Why We Leave Only Extreme Reviews
What studies of hundreds of millions of ratings say about the five-star and one-star pile-ups, and why the middle goes quiet.
Answer to the question
Why are online reviews mostly five stars or one star?
Mostly because of who chooses to write. Research on more than 280 million online reviews found that people with extreme evaluations are more likely to review, and a later model proposed that moderate reviewers drop out more often before posting. Both help explain the J-shaped pattern of ratings.
Key takeaways
- Nan Hu, Jie Zhang and Paul Pavlou found in 2005 data on books, DVDs and videos that between 72 and 78 per cent of ratings stood at four stars or higher.
- Verena Schoenmueller, Oded Netzer and Florian Stahl analysed more than 280 million reviews from 25 platforms and found that self-selection of people with extreme evaluations is an important driver of polarity.
- Leif Brandes, David Godes and Dina Mayzlin proposed differential attrition: moderate reviewers are more likely to stop before posting.
- Eugene Anderson's 1998 work on satisfaction and word of mouth did not allow an asymmetric U-shape to be rejected, with dissatisfied customers talking somewhat more than satisfied ones.
- An unincentivised email asking customers to review reduced review extremity in a large field experiment, which suggests the shape moves with who is asked.
Online reviews for most products pile up in one shape: a tall stack of five-star ratings, a smaller stack of one-star ratings and a thin scatter in between. Researchers call it a J-shaped distribution, and it shows up in data on books and videos from the mid-2000s as well as in hundreds of millions of online reviews gathered since. The research points less to products being either wonderful or awful and more to who decides to write.
Why are online reviews mostly five stars or one star?
Nan Hu, Jie Zhang and Paul Pavlou described the pattern in 2009 using ratings for books, DVDs and videos collected between February and July 2005. Between 72 and 78 per cent of the ratings stood at four stars or higher. They concluded that the average rating is a weak proxy for quality. The J-Shaped Review Distribution is the shape they described, and the studies that follow ask what produces it.
A far larger test came in 2020, when Verena Schoenmueller, Oded Netzer and Florian Stahl analysed more than 280 million online reviews from 25 platforms. Most platforms showed high polarity, though the degree varied with how selective customers were about reviewing. Using platform data, experiments and a survey, the authors found that self-selection, where people with extreme evaluations are more likely to write a review, is an important driver of the shape. That helps explain why online reviews can look so cheerful in aggregate: customers who feel strongly are the ones who write, and those who felt only that a product was fine are mostly absent.
Three accounts of the J sit side by side in the research. They do not compete so much as describe different steps between buying a product and posting about it, and each is stated here as its authors stated it.
| Account | What it says | Where it comes from |
|---|---|---|
| Self-selection | People with extreme evaluations are more likely to write | Schoenmueller, Netzer and Stahl, 2020 |
| Differential attrition | People with moderate experiences drop out before posting | Brandes, Godes and Mayzlin, 2022 |
| Satisfaction and talk form a U | Very pleased and very displeased customers talk more than those in between | Anderson, 1998 |
Who writes a review and who stays silent about it?
The idea is older than the web. In 1998 Eugene Anderson asked whether dissatisfied customers spread more word of mouth than satisfied ones, since theory and earlier evidence pointed both ways. He proposed an asymmetric U-shaped relationship between satisfaction and talk, and data from the United States and Sweden did not allow that shape to be rejected, with similar estimates in both countries. Dissatisfied customers did talk more than satisfied ones, though by a smaller margin than is often assumed. That detail matters for online reviews, because it suggests the pleased end of the scale was never silent, which fits a J with a tall stack at five stars.
Leif Brandes, David Godes and Dina Mayzlin added a mechanism in 2022 that they call differential attrition. Moderate reviewers, they proposed, are more likely than extreme ones to stop somewhere along the path to posting. Their model combined attrition with differences in the utility of writing and with base rates, and their analyses found evidence for both attrition and utility in review data.
Jonah Berger's Contagious supplies the everyday logic behind these models. His case is that people pass on what is remarkable, meaning worth remarking on, because sharing it makes the sharer look interesting. A meal that was fine or a gadget that did what the box said gives the writer little to remark on. Read that way, an online review is word of mouth with a star count attached. Berger wrote about talk and not about star ratings, so applying his point to the J is our extension, and the papers above are the evidence for the pattern itself. The two fit together: ordinary experiences give little reason to open the page at all.
What does the average star rating hide from the reader?
Schoenmueller and colleagues also reported that self-selection and the polarity it creates make online reviews less informative. Hu and colleagues made the related point years earlier that the average is a weak proxy for quality. Neither paper says that online reviews are false. Each says that a set of online reviews written by the most moved customers is a partial portrait, and a partial portrait is easy to misread when the headline number is a single average.
We would read the spread before the average. Count how many ratings sit at each level, and treat the three-star reviews as the rarest voices, since the research suggests they come from the group most likely to have dropped out. A mild criticism inside an otherwise warm review is ordinary, and research gathered in Why a Minor Flaw Makes Brands Likeable shows a small flaw can change how a brand is judged. The bottom stack deserves a read too, since Why Does One Bad Review Hurt So Much asks how much weight a single low score carries.
The shape also matters beyond a single product page. Visible ratings guide the next person, and What is social proof in marketing? covers how signs of other people's choices steer later choices. A lopsided set can therefore reach further than the few who wrote it.
The strongest sign that the J is produced by who writes comes from the field experiment in the Brandes, Godes and Mayzlin paper. An unincentivised email inviting customers to post, sent by an online travel platform, reduced extremity in a large field experiment. The authors report that only the utility-based and attrition-based mechanisms in their model can explain that result. In plain terms, asking a wider group of customers to write, without paying them to, pulled the scores towards the middle. On this evidence the J is less a fixed trait of the goods than a record of whom the platform asked.
Common questions
Why are online reviews mostly five stars or one star?
Mainly because of who chooses to write. Research on more than 280 million reviews found that people with extreme evaluations are more likely to review, and a later model proposed that moderate reviewers drop out more often before posting.
What is a J-shaped distribution of reviews?
It is the pattern in which most ratings sit at five stars, a smaller group sits at one star and few fall in between. In 2005 data on books, DVDs and videos, between 72 and 78 per cent of ratings were four stars or higher.
Does this mean online reviews are wrong?
No. The research says the set is a partial picture written mostly by customers with strong views, which makes the average less informative.
How should I read a set of ratings?
Read the spread across every star level before the headline number, and give the rarest middle scores a careful look.
Sources
Berger, J. Contagious. Seed book; vault hub Books/Contagious.
Hu, N., Zhang, J., & Pavlou, P. A. (2009). Overcoming the J-shaped distribution of product reviews. Communications of the ACM, 52(10), 144-147. doi:10.1145/1562764.1562800 https://doi.org/10.1145/1562764.1562800
Schoenmueller, V., Netzer, O., & Stahl, F. (2020). The polarity of online reviews: Prevalence, drivers and implications. Journal of Marketing Research, 57(5), 853-877. doi:10.1177/0022243720941832 https://doi.org/10.1177/0022243720941832
Brandes, L., Godes, D., & Mayzlin, D. (2022). Extremity bias in online reviews: The role of attrition. Journal of Marketing Research, 59(4), 675-695. doi:10.1177/00222437211073579 https://doi.org/10.1177/00222437211073579
Anderson, E. W. (1998). Customer satisfaction and word of mouth. Journal of Service Research, 1(1), 5-17. doi:10.1177/109467059800100102 https://doi.org/10.1177/109467059800100102