Unveiling Stars: How Graphical Displays of Online Consumer Ratings Affect Consumer Perception and Judgment

With the burgeoning of e-commerce, nearly everything is reviewed on third-party platforms and e-commerce websites, offering consumers access to a vast array of online word of mouth (WOM). However, given the enormous number of reviews, it is impossible for consumers to read and fully incorporate all the reviews that a product or service receives into their decisions. Review platforms, therefore, extract meaningful summary information (e.g., products’ average rating, the number of reviews, and rating distributions) to facilitate decision making. While most platforms offer the same types of summary information, this information is displayed in a variety of visual formats, such as rating distribution bar graphs in different forms, orientations, colors, and sizes and the weighted averages in stars, bubbles, or squares.

Prior studies have extensively investigated the impact of the various attributes of online WOM, including the textual content of reviews, the number of reviews, average ratings, and their variance, on consumer decision making (e.g., Chevalier and Mayzlin 2006; Rocklage and Fazio 2020; Schoenmueller, Netzer, and Stahl 2020; Watson, Ghosh, and Trusov 2018). However, the research corpus examining how the visual display of summary information affects consumers’ evaluation of target products is quite limited (e.g., Fisher, Newman, and Dhar 2018; Rozenkrants, Wheeler, and Shiv 2017).

The primary question we examine is whether online rating distributions displayed in different formats influence consumers’ product evaluations differently. Rating distribution graphs used by various platforms differ on multiple attributes, such as the color of the bars, rating labels (e.g., star rating scale vs. descriptive labeling), how bar values are labeled (e.g., percentage vs. frequencies), and the range of the x-axis (full range vs. a truncated range). While all these differences can affect consumers’ decisions, this research focuses on the role of the x-axis (i.e., whether the x-axis has a full or truncated range) in shaping consumers’ judgments as the range of the x-axis affects the reference points against which bar values are displayed and therefore perceived. Specifically, we examine the impact of two prominent classes of distribution display formats: proportional bar graphs (x-axes range from zero to 100% of the ratings; Figure 1, Panel A) and simple bar graphs (x-axes range from zero to the share of the most frequent rating score, i.e., the peak value; Figure 1, Panel B). Note that “proportional” and “simple” refer only to the graphical aspect of the ratings (i.e., the x-axis) and do not describe value labels in any way. Figure 2 shows simplified versions of these two graphical formats.

OPEN IN VIEWER

Figure 1.

Proportional and Simple Bar Graph Examples.

OPEN IN VIEWER

Figure 2.

A Typical Rating Distribution Displayed in Proportional and Simple Bar Graphs.

The graphical presentation aspect of online WOM is crucial for at least three reasons: (1) previous research has shown that consumers rely heavily on rating distributions in their decision making (Fisher, Newman, and Dhar 2018; He and Bond 2015; Rozenkrants, Wheeler, and Shiv 2017), (2) surveys have shown that the rating distribution summary is the most utilized feature of online WOM by consumers (Baymard 2017), and (3) unlike many other attributes, the presentation aspects of online WOM summary information are fully controlled by the review platforms and can be changed at relatively negligible cost.

As Figure 2 shows, the fundamental difference between a proportional bar graph and a simple one is that in the former (Figure 2, Panel A), the x-axis ranges from 0% to 100% of the ratings, resulting in a part-to-whole portrayal of the bar values. In the latter (Figure 2, Panel B), however, the x-axis ranges from 0% to peak value (i.e., 53% of the ratings), resulting in a simpler graphical presentation of the bar values. Therefore, the full-range x-axis in a proportional bar graph provides two fixed reference points—0% and all (100%) of the ratings—against which the magnitude of each bar can be visually assessed. The value of each star rating bar in a proportional bar graph is shown using two areas. In one area, the proportion of people who have given a specific rating score (e.g., five stars) is highlighted. In the other area, the space between the end of the highlighted section and the right end of the graph, shows those who have not assigned that rating, making part-to-whole relationships immediately graphically apparent.

In contrast, the truncated-range x-axis in a simple bar graph offers one fixed reference point (i.e., 0%) and one variable reference point, the peak value (i.e., the share of the most frequent rating against which the value of each rating bar can be visually judged). Since the second reference point in a simple bar graph is always smaller than 100% of the ratings, a simple bar graph displays rating bars as individual values rather than as parts of a whole.

Drawing on visual and graphical perception research (e.g., Cleveland and McGill 1985; Fischer 2000; Hegarty 2011; Lurie and Mason 2007; Talbot, Setlur, and Anand 2014), we contend that consumers evaluate an item more favorably when its rating distribution is displayed in a simple rather than proportional format. Specifically, we draw attention to the unique aspect of simple bar graphs: a truncated response range for each star rating bar. We propose that simple bar graphs elicit a more extreme response integrated toward the end point of the response range (i.e., peak value) relative to proportional bar graphs (i.e., 100%). This occurs because prior research has shown that consumers rely on visual reference points to assess numerical information and make sense of graphical stimuli (Talbot, Setlur, and Anand 2014; Thomas and Kyung 2019), and the visual reference point offered by a simple bar graph (i.e., the peak value) is usually smaller than that offered by a proportional bar graph (i.e., 100%).

In other words, any data point displayed on the axis representing the variable of interest (i.e., the x-axis) would be assessed with respect to two reference points: the baseline (or the minimum value that can be displayed in the graph, e.g., x = 0%) and the largest value that can potentially be displayed (e.g., x = 100%). In both formats, the baseline values are the same (i.e., 0%). However, the largest values for which a graphical representation exists in a simple bar graph (i.e., the peak value) and in a proportional bar graph (i.e., 100%) are not equal. Our premise is that this seemingly minimal difference in the reference points of the two graphical display formats is enough to substantially and positively influence the evaluation of the products whose ratings distributions are presented in a simple bar graph.

The current research has important theoretical and managerial implications. First, it extends the literature by studying a neglected feature of online WOM. Although prior research has suggested that presenting rating distributions in graphical displays affects consumers’ evaluation of ratings (e.g., Fisher, Newman, and Dhar 2018; He and Bond 2015), it is still unclear how various graphical representations of ratings may affect consumers’ interpretations of ratings. The current research scrutinizes the visualization aspect of online ratings to provide valuable insights into how and why rating distributions presented in various graphical formats may engender different outcomes. Second, based on the finding that the graphical format of rating distributions is an essential factor in forming consumers’ judgments of online ratings, we present recommendations for how online review platforms and businesses can improve their performance. Third, by demonstrating how different visual representations of ratings affect consumers’ perceptions of those ratings, the findings of the current research contribute to graphical perception research.

In the following sections, we discuss the relevant literature and develop our conceptual framework and specific hypotheses. We then report the findings of six studies designed to test those hypotheses and the underlying processes. Finally, we discuss the theoretical and managerial implications of this work and some avenues for future research.

Conceptual Background

Displaying rating distributions graphically, as opposed to presenting statistics such as average ratings, can lead to very different outcomes (e.g., Fisher, Newman, and Dhar 2018; He and Bond 2015; Rozenkrants, Wheeler, and Shiv 2017). Prior research, however, has focused mainly on how consumers incorporate distribution summary information into their decisions without studying the differential impact of distributions’ various presentation formats. For instance, consumers tend to engage with rating distributions via categorization, summation, and deduction processes rather than via the mere extraction of information (Fisher, Newman, and Dhar 2018). Consumers use distribution graphs to infer the level of dispersion (or consensus) in ratings, which influences their evaluations (Khare, Labrecque, and Asare 2011). Moreover, the same rating distribution affects consumers’ decisions and choices differently depending on the product category (He and Bond 2015) or situational factors such as the activation of self-expression goals (Rozenkrants, Wheeler, and Shiv 2017).

Utility maximization models suggest that the distribution of a set of ratings should generate equivalent evaluations, regardless of its presentation format (Tversky and Kahneman 1986). Empirical evidence (e.g., Kahneman and Tversky 1979; Stone, Yates, and Parker 1997), however, shows that varied representations of the same information highlight different aspects and thus alter cognitive processes, influencing task performance (Bettman and Kakkar 1977; Kim and Lakshmanan 2021) and consumer preferences (Bettman, Luce, and Payne 1998). Advancing this line of reasoning, we predict that the truncated x-axis in a simple bar graph affects consumers’ cognitive processes and related outcomes. For deeper insight into the underlying processes, we examine the two graphical formats through the lens of visual perception to develop our hypotheses.

