight into a key driver of consumer decision making. It is worth noting, however, that the psychology underlying the effects we observed might apply to deviations from defaults in other settings as well. For instance, people buying a new car often face the decision of whether to alter the car’s default features, people purchasing restaurant meals often have the option to change the default side dish, and social media users can opt (or not) to change their default privacy settings. Based on the current insights, consumers who observe others deviating from such defaults might infer that these decisions are more thoughtful and accurate, and thus be more influenced by them. Future research could investigate this possibility. Strategic Defaults Finally, our research has strategic implications for marketers. For example, our findings suggest that marketers may increase sales by publicizing positive reviews that have clearly deviated from a default evaluation. Similarly, when a product receives a moderately positive review in the context of an extremely positive default, firms might enhance that review’s positive impact (e.g., on sales) by highlighting that the default is extremely positive. In other words, making the default salient could increase perceptions of a moderately positive review’s accuracy, and heighten sales as a result. In addition, if marketers are certain (through market testing, historical review data, or other means) that reviewers will rate their products highly, then implementing a default rating that differs from reviewers’ normative ratings could boost sales when consumers see that reviewers deviated from that default but still gave the product a favorable rating. In other words, consumers might be more persuaded to purchase a favorably reviewed product when those favorable reviews are given in the context of a different default, thus indicating that reviewers put thought into their favorable ratings. In a similar vein, it is possible that consumers not only perceive others as thinking more carefully when others deviate from a default, but also perceive themselves as thinking more carefully when they themselves deviate from a default. Therefore, the strategic use of defaults could boost sales not only among consumers who read reviews, but also among the reviewers themselves. People are more certain of their attitudes when they perceive that they have thought more about them (Barden and Petty 2008). Attitude certainty, in turn, increases attitudes’ persistence over time, resistance to attack, and impact on behavior (Rucker et al. 2014). Thus, if marketers’ strategic use of defaults boosts reviewers’ perceptions of their own thoughtfulness—and thus their attitude certainty—defaults may generate substantially more favorable future attitudes and behavior. DATA COLLECTION INFORMATION The first author supervised the collection of data for the first study by research assistants at Boston University in 2017–2018. The first author managed the collection of data for studies 2, 4a, 4b, and 6 using Mechanical Turk. The first author managed the collection of data for study 3 using SurveyMonkey, and for study 5 using SSI. The data were analyzed by the first author. The authors thank Sofia Nikolakaki and Ewart Thomas for their help with scraping and analyzing the secondary data. Supplementary materials are included in the web appendix accompanying the online version of this article. References Aaker Jennifer, Vohs Kathleen D., Mogilner Cassie ( 2010), “Nonprofits Are Seen as Warm and For-Profits as Competent: Firm Stereotypes Matter,” Journal of Consumer Research , 37 ( 2), 224– 37. Google Scholar