Visual Perception of Graphical Displays

Graphs comprise common structural components (Kosslyn 1989), such as the framework of a graph, specifiers, and so forth. Simkin and Hastie (1987) argue that to understand how these structural components are perceived in the human mind, we must ascertain the mental processes that affect (1) how visual components are represented in the mind, (2) how representations help make inferences, and (3) how these representations are integrated within a context to define a response.

Early theories of graph perception (Cleveland 1985; Cleveland and McGill 1984, 1985) identified several graphical perception tasks essential for decoding a display design's visual elements to form a judgment. These tasks are ordered in descending level of known accuracy. Building on this taxonomy, Simkin and Hastie (1987) argued that it is not only the display design elements that influence perceptual judgment but also the task at hand. For instance, when people were asked to make comparison judgments, bar charts showed the best comprehension performance; for proportion judgments, though, participants were most accurate with pie charts.

To elaborate on the role of context in judgment tasks, Gillan and Lewis (1994) argue that people apply certain processes such as looking and encoding for spatial locations of specifiers (e.g., using axes and labels associated with them). Relatedly, Carpenter and Shah (1998) argue that a key aspect of graph comprehension is how people relate graph features to their referents by both finding the patterns and reading and rereading information from the axes (e.g., visual reference points) and other regions of a graph. Accordingly, research on bar graphs has shown that even small changes such as adding a random bar, introducing irrelevant visual cues, displaying bars separately or adjacent to each other, and using aligned or unaligned baselines can significantly affect the perception of the focal information (Fischer 2000; Talbot, Setlur, and Anand 2014; Zacks et al. 1998).

Consumers’ Processing of Ratings Distribution and Product Evaluation

Based on these findings, we argue that a key feature in comprehending review bar graphs is the use of the start points and end points of the bars as references. Whereas the start points for both simple and proportional bar graphs are fixed at 0%, as indicated previously, the end points differ across simple and proportional bar graphs. In proportional bar graphs, the end point is fixed and always refers to the scale maximum (i.e., 100% of ratings), which enables consumers to notice the lack of specific ratings in a rating distribution. In simple bar graphs, however, the end point varies with the distribution's peak value (i.e., the share of the most frequent rating score). The integration of this smaller and variable end reference point that eliminates the graphical cues representing those who did not give a specific rating (say, a five-star rating) can systematically and substantially affect consumers’ evaluation of the items whose ratings are presented in simple bar graphs.

We contend that the truncated x-axis in simple bar graphs inflates consumers’ perception of ratings by reducing the reference point against which the bars are compared. As described previously, when consumers face proportional bar graphs to evaluate potential purchases, they assess the magnitude of each bar—especially the four- and five-star bars, the most crucial bars according to the positivity imbalance literature (e.g., Hu, Pavlou, and Zhang 2017; Moe, Netzer, and Schweidel 2017; Schoenmueller, Netzer, and Stahl 2020)—relative to the full response range by considering the visual distances of the highlighted section from the start point (0%) and the end point of the response range (100%). Thus, the two reference points with which a bar is compared are fixed at 0% and 100% in proportional bar graphs. In contrast, when consumers see simple bar graphs, they assess the magnitude of each bar relative to a truncated response range, as the peak value of the bar is generally less than 100%. In other words, because of this propensity to use visual distance from the endpoints (Thomas and Kyung 2019), the assessment of bar graphs is affected by the second reference point, which is not fixed in simple bar graphs. Therefore, the shift in the visual reference point against which the bars are assessed from 100% to the peak value in simple bar graphs can result in a systematically higher evaluation of the target product.

Let us illustrate our reasoning with a concrete example. Consider the distribution displayed in Figure 2, where the ratings are distributed as follows: 53% (five-star), 34% (four-star), 1% (three-star), 2% (two-star), and 10% (one-star). In the proportional format, the length of each bar matches its value. For instance, the highlighted portion of the five-star bar takes 53% of a full bar, offering a visual representation for 53% of the ratings that are five-star as well as a visual representation for the 47% of the ratings that are not five-star in the same bar. In the simple format, however, the peak value (i.e., 53%) will be the reference point because 53% is the largest value for which a visual representation exists in the graph. Therefore, the bar associated with this value is displayed as a full bar as if it has taken the maximum possible value. In other words, the bar associated with the most frequent rating score in a distribution (i.e., the tallest bar) will always fill 100% of the available space in a simple bar graph. This shift in reference point—from 100% in the proportional bar graph to 53% in the simple bar graph—against which each bar is visualized distorts the graphical representations of all bars. Thus, the five-star bar will be displayed as 53/53 = 100% of the reference point, and four-, three-, two-, and one-star bars will take 34/53 = 64%, 1/53 = 2%, 2/53 = 4%, and 10/53 = 19% of the graph's reference point, respectively, inflating the perception of all ratings proportionate to the peak value (i.e., 53%).

Prior research (e.g., Hu, Pavlou, and Zhang 2017; Moe, Netzer, and Schweidel 2017; Schoenmueller, Netzer, and Stahl 2020) as well as real-world data show that online consumer ratings are overwhelmingly positive. For instance, half of all ratings on Amazon are five-star ratings (Rocklage, Rucker, and Nordgren 2021). This positivity problem, however, is not unique to Amazon. According to Yelp (2022), over 68% of all ratings on Yelp are either four- or five-star, leading to an average rating of at least 3 out of 5 for over 80% of businesses. Given this significant imbalance in online ratings, we suggest that the inflation in the perceptual representation of ratings caused by the shift in the reference point from 100% to the peak value in simple bar graphs will lead to an inflated perception of positive ratings and, therefore, a higher evaluation of the target item. Formally, we propose the following:

H₁: Displaying rating distributions in simple versus proportional bar graphs (i.e., shrunken vs. full x-axis) leads to a higher evaluation of the target item.

The moderating role of the peak value

Using simple (vs. proportional) bar graphs does not distort the graphical representations of all distributions equally. As mentioned previously, the visual representations of all values (the highlighted bars) are inflated by the inverse of the peak value (i.e., 1/53% = 1.89 in our previous example). The average rating of the distribution used in this example (i.e., Figure 2) is 4.2 out of 5. Now consider another set of ratings (76% five-star, 0% four-star, 5% three-star, 6% two-star, and 13% one-star) with the same mean (4.2 out of 5) displayed in Figure 3. Compared with the distribution in Figure 2, using a simple bar graph to display the distribution in Figure 3 results in a smaller change in reference points from 100% to 76% (vs. from 100% to 53%). Shifting the reference point of a distribution graph from 100% to 76% inflates the visual representations of all values by 1/76% = 1.32, which is smaller than 1/53% = 1.89. Therefore, a distribution's peak value determines the extent to which using a simple bar graph influences the visual representations of the bars such that smaller peak values lead to greater shifts in reference points and, therefore, higher levels of visual distortion compared with proportional bar graphs.

OPEN IN VIEWER

Figure 3.

A Typical Distribution Displayed in Proportional and Simple Bar Graphs.

This implies that the extent to which a simple bar graph visually inflates ratings decreases as the peak value approaches 100%, suggesting that displaying rating distributions that have smaller (vs. larger) peak values in simple (vs. proportional) bar graphs should lead to a stronger positive effect. In other words, we propose:

H₂: As the peak value increases, the positive effect of displaying rating distributions in simple bar graphs (as described in H₁) becomes weaker.

The moderating role of the imbalance score

Given that simple bar graphs enlarge both negative and positive ratings, a reasonable expectation is that the inflation in negative ratings would outweigh the inflation in positive ratings, leading to lower evaluations of simple bar graphs. This is unlikely to be the case for most distributions because even though both positive and negative ratings are inflated in simple bar graphs, their perceived difference is also exaggerated by the same ratio. In other words, because of the truncated x-axis, the perceptual difference between positive and negative ratings (known as the imbalance score) is higher when ratings are displayed in a simple (vs. proportional) bar graph. Consider the distribution in Figure 2. While the perceptual difference between positive (i.e., five-star and four-star) and negative (i.e., two-star and one-star) ratings in a proportional bar graph is 75% of the graph's reference point (i.e., 75% of a full bar; 53% + 34% − 2% − 10% = 75%), this difference is vastly inflated in a simple bar graph, at 142% of a full bar (53/53 + 34/53 − 2/53 − 10/53 = 142%).

Although the imbalance score can be thought of as a mathematical score, Fisher, Newman, and Dhar (2018) suggest that it is a psychological construct and find that this construct positively affects consumers’ evaluation of an item even after controlling for the effect of statistical characteristics such as the mean and variance. We suggest that because simple bar graphs enhance the perceptual difference between positive and negative ratings, the positive impact of using simple (vs. proportional) bar graphs on consumers’ evaluations should be stronger for distributions with higher (vs. lower) imbalance scores.

For instance, consider the two distributions displayed in Figure 4 with the same means (3 out of 5), variances (1.72), and peak values (52%) but different imbalance scores (−17% vs. 17%). While displaying these distributions in simple (vs. proportional) bar graphs distorts the visual representations of both distributions similarly by shifting the graph's reference points from 100% to 52%, it is apparent that it does not affect consumers’ evaluations equally.

OPEN IN VIEWER

Figure 4.

Distributions with the Same Means, Peak Values, and Variances but Different Imbalance Scores.

As previously discussed, while the perceptual difference between the proportion of positive ratings and the proportion of negative ratings is unbiased in proportional bar graphs, it is inflated in simple bar graphs. This means that the −17% (17%) imbalance score of distribution A (distribution B) in Figure 4 has an unbiased graphical representation of −17% (17%) of the graph's reference point when displayed in the proportional bar graph format. When the same distributions are displayed in simple bar graphs, however, the graphical representation of the −17% (17%) imbalance score is equivalent to −33% (33%) of the simple bar graph visual reference points (i.e., ±17/52 = ±33%). Therefore, using a simple (vs. proportional) bar graph to display distribution A in Figure 4, whose imbalance score is −17%, can even have a negative rather than positive effect on consumers’ evaluations. In contrast, since distribution B in Figure 4 has a positive imbalance score (i.e., 17%), displaying it in a simple (vs. proportional) bar graph should lead to higher evaluations, suggesting that the effect of using simple (vs. proportional) bar graphs described in H₁ should also be moderated by the distribution's imbalance score.

We therefore formally propose the following:

H₃: The positive impact of displaying rating distributions in simple (vs. proportional) bar graphs on product evaluation (H₁) is stronger for distributions with high (vs. low) imbalance scores.

The Graphical Format of Prior Ratings and Its Impact on Future Ratings

The previous sections discussed how different graphical presentations of the same ratings may influence consumers’ preconsumption evaluation of an item. Another important question regarding the effect of a distribution's graphical format is whether displaying prior ratings in different forms would affect an item's future ratings (i.e., consumers’ postconsumption rating behavior). This is a crucial question because consumers can see an item's previous ratings when visiting a review platform's website to submit their ratings. The findings of several research streams suggest that consumers’ perceptions of the ratings posted for a product can influence the product's future ratings. In the following section, we discuss the findings of two research streams (i.e., the anchoring effect literature and prior research on online ratings) to determine how displaying ratings in simple and proportional bar graphs may affect consumers’ postconsumption ratings.

The anchoring effect is a robust heuristic in which an individual's judgments are impacted by a specific reference point or anchor (Furnham and Boo 2011). This form of cognitive heuristic occurs when an individual's evaluation of a stimulus is impacted by a value (anchor) to which they are exposed prior to their assessment. Tversky and Kahneman (1974) suggest that estimates are biased toward the anchor value because individuals do not make sufficient downward or upward adjustments when exposed to anchors. The confirmatory hypothesis testing explanation, however, suggests that individuals find the anchor value to be realistic and thus activate the target's characteristics that are consistent with the anchor to make their answers comparable (Furnham and Boo 2011; Mussweiler and Strack 1999).

The prevalence of the anchoring effect is shown in various contexts, such as consumers’ purchase decisions (Wansink, Kent, and Hoch 1998), payment decision making (Jung, Perfecto, and Nelson 2016), and online consumer ratings (Moe and Schweidel 2012; Schlosser 2005; Wang et al. 2022). Specifically, prior research has demonstrated that reviewers’ exposure to a product's previous ratings and reviews can affect how they rate their experience with that product. For example, Schlosser (2005) finds that being exposed to negative opinions about a product may activate reviewers’ concern regarding the social outcomes of their evaluations (e.g., being judged as having low standards in case of positive evaluations), therefore leading them to adjust their evaluations downward and give the product lower ratings. Moe and Schweidel (2012), however, show that less frequent posters tend to adjust their ratings upward when other ratings are mainly positive. In more recent work, Wang et al. (2022) report that a product's prior average rating positively affects its subsequent ratings.

We argue that presenting rating distributions in different graphical formats can influence a product's future ratings in a comparable manner. To the extent that a distribution's graphical format influences consumers’ perception of an item's prior ratings, depending on its graphical format, the same rating distribution can act as a high or low anchor when consumers provide a star rating expressing their judgment of that item. Therefore, we formally predict the following:

H₄: Displaying a product's previous ratings in a simple (vs. proportional) bar graph leads to more favorable future ratings.

Overview of Studies

We report the findings of six studies in the current manuscript and three in the Web Appendices. The purpose and summary of each study are presented in Table 1 and fully discussed in the text. A few noteworthy aspects of the studies are as follows. All studies were done with online participants at Amazon Mechanical Turk (MTurk), CloudResearch, and Prolific. Only those participants who had complied with the instructions were included (Web Appendix A). We employed a large pool of distributions in various studies. In each experiment, a between-subjects design was used for the graphical format, with at least two formats being included. In studies measuring product evaluations, we administered (in randomized order) Fisher, Newman, and Dhar’s (2018) four-item, ten-point Likert scale with the following items:

How do you feel about this product/service? (1 = “unfavorable,” and 10 = “favorable”)

How would you expect your experience of this product/service to be? (1 = “very negative,” and 10 = “very positive”)

How likely would you be to buy this product/service? (1 = “very unlikely,” and 10 = “very likely”)

How much would you be willing to pay for this product/service? (1 = “not a lot,” and 10 = “very much”)

OPEN IN VIEWER

Table 1.

Summary of Studies.

Study	Purpose	Number of Distributions (Stimuli)a	Sample Size	Main Dependent Variable
Study 1	Testing H1, H2, and H3	200 (400)	1,192	Product evaluation
Study 2	Testing the effect of the graphical format on perception of ratings	3 (6)	284	Estimation and estimation errors
Study 3	Testing alternative explanations	21 (126)	607	Product evaluation
Study 4	Testing the effect of the graphical format in the presence of competing information and different numerical information formats	21 (105)	400	Product evaluation
Study 5	Testing H4	1 (2)	921	Website's postconsumption star rating
Study 6	Providing field evidence supporting H4	—	—	Restaurant star ratings

^aMultiple versions (i.e., formats) of each distribution were created depending on the number of conditions. All stimuli can be accessed at https://osf.io/h2nzu/?view_only=cfb7fab2e9624ee8aa6fdbf898bbb9d8.

Study 1: Testing the Effect of the Graphical Format on Product Evaluation

Study 1 aimed to test the effect of the graphical format on consumer evaluation (H₁) in a controlled setting. The additional objectives of this study were to test the interaction effects between the graphical format and (1) peak value and (2) imbalance score described in H₂ and H₃. Per our conceptualization, the peak value of a distribution determines the extent to which using a simple bar graph distorts its visual representation. In addition, our conceptualization predicts that the positive effect of using a simple (vs. proportional) bar graph is moderated by the distribution's imbalance score (H₃).

Method

We recruited 1,192 participants (44% female, 56% male; M_age = 37.29 years) from MTurk in return for a small payment. This study involved a 2 (graphical format: simple vs. proportional bar graph, as displayed in Figure 4) × 2 (product: laptop vs. stapler) × 10 (rating distributions) mixed design. The first two factors varied between subjects, and the third varied within subjects. Each participant saw ten randomly selected distribution graphs in a randomized order and was told that each distribution graph displayed a product's ratings. We measured product evaluation using the previous four-item scale (Fisher, Newman, and Dhar 2018). Since previous research suggests that any conclusion about graphical displays of information should be based on a large number of stimuli (Hegarty 2011), our distribution pool consisted of 200 randomly selected distributions with an average rating between 3 and 4.8 out of 5 from 130,000 randomly generated distributions (see Web Appendix B), resulting in 400 stimuli. Distributions were selected to have an average rating of at least 3 because, as discussed previously, research and field data have shown that the majority of real-world ratings are positive; in fact, the average star rating in most industries is higher than 3.8 (ReviewTrackers 2019). Such a distribution pool was created also because 90% of consumers do not even consider a product whose average rating is below 3.0 (BrightLocal 2018). Participants saw only a rating distribution displayed in a simple or proportional graph for each product evaluation task, as shown in Figure 4. We used staplers and laptops as target products to test the effect of the graphical format on both small and more significant purchases.