CrossRef Search ADS   Active Pest Control ( 2018), Review Us, https://activepestcontrol.com/review-us/. Agnew Julie R., Szykman Lisa R. ( 2005), “Asset Allocation and Information Overload: The Influence of Information Display, Asset Choice, and Investor Experience,” Journal of Behavioral Finance , 6 ( 2), 57– 70. Google Scholar CrossRef Search ADS   Allsop Dee T., Bassett Bryce R., Hoskins James A. ( 2007), “Word-of-Mouth Research: Principles and Applications,” Journal of Advertising Research , 47 ( 4), 398– 411. Google Scholar CrossRef Search ADS   Amazon ( 2014), “Medicasp Therapeutic Shampoo,” https://www.amazon.com/Medicasp-Therapeutic-Shampoo/dp/B00BCYZOIG. Amazon ( 2015), “Buy this book!” (customer review by Sunder), https://www.amazon.com/gp/customer-reviews/R985UQZQ5FH2C/ref=cm_cr_arp_d_rvw_ttl?ie=UTF8&ASIN=B00506VMJ2. Amazon ( 2016), “Nice ribbon in a variety of colors,” https://www.amazon.com/gp/product/B0097K2JF8/ref=s9u_simh_gw_i2?ie=UTF8&fpl=fresh&pd_rd_i=B0097K2JF8&pd_rd_r=d8072b17-381f-11e8-bb6d-df8617f41b33&pd_rd_w=5Ll4c&pd_rd_wg=8wFGg&pf_rd_m=ATVPDKIKX0DER&pf_rd_s=&pf_rd_r=6ARY8T8CW7JZXZZGEH62&pf_rd_t=36701&pf_rd_p=0411ffec-c026-40ae-aac5-2cd3d48aeeac&pf_rd_i=desktop. Anderson Norman H. ( 1981), Foundation of Information Integration Theory , New York: Academic Press. ASUS ( 2015), “Status Bar Transparency Setting” (forum post from user Naresh), https://www.asus.com/zentalk/thread-2941-1-1.html. Bailey Jeffrey J., Nofsinger John R., O’Neill Michele ( 2004), “401(k) Retirement Plan Contribution Decision Factors: The Role of Social Norms,” Journal of Business and Management , 9 ( 4), 327– 44. Barden Jamie, Petty Richard E. ( 2008), “The Mere Perception of Elaboration Creates Attitude Certainty: Exploring the Thoughtfulness Heuristic,” Journal of Personality and Social Psychology , 95 ( 3), 489– 509. Google Scholar CrossRef Search ADS   Barden Jamie, Tormala Zakary L. ( 2014), “Elaboration and Attitude Strength: The New Meta‐cognitive Perspective,” Social and Personality Psychology Compass , 8 ( 1), 17– 29. Google Scholar CrossRef Search ADS   Bargh John A., Chartrand Tanya L. ( 1999), “The Unbearable Automaticity of Being,” American Psychologist , 54 ( 7), 462– 79. Google Scholar CrossRef Search ADS   Berlo David K., Lemert James, Mertz Robert ( 1969), “Dimensions for Evaluating the Acceptability of Message Sources,” Public Opinion Quarterly , 33 ( 4), 563– 76. Google Scholar CrossRef Search ADS   Brett Joan F., Atwater Leanne E. ( 2001), “360° Feedback: Accuracy, Reactions, and Perceptions of Usefulness,” Journal of Applied Psychology , 86 ( 5), 930– 42. Google Scholar CrossRef Search ADS   Briñol Pablo, Petty Richard ( 2009), “ Persuasion: Insights from the self-validation hypothesis,” Advances in Experimental Social Psychology , 41, 69– 118. Google Scholar CrossRef Search ADS   Campbell Margaret C., Kirmani Amna ( 2000), “ Consumers' use of persuasion knowledge: The effects of accessibility and cognitive capacity on perceptions of an influence agent,” Journal of Consumer Research , 27 ( 1), 69– 83. Google Scholar CrossRef Search ADS   Cress ( 2018), http://www.cressfuneralservice.com/review/. Chaiken Shelly, Eagly Alice H. ( 1989), “Heuristic and Systematic Information Processing Within and Beyond the Persuasion Context,” in Unintended Thought , ed. Uleman James S., Bargh John A., New York: Guilford, 212– 47. Chesney Thomas ( 2006), “An Empirical Examination of Wikipedia’s Credibility,” First Monday , 11 ( 11), http://firstmonday.org/issues/issue11_11/chesney/index.html. Cheung Man Yee, Luo Chuan, Sia Choon Ling, Chen Huaping ( 2009), “ Credibility of electronic word-of-mouth: Informational and normative determinants of on-line consumer recommendations,” International Journal of Electronic Commerce , 13 ( 4), 9– 38. Google Scholar CrossRef Search ADS   Chevalier Judith A., Mayzlin Dina ( 2006), “The Effect of Word of Mouth on Sales: Online Book Reviews,” Journal of Marketing Research , 43 ( 3), 345– 54. Google Scholar CrossRef Search ADS   City-Data Real Estate Forum ( 2013), http://www.city-data.com/forum/real-estate/1807095-hiring-realtor-based-zillow-reviews.html. Clary Gil, Tesser Abraham ( 1983), “Reactions to Unexpected Events,” Personality and Social Psychology Bulletin , 9 ( 4), 609– 20. Google Scholar CrossRef Search ADS   Clement Russell W., Krueger Joachim ( 2002), “Social Categorization Moderates Social Projection,” Journal of Experimental Social Psychology , 38 ( 3), 219– 31. Google Scholar CrossRef Search ADS   Conchie Stacey M., Burns Calvin ( 2009), “Improving Occupational Safety: Using a Trusted Information Source to Communicate about Risk,” Journal of Risk Research , 12 ( 1), 13– 25. Google Scholar CrossRef Search ADS   Connors Laura, M. Mudambi Susan, Schuff David ( 2011), “Is it the review or the reviewer? A multi-method approach to determine the antecedents of online review helpfulness.” System Sciences (HICSS), 2011 44th Hawaii International Conference on IEEE. Crotty David ( 2009), “How Meaningful Are User Ratings?” https://scholarlykitchen.sspnet.org/2009/11/16/how-meaningful-are-user-ratings-this-article-4-5-stars/. Davies Martin ( 1994), “Private Self-Consciousness and the Perceived Accuracy of True and False Personality Feedback,” Personality and Individual Differences , 17 ( 5), 697– 701. Google Scholar CrossRef Search ADS   Deutsch Morton, Gerard Harold B. ( 1955), “A Study of Normative and Informational Social Influences upon Individual Judgment,” Journal of Abnormal and Social Psychology , 51 ( 3), 629– 36. Google Scholar CrossRef Search ADS   Dimaggio Giancarlo, Lysaker Paul H., Carcione Antonino, Nicolò Giuseppe, Semerar Antonio ( 2008), “Know Yourself and You Shall Know the Other…to a Certain Extent: Multiple Paths of Influence of Self-Reflection on Mindreading,” Consciousness and Cognition , 17 ( 3), 778– 89. Google Scholar CrossRef Search ADS   Doh Sun-Jae, Hwang Jang-Sun ( 2009), “How Consumers Evaluate eWOM (Electronic Word-of-Mouth) Messages,” Cyberpsychology & Behavior: The Impact of the Internet, Multimedia and Virtual Reality on Behavior and Society , 12 ( 2), 193– 7. Dr. Oogle ( 2016), https://www.doctor-oogle.com/. Durham Volkswagen ( 2018), https://www.durhamvw.com/review-us/ Ein-Gar Danit, Shiv Baba, Tormala Zakary L. ( 2012), “When Blemishing Leads to Blossoming: The Positive Effect of Negative Information,” Journal of Consumer Research , 38 ( 5), 846– 59. Google Scholar CrossRef Search ADS   Eisend Martin ( 2006), “Two-Sided Advertising: A Meta-analysis,” International Journal of Research in Marketing , 23 ( 2), 187– 98. Google Scholar CrossRef Search ADS   Everett Jim A. C., Caviola Lucius, Kahane Guy, Savulescu Julian, Faber Nadira S. ( 2015), “Doing Good by Doing Nothing? The Role of Social Norms in Explaining Default Effects in Altruistic Contexts,” European Journal of Social Psychology , 45 ( 2), 230– 41. Google Scholar CrossRef Search ADS   Faraway Julian ( 2006), Extending the Linear Model with R: Generalized Linear, Mixed Effects and Nonparametric Regression Models , Boca Raton, FL: Chapman & Hall/CRC Texts in Statistical Science. Filieri Raffaele, McLeay Fraser ( 2014), “E-WOM and Accommodation: An Analysis of the Factors That Influence Travelers’ Adoption of Information from Online Reviews,” Journal of Travel