Results

We used a linear mixed-effects regression model with product evaluation as the outcome variable and graphical format (−1 = proportional bar graph and 1 = simple bar graph; H₁), participants’ mean-centered average rating, variance, peak value, graphical format × peak value (H₂), the imbalance score and its interaction with the graphical format (H₃), and the product (−1 = staple and 1 = laptop) and its interaction with the graphical format as fixed effects. To test our hypotheses, we followed Hamaker and Muthén's (2020) recommendations to disentangle between-subjects and within-subjects slopes and interactions. We included variance in our model because research has found a negative impact of variance on consumers’ evaluations (Sun 2012). Due to the mixed design of this study, we included random intercepts for participants. Given the importance of the sampling of the stimuli (Judd, Westfall, and Kenny 2012), we also included random intercepts for the rating distributions that were used.

We used the following equation to model how subject i evaluated stimulus j:

Evaluation ij = γ 00 + β 1 Graphical format i + β 2 ( Rating ij − Rating ¯ i ) + β 3 ( Variance ij − Variance ¯ i ) + β 4 ( Peak value ij − Peak value ¯ i ) + β 5 ( Imbalance score ij − Imbalance score ¯ i ) + β 6 Graphical format i × ( Peak value ij − Peak value ¯ i ) + β 7 Graphical format i × ( Imbalance score ij − Imbalance score ¯ i ) + β 8 Product i + β 9 Graphical format i × Product i + u 0 i + δ j + e ij ,

(1)

where u_0i and δ_j represent between-subjects level and between-stimuli level random intercepts, respectively. Moreover, Rating ¯ i , Variance ¯ i , Peak value ¯ i , and Imbalance score ¯ i represent the means for the average rating, variance, peak value, and imbalance score of the rating distributions evaluated by subject i. By mean-centering continuous independent variables for each participant (i.e., using I V ij − I V i ¯ instead of I V ij ), β₁ represents the main effect of the graphical format (H₁) when all continuous independent variables (i.e., average rating, variance, peak value, and imbalance score) are equal to their means at the within-subjects level (i.e., I V ij − I V i ¯ = 0 ) . In this equation, β₂, β₃, β₄, and β₅ test the within-subjects effects of the average rating, variance, peak value, and imbalance score on the product evaluation, respectively. Furthermore, β₆ and β₇ enable us to test H₂ and H₃. While β₆ indicates how the effect of the graphical format (β₁) is influenced by the distribution's peak value, β₇ shows how the imbalance score moderates this effect. Further, a negative β₆ would suggest that an increase in the distribution's peak value mitigates the positive effect of simple bar graphs (H₂), and a positive β₇ would suggest that the effect of the graphical format is stronger for distributions with higher imbalance scores (H₃).

The mixed-effects regression results are displayed in Table 2. As proposed, displaying ratings in simple bar graphs led to higher product evaluations (β₁ = .14, SE = .03, p < .001, 95% CI = [.08, .20]). We also found that the effect of peak value on product evaluation was positive (β₄ = 1.42, SE = .19, p < .001, 95% CI = [1.07, 1.80]). As predicted in H₂, as the distribution's peak value at the within-subjects level increased, the positive impact of the graphical format on the evaluation weakened (β₆ = −.48, SE = .08, p < .001, 95% CI = [−.64, −.33]). We conducted Johnson–Neyman floodlight analysis to identify the regions where simple bar graphs led to higher evaluations. The mean-centered peak values at the within-subjects level ranged from −.41 to .41, and the effect of the graphical format was significant from −.41 to .15, spanning 68% of the range (Figure 5).

OPEN IN VIEWER

Figure 5.

Moderating Roles of the Peak Value and Imbalance Score in Study 1.

OPEN IN VIEWER

Table 2.

Statistical Summary of the Mixed-Effects Regression Model in Study 1.

Coefficients	Estimate	SE	t-Value	p-Value	95% CI
γ00 (intercept)	6.04	.04	160.59	<.001	5.97	6.13
β1 (H1: graphical format)	.14	.03	4.48	<.001	.08	.20
β2 (average rating)	1.72	.11	15.85	<.001	1.49	1.93
β3 (variance)	.13	.04	2.97	.003	.04	.21
β4 (peak value)	1.42	.19	7.58	<.001	1.07	1.80
β5 (imbalance score)	1.54	.20	7.70	<.001	1.17	1.96
β6 (H2: graphical format × peak value)	−.48	.08	−5.94	<.001	−.64	−.33
β7 (H3: graphical format × imbalance score)	.19	.04	4.37	<.001	.10	.28
β8 (product)	−.08	.03	−2.46	.014	−.14	−.01
β9 (graphical format × product)	−.02	.03	−.54	.590	−.08	.04
Grouping Variable	Random Effects (SD)
Subject	.98
Stimulus	.31
AIC	41,519

We operationalized the imbalance score by calculating the difference between the proportion of positive (i.e., sum of four- and five-star) and negative (i.e., sum of one- and two-star) ratings. Our analysis supported the positive impact of the imbalance score on product evaluations reported in previous research (β₅ = 1.54, SE = .20, p < .001, 95% CI = [1.17, 1.96]). More importantly, as predicted in H₃, the positive effect of displaying ratings in simple bar graphs was stronger for distributions with greater imbalance scores (β₇ = .19, SE = .04, p < .001, 95% CI = [.10, .28]). Consistent with prior research, we also found that the average rating had a positive impact on evaluations (β₂ = 1.72, SE = .11, p < .001, 95% CI = [1.49, 1.93]). While we expected a negative within-subjects effect of variance, higher variance led to higher product evaluations (β₃ = .13, SE = .04, p = .003, 95% CI = [.04, .21]), possibly because the peak value and imbalance score captured its negative effect (variance had a negative effect on evaluation in the absence of peak value and imbalance score). Moreover, while laptops were evaluated less positively than staplers (β₈ = −.08, SE = .03, p = .014, 95% CI = [−.14, −.01]), the effect of the graphical format was not moderated by product type (β₉ = −.02, SE = .03, p = .590, 95% CI = [−.08, .04]).

In sum, this study provided support for H₁, H₂, and H₃ by demonstrating the positive effect of displaying ratings in the simple format for an extensive set of distributions (H₁) and the moderating roles of the peak value (H₂) and imbalance score (H₃). We further showed that this effect holds for both small and more significant purchases.

Study 2: Consumers’ Perceptions of Ratings (Value Estimation Tasks)

Study 1 showed that consumers evaluate products more positively when rating distributions are displayed in simple bar graphs and that this effect is moderated by the distribution's peak value and imbalance score. We propose that this occurs because the shrunken x-axis in simple bar graphs inflates consumers’ perceptions of positive ratings when the distributions have small peak values. While the negative interaction effect between the peak value and the graphical format supports this theorizing, Study 1 did not provide direct evidence that consumers’ perceptions of ratings are influenced by the graphical format and the distribution's x-axis range. Thus, this study aimed to test the effects of the graphical format on the perceptions of ratings for distributions with different peak values.

Method

Participants (N = 284; 63% female, 35% male, 1% other; M_age = 37.23 years) were recruited via Prolific. In the instruction task, they were informed that they would estimate the number of consumers who had given the product five-star scores. In a mixed-design experiment, they were randomly assigned to simple or proportional bar graph conditions, where each viewed three distributions displayed in Figure 6 in a randomized order and entered their best and quick estimates of the number of five-star ratings for each distribution.

OPEN IN VIEWER

Figure 6.

Stimuli Used in Study 2.