Research , 53 ( 1), 44– 57. Google Scholar CrossRef Search ADS   Flanagin Andrew J., Metzger Miriam J. ( 2000), “Perceptions of Internet Information Credibility,” Journalism & Mass Communication Quarterly , 77 ( 3), 515– 40. Google Scholar CrossRef Search ADS   Forman Chris, Ghose Anindya, Wiesenfeld Batia ( 2008), “Examining the Relationship between Reviews and Sales: The Role of Reviewer Identity Disclosure in Electronic Markets,” Information Systems Research , 19 ( 3), 291– 313. Google Scholar CrossRef Search ADS   Goldman Alvin I., Jordan Lucy C. ( 2013), “Mindreading by Simulation: The Roles of Imagination and Mirroring,” in Understanding Other Minds: Perspectives from Developmental Social Neuroscience , ed. Baron-Cohen Simon, Lombardo Michael, Tager-Flusberg Helen, Oxford, UK: Oxford University Press, 448– 66. Google Scholar CrossRef Search ADS   Goldstein Noah J., Cialdini Robert B., Griskevicius Vladas ( 2008), “A Room with a Viewpoint: Using Social Norms to Motivate Environmental Conservation in Hotels,” Journal of Consumer Research , 35 ( 3), 472– 82. Google Scholar CrossRef Search ADS   Grohmann Bianca, Spangenberg Eric R., Sprott David E. ( 2007), “The Influence of Tactile Input on the Evaluation of Retail Product Offerings,” Journal of Retailing , 83 ( 2), 237– 45. Google Scholar CrossRef Search ADS   Groupon ( 2016), “$31 for $70 Worth of Services—The Laboratory Production Studios,” https://www.groupon.com/deals/the-laboratory-production-studios. Hayes Andrew F. ( 2013), Introduction to Mediation, Moderation, and Conditional Process Analysis , New York: Guilford. Hovland Carl, Janis Irving, Kelley Harold ( 1953), Communication and Persuasion , New Haven, CT: Yale University Press, 19– 48. Hu Nan, Zhang Jie, Pavlou Paul A. ( 2009), “Overcoming the J-Shaped Distribution of Product Reviews,” Communications of the Acm , 52 ( 10), 144– 7. Google Scholar CrossRef Search ADS   Huber Oswald, Seiser Gabriele ( 2001), “Accounting and Convincing: The Effect of Two Types of Justification on the Decision Process,” Journal of Behavioral Decision Making , 14 ( 1), 69– 85. Google Scholar CrossRef Search ADS   Huh Young Eun, Vosgerau Joachim, Morewedge Carey K. ( 2014), “Social Defaults: Observed Choices Become Choice Defaults,” Journal of Consumer Research , 41 ( 3), 746– 60. Google Scholar CrossRef Search ADS   Hyken Shep ( 2017), “Sixty-Four Percent of U.S. Households Have Amazon Prime,” Forbes , https://www.forbes.com/sites/shephyken/2017/06/17/sixty-four-percent-of-u-s-households-have-amazon-prime/#23a811844586/. Jin Liyin, He Yanqun, Zhang Ying ( 2014), “How Power States Influence Consumers’ Perceptions of Price Unfairness,” Journal of Consumer Research , 40 ( 5), 818– 33. Google Scholar CrossRef Search ADS   Johnson Eric J., Goldstein Daniel ( 2003), “Do Defaults Save Lives?” Science , 302 ( 5649), 1338– 9. Google Scholar CrossRef Search ADS   Kane Kat ( 2015), “The Big Hidden Problem with Uber? Insincere 5-Star Ratings,” Wired, http://www.wired.com/2015/03/bogus-uber-reviews/. Kaufman Douglas, Stasson Mark, Hart Jason ( 1999), “Are the Tabloids Always Wrong or Is That Just What We Think? Need for Cognition and Perceptions of Articles in Print Media,” Journal of Applied Social Psychology , 29 ( 9), 1984– 2000. Google Scholar CrossRef Search ADS   Khan Uzma, Kupor Daniella M. ( 2016), “Risk (Mis)Perception: When Greater Risk Reduces Risk Valuation,” Journal of Consumer Research , 43 ( 5), 769– 86. Krueger Joachim ( 2000), “The Projective Perception of the Social World: A Building Block of Social Comparison Processes,” in Handbook of Social Comparison: Theory and Research , ed. Suls