Results

We calculated the estimation errors (as percentages) by subtracting true values (i.e., 118, 725, or 1,166) from participants’ estimated values and dividing them by the total number of ratings displayed in the distribution (i.e., 1,180, 1,250, or 1,310). Since each participant estimated three values, we conducted a two-way mixed analysis of variance (ANOVA) to test the effects of the graphical format, stimulus (a three-level within-subjects factor representing the distributions with various peak values shown in Figure 6), and their interactions on the estimate errors. Our analysis revealed that both graphical format (F(1, 282) = 14.89, p < .001, η p 2  = .05) and stimulus (F(2, 564) = 110.96, p < .001, η p 2  = .28) had significant effects on the estimation errors, which were qualified with a significant interaction (F(2, 564) = 10.28, p < .001, η p 2  = .04).

Given that the shift in reference points for distributions with high peak values is small, we expected participants who viewed simple (vs. proportional) bar graphs to overestimate the target value only when the distribution had a small peak value (i.e., distributions A and B). Consistent with this expectation, we found that participants in the simple (vs. proportional) bar graph condition overestimated the number of five-star ratings in distributions A (true value = 118; M_simple = 204.10, M_proportional= 150.47; t(282) = 4.47, p < .001, Cohen's d = .53) and B (true value = 725; M_simple = 840.17, M_proportional = 755.11; t(282) = 4.72, p < .001, Cohen's d = .56). When the distribution's peak value was high (i.e., distribution C in Figure 6), however, there was no difference in participants’ estimates (true value = 1,166; M_simple = 1,080.12, M_proportional = 1,093.59; t(282) = .66, p = .511, Cohen's d = .08). Figure 7 displays the participants’ estimates across the conditions. The results of this study supported our proposition that simple bar graphs, compared with proportional bar graphs, inflate consumers’ perceptions of ratings when the peak values are not close to 100% due to larger shifts in the graphs’ reference points.

OPEN IN VIEWER

Figure 7.

Estimates of Target Values in Study 2.

While these findings are crucial in advancing our understanding of how consumers process rating distributions displayed in various formats, in the real world, though, bar values (i.e., percentage or the number of ratings associated with each bar) are usually displayed by review platforms and need not be estimated by consumers. As displayed in Figure 1, while many platforms include this information in distribution graphs, some platforms do not. A critical question, therefore, would be whether consumers’ mental representations and recall of the ratings can also be affected by the graphical formats in which bar values are explicitly displayed.

We conducted two studies to answer this question and reported the results in Web Appendices C and D. We asked participants to recall the percentage of the ratings associated with different bars in Web Appendix C and the number of ratings associated with different bars in Web Appendix D. In both studies, participants in the simple bar graph condition recalled significantly greater values than those in the proportional bar graph condition, especially for the tallest bar (i.e., the peak value). Study 2 and the studies in Web Appendices C and D, together, therefore, suggest that not only do simple bar graphs lead to overestimation of the values, but they also distort consumers’ mental representation of the ratings, regardless of whether bar values are displayed explicitly as percentages, frequencies, or neither, leading to more favorable evaluations of the target products. These findings support our theorizing that simple bar graphs result in an inflated perception of the ratings and, therefore, higher evaluations of the target items.

Study 3: Ruling Out Alternative Explanations

We previously argued that the shift in visual reference points from 100% to the peak value induced by simple bar graphs is the underlying mechanism of the observed effect. In addition to this shift in reference points, however, simple and proportional bar graphs differ in several other aspects, including the bars’ relative and absolute lengths and the number of graphical cues provided. These differences, therefore, may also lead to altered judgments or make the processing of one format easier than the other. For example, when a fixed space is dedicated to displaying rating distributions as in Studies 1 and 2, because of the nature of the two formats, all bars in simple bar graphs appear taller and consequently more visually salient than those in proportional bar graphs. Therefore, one explanation for our main proposition that simple bar graphs lead to higher evaluations (H₁) can be that this difference in the absolute size of the bars and their visual salience may alter the perception of ratings and items’ favorability.

Moreover, graphical perception research finds that the number of elements displayed in a visual stimulus is essential in determining its complexity: the more graphical cues a stimulus has, the more complicated it is (Hegarty 2011). According to this view, processing a distribution displayed in a proportional bar graph could be more complicated than processing a distribution in a simple bar graph. The higher processing fluency caused by the lower number of cues can lead to more favorable product evaluations (Shen, Jiang, and Adaval 2010) and may explain why using simple graphs leads to higher evaluations of the target item (H₁). Conversely, a proportional bar graph offers more information than a simple bar graph does. Therefore, a lack of information can make the processing of a simple bar graph distribution more complex, leading to lower evaluations, which contradicts H₁. These interpretations suggest that processing one of the two formats, either simple or proportional, could require more time and effort than processing the other (Westerman, Lanska, and Olds 2015).

Method

We recruited 607 participants (69% female, 29% male, 2% other; M_age = 35.10 years) using Prolific. The stimuli consisted of 21 distributions with an average rating of 3.0, 3.3, 3.6, 3.9, 4.2, 4.5, or 4.8 randomly selected from a pool of 130,000 randomly generated distributions (see additional details in Web Appendix E). Participants were randomly assigned to one of the six format conditions displayed in Figure 8, where each person was shown seven rating distributions with different means in a randomized order. Formats I and IV were identical to those used in the previous studies. Including Formats II and V enabled us to test the impact of the presence of lightly shaded rectangles (i.e., the number of graphical cues displayed in a distribution graph). The lightly shaded rectangles in Format II do not affect the graph's reference points but increase the number of graphical cues compared with Format I's blank spaces.

OPEN IN VIEWER

Figure 8.

Examples of Formats Used in Study 3.

In contrast, Format III enabled us to test the impact of bars’ visual saliency because it displays information in a simple format, while the bars appear the same size as those in the proportional formats (i.e., Formats IV and V). It should be noted that unlike Format V, the placement of bar values next to the bars in Format III hinders the integration of 100% as a fixed reference point, as consumers do not need to visually scan the distance between the bars and the chart border to access and read the explicitly written values. Therefore, the explanation based on shifts in reference points predicts that participants viewing Formats I, II, and III will evaluate the products more positively than participants viewing Formats IV and V. The table-format condition (Format VI) was included as the control condition in which no graphical cues were provided.

Results

We conducted a one-way repeated-measures ANOVA to test whether the product evaluations differed significantly among the six conditions. As in previous studies, the format in which rating distributions were displayed had a significant effect on the product evaluations (F(5, 601) = 4.33, p = .001, η p 2  = .03). Figure 9 displays the product evaluations across the six conditions for all 21 distributions.

OPEN IN VIEWER

Figure 9.

Product Evaluation as a Function of Graphical Format for all Stimuli in Study 3.

A priori, we expected participants in both the simple and proportional bar graph conditions to evaluate the target items more favorably than those in the table condition because research has shown that providing information in a graphical format can facilitate cognitive processes (Lurie and Mason 2007; Scaife and Rogers 1996) that lead to higher evaluations. We aggregated product evaluations at the within-subjects level to conduct planned contrasts. Planned contrasts on aggregated data revealed that while participants in the simple bar graph (i.e., Formats I, II, and III) conditions evaluated the target products more positively than those in the proportional bar graph (i.e., Formats IV and V) conditions (M_simple = 6.00, M_proportional = 5.61; F(1, 601) = 14.95, p < .001, η p 2  = .02) and table conditions (M_table = 5.59; F(1, 601) = 10.25, p = .001, η p 2  = .02), there was no significant difference in product evaluation between the proportional and table conditions (F(1, 601) = .02, p = .900, η p 2  = .00).

We conducted four linear mixed-effects regression models to test the impact of bar lengths, the number of graphical cues (i.e., the presence of lightly shaded rectangles), and the simple versus the proportional aspect of various formats on evaluations. Since we were only interested in the effects of graphical features, those models were only applied to graphical format conditions (i.e., Formats I, II, III, IV, and V) using modified versions of Equation 1. As shown in Table 3, we included the graphical format (1 = simple, −1 = proportional), its interaction with the peak value, and the imbalance score in Models 1 and 2. Similar to Study 1, we also included random intercepts for participants and distributions in our models. Model 1 was the full model in which all predictors were included. Model 2 was our predicted model, which dropped the effects of the number of visual cues and bar size. Models 3 and 4 were covariates-only and base models in which graphical format or all graphical features were excluded, respectively. Model 2 outperformed Models 1, 3, and 4 according to the Akaike information criterion (AIC). This finding indicated that the proportional (vs. simple) aspect of graphical formats better explains the observed differences in product evaluations across the conditions.