Jerry, Wheeler Ladd, New York: Plenum/Kluwer, 323– 51. Google Scholar CrossRef Search ADS   Kuo Ying-Feng, Wu Chi-Ming, Deng Wei-Jaw ( 2009), “The Relationships among Service Quality, Perceived Value, Customer Satisfaction, and Post-Purchase Intention in Mobile Value-Added Services,” Computers in Human Behavior , 25 ( 4), 887– 96. Google Scholar CrossRef Search ADS   Kupor Daniella, Flynn Frank, Norton Michael ( 2017), “Half a Gift Is Not Half-Hearted: A Giver-Receiver Asymmetry in the Thoughtfulness of Partial Gifts,” Personality & Social Psychology Bulletin , 43 ( 12), 1686– 95. Google Scholar CrossRef Search ADS   Kupor Daniella M., Laurin Kristin, Levav Jonathan ( 2015), “Anticipating Divine Protection? Reminders of God Can Increase Nonmoral Risk Taking,” Psychological Science , 26 ( 4), 374– 84. Google Scholar CrossRef Search ADS   Kupor Daniella M., Tormala Zakary L. ( 2015), “Persuasion, Interrupted: The Effect of Momentary Interruptions on Message Processing and Persuasion,” Journal of Consumer Research , 42 ( 2), 300– 15. Kupor Daniella M., Tormala Zakary, Norton Michael I. ( 2014), “The Allure of Unknown Outcomes: Exploring the Role of Uncertainty in the Preference for Potential,” Journal of Experimental Social Psychology , 55 ( 6), 210– 6. Google Scholar CrossRef Search ADS   Kupor Daniella M., Tormala Zakary L., Norton Michael I., Rucker Derek D. ( 2014), “Thought Calibration: How Thinking Just the Right Amount Increases One’s Influence and Appeal,” Social Psychological and Personality Science , 5 ( 3), 263– 70. Google Scholar CrossRef Search ADS   Larson Lindsay RL, Denton L. Trey ( 2014), “ eWOM Watchdogs: Ego-Threatening Product Domains and the Policing of Positive Online Reviews,” Psychology & Marketing , 31 ( 9), 801– 11. Google Scholar CrossRef Search ADS   Lee Yi-Ching, Lee John D., Boyle Linda Ng ( 2007), “Visual Attention in Driving: The Effects of Cognitive Load and Visual Disruption,” Human Factors , 49 ( 4), 721– 33. Google Scholar CrossRef Search ADS   Lin Lin, Lee Jennifer, Robertson Tip ( 2011), “Reading While Watching Video: The Effect of Video Content on Reading Comprehension and Media Multitasking Ability,” Journal of Educational Computing Research , 45 ( 2), 183– 201. Google Scholar CrossRef Search ADS   McCroskey James ( 1966), “Scales for the Measurement of Ethos,” Speech Monographs , 33 ( 1), 65– 72. Google Scholar CrossRef Search ADS   McGinnies Elliott, Ward Charles D. ( 1980), “Better Liked than Right: Trustworthiness and Expertise as Factors in Credibility,” Personality and Social Psychology Bulletin , 6 ( 3), 467– 72. Google Scholar CrossRef Search ADS   MerchantCircle.com ( 2016), “M D Script,” http://www.merchantcircle.com/m-d-script-fall-city-wa. Miller Dale T. ( 1999), “The Norm of Self-Interest,” American Psychologist , 54 ( 12), 1053– 60. Google Scholar CrossRef Search ADS   Miller Dale T., Prentice Deborah A. ( 1996), “The Construction of Social Norms and Standards,” in Social Psychology: Handbook of Basic Principles , ed. Higgins E. Tory, Kruglanski Arie W., New York: Guilford, 799– 829. Murray Keith B. ( 1991), “A Test of Services Marketing Theory: Consumer Information Acquisition Activities,” Journal of Marketing , 55 ( 1), 10– 25. Google Scholar CrossRef Search ADS   Otterbacher Jahna ( 2008), “ Managing Information in Online Product Review Communities: Two Approaches,” ECIS  706– 717. Ong Beng Soo ( 2012), “The Perceived Influence of User Reviews in the Hospitality Industry,” Journal of Hospitality Marketing & Management , 21 ( 5), 463– 85. Google Scholar CrossRef Search ADS   Pelham Brett W., Sumarta Tin T., Myaskovsky Laura ( 1994), “The Easy Path