OPEN IN VIEWER

Table 3.

Statistical Summary of the Mixed-Effects Regression Models Tested in Study 3.

Coefficient	Model 1: Full Model	Model 2: Predicted Model	Model 3: Covariates Only Model	Model 4: Base Model
γ00 (intercept)	5.81 (.10)***	5.80 (.09)***	5.85 (.10)***	5.84 (.09)***
β1 (H1: graphical format)	.15 (.07)**	.20 (.05)***	—	—
β2 (average rating)	1.89 (.40)***	1.89 (.40)***	1.89 (.40)***	1.89 (.41)***
β3 (variance)	.26 (.16)	.26 (.16)	.26 (.16)	.26 (.16)
β4 (peak value)	2.20 (.73)***	2.20 (.73)***	2.12 (.73)***	2.11 (.73)***
β5 (imbalance score)	1.95 (.69)***	1.95 (.69)***	1.98 (.69)***	1.98 (.69)***
β6 (H2: graphical format × peak value)	−.42 (.17)**	−.42 (.17)**	—	—
β7 (H3: graphical format × imbalance score)	.17 (.09)*	.17 (.09)*	—	—
β8 (bar size; −1 = small, 1 = large)	.05 (.07)	—	.16 (.05)***	—
β9 (number of cues; −1 = no gray rectangles, 1 = with gray rectangles)	−.08 (.06)	—	−.12 (.05)**	—
Grouping Variable	Random Effects (SD)
Subject	1.02	1.02	1.02	1.04
Stimulus	.37	.37	.37	.37
AIC	12,576	12,574	12,580	12,589

*p < .1.

**p < .05.

***p < .01.

Based on this model, offering rating distributions in simple bar graphs led to higher product evaluations (β₁ = .20, SE = .05, p < .001, 95% CI = [.09, ­­.30]), supporting H₁. Consistent with Study 1, peak value (β₄ = 2.20, SE = .73, p = .005, 95% CI = [.73, 3.53]) and imbalance score (β₅ = 1.95, SE = .69, p = .007, 95% CI = [.60, 3.38]) had positive effects on product evaluations. The negative and significant interaction between the display format and peak value (β₆ = −.42, SE = .17, p = .016, 95% CI = [−.73, −.07]) showed that an increase in the peak value leads to a smaller positive effect of using simple bar graphs. While the coefficient testing the interaction between the graphical format and imbalance score was marginally significant (β₇ = .17, SE = .09, p = .060, 95% CI = [.004, .347]), it was consistent with our prediction and the results of Study 1. Again, in consonance with previous research, the average rating had a positive effect on product evaluations (β₂ = 1.89, SE = .40, p < .001, 95% CI = [1.06, 2.69]).

Processing Fluency and Response Time

As discussed previously, the idea that processing a proportional bar graph distribution could require more cognitive resources than processing a simple bar graph distribution could be why using simple graphs leads to higher evaluations of the target item. If consumers have more difficulty processing one type of information, then they should require more time to process that piece of information (Westerman, Lanska, and Olds 2015). In this experiment, we measured the response time for each evaluation task. We conducted a one-way repeated-measures ANOVA to test the effect of the information presentation format on processing time (i.e., the amount of time spent on each product evaluation task in seconds). Our analysis revealed that the processing time was not significantly different across the six conditions (F(5, 601) = .67, p = .646). The response time in seconds for each task in Formats I through VI was, respectively, 19.91 (SD = 23.37), 17.79 (SD = 17.87), 19.79 (SD = 39.1), 19.16 (SD = 31.07), 17.05 (SD = 13.34), and 17.21 (SD = 17.49). Analyzing log-transformed processing times yielded similar results (F(5, 601) = 1.17, p = .411). Further, there was no significant correlation between the response time and product evaluation (r = .005, p = .761), negating the role of processing fluency in the current context.

In sum, Study 3 replicated the findings of previous studies and showed that other visual differences and bar size do not explain the impact of graphical formats on consumers’ evaluations. Additionally, no support for a processing fluency explanation was found.

Study 4: Variations in Numerical Information, Products, and Colors

In the preceding studies, the value of the bars had always been reported in percentages to isolate the effect of the graphical format from other attributes such as the format of the numerical information. As displayed in Figure 1, however, review platforms display bar values using both frequencies and percentages. Moreover, we only tested our predictions with a limited variation in products and their descriptions, where rating distributions were the main part of the stimuli viewed by participants. The main objectives of this study were, therefore, threefold: we wanted to (1) test whether the effect of the graphical format on evaluation holds when additional pieces of information (e.g., product descriptions, average ratings, product images, and ratings volume) compete for consumers’ attention as in the real world, (2) examine whether a match (i.e., simple bar graphs with frequencies and proportional bar graphs with percentages) or mismatch (i.e., simple bar graphs with percentages and proportional bar graphs with frequencies) between the graphical format and explicitly written values affects the results, and (3) replicate the results with more externally valid stimuli and a wide variety of products.

Method

We recruited 400 MTurk participants (44% female, 56% male; M_age = 39.31 years). The same pool of 21 distributions and measurement scale employed in Study 3 (see Web Appendix E for more details) were used. We randomly assigned participants to one of the five formats exhibited in Figure 10 in a mixed design, where each participant evaluated seven different products (i.e., rechargeable LED lamp, face mask, charcoal grill, coffee maker, toaster, air circulator, and microwave oven) with different rating distributions in a randomized order (see Figure W5 of Web Appendix F for actual stimuli). We designed our stimuli to resemble a product page on Amazon.com. Formats I (simple with frequencies), II (simple with percentages), III (proportional with frequencies), and IV (proportional with percentages) allowed us to test the impact of the type of numerical information provided in a distribution graph (written frequencies vs. percentages for the value of each bar). Format V (proportional with percentages and high contrast) was included to test the impact of the contrast between the highlighted and empty portion of the bars in proportional bar graphs on evaluations.

OPEN IN VIEWER

Figure 10.

Examples of Formats Used in Study 4.

Results

We conducted a two-way repeated measures ANOVA to test the effects of the graphical format, the format of the numerical information, and their interaction on product evaluation. ¹ Our analysis revealed that the effect of using different graphical formats (i.e., simple vs. proportional bar graphs) was significant (F(1, 396) = 13.19, p < .001, η p 2  = .03). Importantly, neither the main effect of the format of numerical information (F(1, 396) = .16, p = .687, η p 2  = .00) nor its interaction with the graphical format (F(1, 396) = .03, p = .876, η p 2  = .00) was significant. Consistent with our prediction and as in Study 3, planned contrasts on the within-subjects level aggregated data revealed that product evaluation in the simple bar graph conditions (i.e., Formats I and II; M_simple = 6.68) was higher than in the proportional bar graph conditions (i.e., Formats III, IV, and V; M_proportional = 6.24; F(1, 395) = 13.11, p < .001, η p 2  = .03). Figure 11 displays participants’ evaluations of the products across the five conditions for all stimuli.

OPEN IN VIEWER

Figure 11.

Product Evaluation Across Graphical Formats for All Stimuli in Study 4.

These findings support the idea that the positive effect of using a simple bar graph holds even when (1) other information (e.g., product description and images, rating volume, and graphical and verbal presentation of the average ratings) is provided and (2) bar values are provided as frequencies and percentages. This study demonstrates the robustness of the results of Study 1 and Study 3 to differences in numerical information format and other attributes such as average rating. Thus far, we have shown that the graphical format of rating distributions affects consumers’ perception of the ratings and their preconsumption evaluations of various products. In the next two studies, we examine the effect of distributions’ graphical format on items’ future ratings (H₄).

Study 5: The Effect of the Graphical Format of Prior Ratings on Future Ratings

As mentioned previously, consumers’ perceptions of an item's prior ratings may influence the valence of its future ratings. Study 2 demonstrated that consumers perceive ratings differently depending on the graphical format in which they are displayed. We have also provided additional evidence showing that consumers’ perception and recall of ratings are influenced by the graphical format in Web Appendices C and D. Therefore, Study 5 aimed to test whether displaying rating distributions in simple and proportional bar graphs would affect how consumers rate a consumed product or service.