from Many to Much: The Numerosity Heuristic,” Cognitive Psychology , 26 ( 2), 103– 33. Google Scholar CrossRef Search ADS   Petty Richard E., Cacioppo John T. ( 1984), “The Effects of Involvement on Responses to Argument Quantity and Quality: Central and Peripheral Routes to Persuasion,” Journal of Personality and Social Psychology , 46 ( 1), 69– 81. Google Scholar CrossRef Search ADS   Petty Richard E., Cacioppo John T., Schumann David ( 1983), “Central and Peripheral Routes to Advertising Effectiveness: The Moderating Role of Involvement,” Journal of Consumer Research , 10 ( 2), 135– 46. Google Scholar CrossRef Search ADS   Peysakhovich Alexander, Karmarkar Uma R. ( 2016), “Asymmetric Effects of Favorable and Unfavorable Information on Decision Making under Ambiguity,” Management Science , 62 ( 8), 2163– 78. Google Scholar CrossRef Search ADS   PrestaShop ( 2015), “Default Star Rating to 5 Instead of 3,” https://www.prestashop.com/forums/topic/449039-default-star-rating-to-5-instead-of-3/. Priester Joseph R., Petty Richard E. ( 1995), “Source Attributions and Persuasion: Perceived Honesty as a Determinant of Message Scrutiny,” Personality and Social Psychology Bulletin , 21 ( 6), 637– 54. Google Scholar CrossRef Search ADS   Priester Joseph R., Petty Richard E. ( 2003), “The Influence of Spokesperson Trustworthiness on Message Elaboration, Attitude Strength, and Advertising Effectiveness,” Journal of Consumer Psychology , 13 ( 4), 408– 21. Google Scholar CrossRef Search ADS   Pieters Rik ( 2017), “ Meaningful mediation analysis: Plausible causal inference and informative communication,” Journal of Consumer Research , 44( 3), 692– 716. Google Scholar CrossRef Search ADS   Printing Center USA ( 2018), https://www.printingcenterusa.com/ Probst C. Adam, Shaffer Victoria A., Chan Y. Raymond ( 2013), “The Effect of Defaults in an Electronic Health Record on Laboratory Test Ordering Practices for Pediatric Patients,” Health Psychology , 32 ( 9), 995– 1002. Google Scholar CrossRef Search ADS   Quora ( 2013), “Lyft (Company): How Critical Are You as a Passenger When Rating a Ridesharing Driver?” https://www.quora.com/Lyft-company-How-critical-are-you-as-a-passenger-when-rating-a-ridesharing-driver. Reich Taly, Kupor Daniella M., Smith Rosanna K. ( 2017), “Made by Mistake: When Mistakes Increase Product Preference,” Journal of Consumer Research , 44 ( 5), 1085– 1103. Reich Taly, Tormala Zakary L. ( 2013), “When Contradictions Foster Persuasion: An Attributional Perspective,” Journal of Experimental Social Psychology , 49 ( 3), 426– 39. Google Scholar CrossRef Search ADS   Reichelt Jonas, Sievert Jens, Jacob Frank ( 2014), “How Credibility Affects eWOM Reading: The Influences of Expertise, Trustworthiness, and Similarity on Utilitarian and Social Functions,” Journal of Marketing Communications , 20 ( 1–2), 65– 81. Google Scholar CrossRef Search ADS   Rosario Ana Babić, Sotgiu Francesca, De Valck Kristine, Bijmolt Tammo H. A. ( 2016), “The Effect of Electronic Word of Mouth on Sales: A Meta-Analytic Review of Platform, Product, and Metric Factors,” Journal of Marketing Research , 53 ( 3), 297– 318. Google Scholar CrossRef Search ADS   Ross Lee, Greene David, House Pamela ( 1977), “The ‘False Consensus Effect’: An Egocentric Bias in Social Perception and Attribution Processes,” Journal of Experimental Social Psychology , 13 ( 3), 279– 301. Google Scholar CrossRef Search ADS   Rucker Derek D., Petty Richard E., Briñol Pablo ( 2008), “What’s in a Frame Anyway? A Meta-Cognitive Analysis of the Impact of One versus Two Sided Message Framing on Attitude Certainty,” Journal of Consumer Psychology , 18 ( 2), 137– 