Method

In exchange for a small payment, we recruited 921 CloudResearch-approved participants (62% female, 36% male, 2% other; M_age = 40.14 years). They were asked to imagine being in the following scenario:

Imagine that you have recently booked a flight to go on a vacation in a few days. Your carry-on bag is damaged, and you must purchase a brand-new one. However, you are not sure about what to buy. You decide to use the Internet to find a high-quality carry-on bag with free shipping. Editors of the Strategist website collect deals, gift guides, and product reviews from around the web and list high-demand items visited by many shoppers. You do your research online, come across this website, and visit the Strategist website to see if you can find a brand-new carry-on bag.

Next, they were informed:

Many people visit the Strategist every month to find new products and ideas that are picked by experts. These users have turned to Google to write reviews and rate this website. Below you can see how users have rated the Strategist so far.

At this point, we randomly assigned participants to either simple or proportional bar graph conditions as displayed in Figure 12. Then, they were asked to visit the Strategist website, navigate through it, and try to find a carry-on bag that they might purchase.

OPEN IN VIEWER

Figure 12.

Stimuli Used in Study 5.

After visiting the Strategist website, they were asked to report the characteristics of the bag that they liked the most (name, color, price), followed by the likelihood of buying it. Then, using a continuous star-rating scale, they were asked to rate the Strategist website as if they were doing it on Google while viewing its prior ratings displayed in a simple or proportional bar graph. Finally, we measured the extent to which they (1) were familiar with the Strategist website before they participated in our study (1 = “not familiar at all,” and 7 = “very familiar”) and (2) thought that the Strategist website was rated fairly by other customers (1 = “underrated,” 4 = “fairly rated,” and 7 = “overrated”), followed by basic demographic questions.

Results

To ensure that participants rated the website (and not their preferred carry-on bag), we asked them to indicate what they rated after the website rating task. Of 921 participants who completed the study, 130 reported that they rated their preferred bag (128 participants) or did not recall what they rated (2 participants). Excluding those participants from our analysis, we found that participants in the simple condition rated the Strategist website more positively (M_simple = 3.71, SD = .95) than those in the proportional bar graph condition (M_proportional= 3.53, SD = .99; t(789) = 2.66, p = .008, Cohen's d = .19). This finding was supported even after we controlled for the effects of familiarity with the website and perceived fairness of prior ratings in a regression analysis, where we found that using a simple (vs. proportional) bar graph to display prior ratings led to more favorable ratings (b = .21, SE = .07, p = .002). This analysis also revealed that (1) familiarity had a positive effect on future ratings (b = .12, SE = .03, p < .001), and (2) a higher score in the perceived fairness of prior ratings led to lower postconsumption ratings (b = −.29, SE = .04, p < .001). Furthermore, familiarity with the Strategist website (M_simple = 1.46, M_proportional= 1.39; t(789) = 1.10, p = .273) was not different between the two conditions.

Although comparing the perceived fairness of prior ratings with the scale's central point (i.e., 4 = “fairly rated”) using two one-sample t-tests revealed that participants in both conditions believed that prior customers had overrated the Strategist website (M_simple = 4.36, t(395) = 8.65, p < .001; M_proportional= 4.30, t(394) = 6.80, p < .001), the perceived fairness of prior ratings was not different across the two conditions (t(789) = .94, p = .345). Importantly, neither the prices of the products that participants preferred (M_simple = $168.38, M_proportional= $155.24; t(789) = .81, p = .418) nor the willingness to buy those products (M_simple = 4.19, M_proportional= 4.24; t(789) = .39, p = .694) was significantly different between the two conditions, suggesting that participants’ experiences with the Strategist website in simple and proportional bar graph conditions were comparable. In sum, this study showed that the graphical format of a product's ratings could affect its future ratings, as described in H₄. In Study 6, we compare the ratings that a number of businesses have received on two different platforms to test if any field evidence can be found for the results of Study 5.

Study 6: Do the Ratings of the Same Business Differ on Various Platforms?

Study 5 found that consumers tend to rate an item more positively when its prior ratings are displayed in a simple (vs. proportional) bar graph (H₄). Since Google and TripAdvisor display rating distributions in simple and proportional graphical formats, respectively, rating distributions should function as higher anchors when consumers rate businesses on Google. Therefore, businesses should receive more favorable ratings on Google than on TripAdvisor. This study aimed to test the extent to which the ratings of a product or service differ across platforms to provide further evidence for H₄.

Method

We searched Google Maps, which uses the simple bar graph format (Figure 1), for restaurants in New York City. The search yielded 429 locations whose rating distributions and price ranges were collected. We then searched all 429 locations on TripAdvisor—which uses the proportional bar graph format (Figure 1)—and found matches for 343 restaurants. From TripAdvisor restaurant pages, we collected the number of reviews and their distributions.

To test whether businesses received different ratings on the two platforms, we conducted a mixed-effects regression model using individual rating scores as the dependent variable and platform as the independent variable. We included price level and its interaction with the platform as covariates. We treated price level as a continuous variable that ranged from one to four depending on the number of dollar signs displayed on each restaurant page on Google. A random intercept for each restaurant (SD = .19) was also included in the model.

Results

The positive main effect of the platform on ratings revealed that restaurants in our data set received higher ratings on Google (b = .18, SE = .01, p < .001), providing additional evidence supporting H₄. More expensive restaurants were rated more favorably on TripAdvisor (b = .03, SE = .01, p = .015), but the negative interaction between platform and price (b = −.007, SE = .003, p = .011) revealed that the positive effect of pricing level was smaller on Google (see Table 4).

OPEN IN VIEWER

Table 4.

Study 6 Mixed-Effects Regression Results.

Predictor	Estimates	SE	p-Value
(Intercept)	4.18	.03	<.001
Platform (Google vs. TripAdvisor)	.18	.01	<.001
Price	.03	.01	.015
Platform × Price	−.007	.003	.011
Random Effects	Variable		SD
Grouping variable: Restaurant	Intercept		.19
AIC			2,664,219

These results suggest that a business or product is likely to be rated more positively on Google, which uses simple bar graphs to display rating distributions. While these findings are consistent with Study 5 and support H₄, other differences between Google and TripAdvisor that may have led to this effect should not be overlooked. For example, Google and TripAdvisor users may systematically differ in terms of several attributes, such as age, gender, education, expertise, and income, which can lead to different rating behaviors. Moreover, consumers might use Google for more casual events when expectations are lower but TripAdvisor for more important occasions when expectations are higher, and rating the same experience while having lower standards could lead to more positive ratings. Another explanation for the observed results could be that Google reviews are monitored more seriously by businesses, and therefore consumers may feel pressure to provide higher rating scores. Therefore, the results of this study should not be attributed solely to the graphical formats used by Google and TripAdvisor.

General Discussion

This research has focused on an important yet neglected aspect of online WOM. We provide a systematic examination of how displaying ratings in various bar graphs significantly affects consumers’ evaluations of products. Across nine studies (six reported here and three in the Web Appendices), we find compelling evidence that consumers report a higher evaluation of an item when its ratings are displayed in a simple bar graph. We attribute these findings to consumers’ integration of the graphs’ reference points. Utilizing an extensive set of rating distributions, we establish a positive effect of displaying ratings in simple (vs. proportional) bar graphs on consumers’ pre- and postconsumption evaluations.

Importantly, we find that the degree to which simple bar graphs can boost consumers’ evaluations is not equal across all distributions. Studies 1, 2, and 3 reveal that the magnitude of the shifts in graphs’ reference points in simple bar graphs (which has a negative relationship with the distribution's peak value) moderates this effect, suggesting a smaller effect for distributions with greater peak values. The results of Study 2 and Web Appendices C and D demonstrate that consumers’ perceptions and recall of an item's ratings are significantly affected by the graphical format in which those ratings are displayed, providing support for our proposed mechanism.

We also find that the imbalance score of a set of ratings moderates the effect of using different graphical formats. Studies 1 and 3 show that consumers are more influenced by simple bar graphs when rating distributions have greater imbalance scores, as is the case for the majority of products and services (Rocklage, Rucker, and Nordgren 2021). Study 3 shows that neither the differences in relative and absolute sizes of the bars nor the presence of lightly shaded rectangles could explain our findings. Study 4 demonstrates the robustness of the effect of the graphical format to other factors such as the format in which bar values are displayed. Finally, Studies 5 and 6 demonstrate that an item's future ratings could also be influenced by the graphical format of its current ratings.

The findings of the current work might appear surprising to many laypeople. How can a minor change in the graphical presentation format of rating distributions impact not only the target item's evaluation but also consumers’ future rating behavior? However, it should not come as a surprise to those who study visual perception. The visual perception literature is replete with such outcomes in various contexts. In the context of risk communication, for example, Stone et al. (2003) find that using various graphical formats to communicate risk information affects individuals’ perception of the risk and their willingness to engage in preventive actions. Similarly, in marketing research, Thomas and Kyung (2019) find that using a slider scale with two visual reference points (vs. text boxes) to elicit consumers’ willingness to pay for an item leads to more extreme responses assimilating toward those reference points.

Studying how individuals process bar graphs, Zacks et al. (1998) found that adding a third dimension to simple bar graphs can significantly affect individuals’ judgment of the bars. Importantly, they found that even adding a context bar (i.e., a bar not related to the target bar) to a graph leads to biased perceptions of the target bars assimilating toward the context bar. As discussed previously, we know that adding irrelevant depth cues to a bar graph affects both perception and encoding of the bars (Fischer 2000) and even displaying bars separately (vs. adjacent to each other) or in an aligned (vs. unaligned) position with respect to the baseline significantly affects how bar graphs are processed and perceived (Talbot, Setlur, and Anand 2014).

Implications

From a theoretical perspective, this research makes several important contributions. First, we contribute to the large body of research on online WOM. Extant research on consumers’ interpretations of online consumer ratings has been focused mainly on the impact of attributes such as average ratings, rating variance, and rating volume. Even the limited research that has studied the consequences of displaying rating distributions in a graphical format (e.g., Fisher, Newman, and Dhar 2018) does not differentiate between various forms of displaying online consumer ratings. Our findings suggest that the format used to display rating distributions influences consumers’ perception and recall of ratings, their evaluation of products, and their postconsumption ratings. As a result, by identifying the graphical format of rating distributions as a key factor in consumers’ evaluations, these findings enhance our understanding of how online ratings can affect consumers and businesses.

This work also contributes to graphical perception and visual information processing research (e.g., Cleveland and McGill 1984; Hegarty 2011; Lurie and Mason 2007; Talbot, Setlur, and Anand 2014). We augment this research stream by scrutinizing differences between interpretations of graphical formats in an online shopping context and demonstrating how consumers make sense of different graphical representations of the same information. Our results are consistent with the idea that even small changes in relatively subtle aspects of graphical displays can affect task performance dramatically (e.g., Fischer 2000; Hegarty 2011; Lurie and Mason 2007). This research extends the understanding of how different reference points introduced by simple and proportional bar graphs can affect consumers’ perceptions of ratings. Contributing to extant research on visual perception, we also find that individuals’ estimates and recall of rating information are affected by the ratings’ graphical format.

From a managerial perspective, given the astronomical number of reviews being posted, the present work has important implications. Consumers are often overwhelmed in extracting meaningful insights from all the reviews submitted for an item. Summary information, such as rating distributions and average ratings, empowers consumers to evaluate and compare alternatives more efficiently. This work demonstrates that the relatively subtle features of online WOM, such as the graphical format of rating distributions, can substantially influence consumer judgments. Marketers should not ignore this aspect of online WOM, as it is one of the few attributes they fully control and can modify at no additional cost.

Our findings suggest that marketers can improve consumer evaluations of their offerings by displaying ratings in simple bar graphs, which is especially important for retailers or platforms that compete for more sales and transactional fees. For instance, platforms such as Google Maps and TripAdvisor operate in the same industry and both derive advertising and transactional income by enabling business owners to advertise their services and allowing consumers to place online orders or make reservations. Our results suggest that platforms that display ratings in a proportional format (e.g., TripAdvisor) would benefit from switching to a simple bar graph format.

As TripAdvisor and Google use different graphical formats, we conducted a simulation to predict users’ evaluations of the restaurants examined in Study 6 to gain additional insights. We used the results of Study 1 to estimate consumers’ evaluations of the same restaurants on Google and on TripAdvisor. On the ten-point scale measuring product evaluations used in the previous experiments, on average, TripAdvisor users’ evaluations of restaurants in our data set would be 6.59, whereas Google users’ would be 7.44 (i.e., 13% higher; t(342) = 28.56, p < .001). According to the National Restaurant Association, the industry generated $864.3 billion in sales in 2021 in the United States alone (FinanceOnline 2022). With this market size, even a small difference in consumers’ evaluations of restaurants can have a substantial financial impact on both restaurants and review platforms. This suggests that businesses would benefit from promoting their products and services on platforms that use simple bar graphs. We also find that the same product or service receives more favorable ratings when its previous ratings are displayed in simple bar graphs. Retailers and manufacturers who sell their products online can improve their products’ future ratings by using simple bar graphs on their websites.

The current findings, however, do not advocate that all platforms display ratings in simple bar graphs. For instance, consider a product category for which a suboptimal choice can have substantial negative consequences. For example, consumers and physicians may read patients’ reviews for prescription drugs and their side effects on various platforms, such as the Ask a Patient website. The cost of overestimating the benefits when deciding whether to take or prescribe a medication, either by a physician or a patient, would be greater than in the case of choosing a restaurant or a product such as a toaster. Thus, although simple bar graphs may lead to higher product evaluations and potentially higher sales, they should not be used for product categories where a suboptimal choice may lead to potentially grave consequences.

In a similar scenario, the anchoring effect may lead to a biased postconsumption evaluation of a consumed prescription drug when its prior ratings are displayed. Platforms that operate in such categories may consider not displaying rating information on their review solicitation pages. Indeed, the possibility of using graphical displays in an unethical manner to mislead consumers does exist. Therefore, firms should use graphical displays cautiously and consider the ethical repercussions of their usage.

Limitations and Future Work

Although our findings are robust, they need to be viewed in light of two caveats. First, while we generated a population of 130,000 distributions and used samples from these, it was impossible to exhaust all possible distributions. Second, while consumers incorporate distribution information into their judgments, their evaluations are also influenced by other factors, such as prices, brand perception, and the textual content of reviews. For example, research has shown that the textual content of individual reviews can affect consumer decisions beyond other attributes of online WOM (Chen and Lurie 2013; Grewal and Stephen 2019; Kronrod and Danziger 2013; Rocklage and Fazio 2020), but we did not investigate how the impact of using different graphical formats is influenced by the textual content of reviews. Moreover, consumers may put more (or less) weight on the rating distribution summary depending on its graphical format and the levels of other attributes (such as average rating, rating volume, price, or product category), possibly leading to different outcomes.

While we found consistent evidence supporting our proposition that a distribution's graphical format influences consumer judgments, we did not study the long-term effects of those formats on consumers and review platforms. Moreover, we do not claim that graphical formats influence all individuals in the same way. There can be individual variations, such as differences in numeracy levels or information processing styles (verbal vs. visual; Bagozzi 2008; Childers, Houston, and Heckler 1985), that can lead to different outcomes. Future work can also investigate additional graphical aspects of rating distributions, such as the color and orientation of the bar graphs. Whether consumers perceive vertical and horizontal bar graphs differently is a worthwhile research question. Future research may investigate how a bar graph's aspect ratio can influence consumers’ perception in an online shopping context.

In conclusion, this research sheds new light on how consumer judgments are influenced by online rating distributions. We investigated the role of graphical displays in online WOM and found that consumers’ perceptions and recall of rating information are affected by the ratings’ graphical format, leading to diverse outcomes. We contribute to the marketing literature and visual perception and information visualization research by showing how consumers perceive rating distributions differently depending on their graphical formats. In addition, this research has significant implications for consumers, academics, and marketing practitioners by enhancing the understanding of rating distributions and their differential impact on consumer evaluations.

Supplemental Material

sj-pdf-1-mrj-10.1177_00222437231179186 - Supplemental material for Unveiling Stars: How Graphical Displays of Online Consumer Ratings Affect Consumer Perception and Judgment

Supplemental material, sj-pdf-1-mrj-10.1177_00222437231179186 for Unveiling Stars: How Graphical Displays of Online Consumer Ratings Affect Consumer Perception and Judgment by Javad Mousavi, Surendra N. Singh and Promothesh Chatterjee, Tamara Masters in Journal of Marketing Research