49. Google Scholar CrossRef Search ADS   Rucker Derek D., Tormala Zakary L., Petty Richard E., Briñol Pablo ( 2014), “Consumer Conviction and Commitment: An Appraisal-Based Framework for Attitude Certainty,” Journal of Consumer Psychology , 24 ( 1), 119– 36. Google Scholar CrossRef Search ADS   Sakuraba Hisako ( 2012), “Do Simultaneously Presented Visual and Auditory Stimuli Attract Our Attention? The Effect of Divided Attention on Memory,” master’s thesis, Department of Psychological Science, University of Central Missouri, Warrensburg, MO 64093. Schlosser Ann E. ( 2011), “ Can including pros and cons increase the helpfulness and persuasiveness of online reviews? The interactive effects of ratings and arguments,” Journal of Consumer Psychology  21( 3), 226- 239. Google Scholar CrossRef Search ADS   Schrift Rom Y., Amar Moty ( 2015), “Pain and Preferences: Observed Decisional Conflict and the Convergence of Preferences,” Journal of Consumer Research , 42 ( 4), 515– 34. Siegler M. G. ( 2009), “YouTube Comes to a 5-Star Realization: Its Ratings Are Useless,” TechCrunch, https://techcrunch.com/2009/09/22/youtube-comes-to-a-5-star-realization-its-ratings-are-useless/. Simonson Itamar ( 1989), “Choice Based on Reasons: The Case of Attraction and Compromise Effects,” Journal of Consumer Research , 16 ( 2), 158– 74. Google Scholar CrossRef Search ADS   Stephen Andrew T., Grewal Lauren ( 2015), “In Mobile We Trust: How Mobile Reviews Can Overcome Consumer Distrust of User-Generated Reviews,” in NA-Advances in Consumer Research , Vol. 43, ed. Diehl Kristin, Yoon Carolyn, Duluth, MN: Association for Consumer Research, 117– 21. Tormala Zakary L., Clarkson Joshua ( 2008), “Source Trustworthiness and Information Processing in Multiple Message Situations: A Contextual Analysis,” Social Cognition , 26 ( 3), 357– 67. Google Scholar CrossRef Search ADS   Tormala Zakary L., Petty Richard E., Briñol Pablo ( 2002), “Ease of Retrieval Effects in Persuasion: A Self-Validation Analysis,” Personality and Social Psychology Bulletin , 28 ( 12), 1700– 12. Google Scholar CrossRef Search ADS   Wangenheim Florian, Bayón Tomás ( 2004), “The Effect of Word of Mouth on Service Switching,” European Journal of Marketing , 38 ( 9/10), 1173– 85. Google Scholar CrossRef Search ADS   Willemsen Lotte M., Neijens Peter C., Bronner Fred ( 2012), “ The ironic effect of source identification on the perceived credibility of online product reviewers,” Journal of Computer-Mediated Communication , 18( 1), 16– 31. Google Scholar CrossRef Search ADS   Wood Stacy, McInnes Melayne Morgan, Norton David A. ( 2011), “The Bad Thing about Good Games: The Relationship between Close Sporting Events and Game-Day Traffic Fatalities,” Journal of Consumer Research , 38 ( 4), 611– 21. Google Scholar CrossRef Search ADS   Yelp ( 2015), “PeakFit Strength and Conditioning,” https://www.yelp.com/biz/peakfit-strength-and-conditioning-san-dimas-2. Zhang Yan, Epley Nicholas ( 2012), “Exaggerated, Mispredicted, and Misplaced: When ‘It’s the Thought That Counts’ in Gift Exchanges,” Journal of Experimental Psychology: General , 141 ( 4), 667– 81. Google Scholar CrossRef Search ADS   Zillow Advice Thread ( 2016), https://webcache.googleusercontent.com/search?q=cache:NF-gzkL-6jAJ:https://www.zillow.com/advice/PA/pro-to-pro-agents/question-discussion-guide/+&cd=1&hl=en&ct=clnk&gl=us. © The Author(s) 2018. Published by Oxford University Press on behalf of Journal of Consumer Research, Inc. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal

Journal of Consumer ResearchOxford University Press

Published: Mar 14, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off