Explainers / RESEARCH
Why are almost all online ratings above 4 stars?
Because the people who write reviews are mostly those with strong opinions, and most strong opinions are positive. The resulting J-shaped distribution, with many five-star ratings, some one-star ratings and few in the middle, pushes most averages into a narrow band near the top of the scale.
On a five-point scale, three is the middle. On a review site, three is a disaster. Most of the businesses you will ever compare sit in the fours, and the whole contest takes place in a few tenths of a star. To read ratings well you have to recalibrate, and it helps to know why the scale is compressed.
The J shape
In 2009, Hu, Zhang and Pavlou examined Amazon ratings for books, DVDs and videos collected in 2005. They found that 78%, 73% and 72% of ratings in those three categories were four stars or higher, and that the distribution was J-shaped: "mostly 5-star ratings, some 1-star ratings, and hardly any ratings in between".11 Hu, Zhang and Pavlou, Communications of the ACM 52(10), 2009, Overcoming the J-shaped distribution of product reviews
They checked two obvious explanations and rejected both. The J shape was not a small-sample artifact: products with over 2,000 reviews had it too. And it was not confined to good products: even items with a middling average showed the same two-humped pattern.
Their statistical point is the one that should bother anyone who relies on averages. The mean is a sensible summary of a one-humped distribution. For a lopsided two-humped one, they wrote, "the average is a poor proxy of product quality". A 4.3 could be a business where nearly everyone is fairly pleased, or one where most customers are thrilled and one in ten has a terrible time. The average cannot tell those apart. The histogram can.
Why the middle is missing
The leading explanation is about who bothers. Hu and colleagues set out to identify the biases in review systems that produce the shape, and later work has put self-selection at the center: buyers were already disposed to like what they chose, and among buyers, those with extreme experiences are the ones moved to report.
Schoenmueller, Netzer and Stahl tested this at scale, compiling over 280 million reviews from 25 platforms. Most platforms showed strong polarity. Using cross-platform comparison, experiments and surveys, they confirmed "polarity self-selection", the higher tendency of consumers with extreme evaluations to write, as an important driver.22 Schoenmueller, Netzer and Stahl, Journal of Marketing Research 57(5), 2020, The Polarity of Online Reviews: Prevalence, Drivers and Implications They also found that platforms differ: polarity depended on "how selective customers are in reviewing products on the platform". Where a larger share of customers review, the middle fills in.
That last finding has a practical use. A business that asks every customer for a review will, in principle, show a fuller, less extreme distribution than one whose only reviewers are self-starters. A profile with a real population of three and four star reviews may be a sign of broad sampling, not of mediocrity.
Earlier evidence
The positivity was noticed almost as soon as reviews existed. Chevalier and Mayzlin, studying books at Amazon and Barnes and Noble, reported in 2006 that "reviews are overwhelmingly positive at both sites". They also found that one-star reviews had a greater impact on sales than five-star reviews in most of their samples.33 Chevalier and Mayzlin, Journal of Marketing Research 43(3), 2006, The Effect of Word of Mouth on Sales: Online Book Reviews When nearly everything is praised, criticism carries the information.
The extreme case: where every stay is above average
Zervas, Proserpio and Byers collected ratings for millions of Airbnb listings and found that "nearly 95% of Airbnb properties boast an average star-rating of either 4.5 or 5 stars", and virtually none were below 3.5. Hotels on Tripadvisor, by contrast, averaged 3.8.44 Zervas, Proserpio and Byers, Marketing Letters 32, 2021, A first look at online reputation on Airbnb, where every stay is above average
Airbnb reviews follow a personal interaction between host and guest, and both sides rate each other. Laboratory work on feedback systems suggests why that matters. Bolton, Greiner and Ockenfels found that reciprocity in feedback giving "distorts the production and content of reputation information", and that changing how feedback information flows can produce more accurate reputations.55 Bolton, Greiner and Ockenfels, Management Science 59(2), 2013, Engineering Trust: Reciprocity in the Production of Reputation Information
Local services resemble Airbnb more than they resemble Amazon. You have met the plumber. He was in your kitchen. He may have asked you, in person, for a review. The social cost of leaving three stars is higher than for a toaster.
What compression does to decisions
When most businesses sit bunched in the fours, three things follow.
Small differences get over-read. The gap between 4.6 and 4.8 looks meaningful and is often within the margin of error for either business. See how many reviews are enough.
Rounding matters more. With everyone bunched together, the difference between displaying 4.4 and 4.5 can separate a business from several competitors. Studies that exploit Yelp's rounding found measurable effects on revenue and bookings at those thresholds.66 Michael Luca, Harvard Business School Working Paper 12-016, Reviews, Reputation, and Revenue: The Case of Yelp.com77 Anderson and Magruder, The Economic Journal 122(563), 2012, Learning from the Crowd: Regression Discontinuity Estimates of the Effects of an Online Review Database
Low ratings carry more signal than high ones. If 4.5 is ordinary, a 3.8 from a few hundred reviews is a strong negative signal, far stronger than its distance from the midpoint of the scale suggests.
Recalibrating
A rough translation table for local businesses, offered as our reading of the compressed scale and not as a research result:
| Displayed average, 100 or more reviews | How we would read it |
|---|---|
| 4.8 and above | Very strong, or actively managed; check how reviews are collected |
| 4.5 to 4.7 | Good; the normal range for a well-run business |
| 4.2 to 4.4 | Mixed; read the low reviews for a recurring problem |
| 3.8 to 4.1 | A consistent share of customers are unhappy |
| Below 3.8 | Most businesses avoid this; find out why this one did not |
Then look at the histogram. Two businesses at 4.3 with different shapes are different risks. One with ratings spread across three, four and five stars is reliably fine. One with 85% fives and 12% ones is excellent most of the time and awful the rest, and your question becomes what the ones have in common.
Questions readers ask
Does a J-shaped distribution mean reviews are fake?
No. It is the normal shape of honest, self-selected reviews, documented across many platforms.22 Schoenmueller, Netzer and Stahl, Journal of Marketing Research 57(5), 2020, The Polarity of Online Reviews: Prevalence, Drivers and Implications
Why do so few people leave three-star reviews?
An adequate experience gives little reason to write. Research attributes the missing middle largely to people with moderate opinions choosing not to review.22 Schoenmueller, Netzer and Stahl, Journal of Marketing Research 57(5), 2020, The Polarity of Online Reviews: Prevalence, Drivers and Implications
Are ratings getting higher over time?
One published paper argues that reputation systems suffer from "reputation inflation", which may make them less informative in the long run.88 Filippas, Horton and Golden, Marketing Science 41(4), 2022, Reputation Inflation We have not read a study that establishes the trend for local business reviews specifically.
Method A summary of four peer-reviewed studies. Percentages are quoted from the papers and refer to the data sets and years each paper describes. See how we source and check.
Sources
Every source below was opened and read for this page. Links go to the publisher's own copy.
- Hu, Zhang and Pavlou, Communications of the ACM 52(10), 2009. Overcoming the J-shaped distribution of product reviews
- Schoenmueller, Netzer and Stahl, Journal of Marketing Research 57(5), 2020. The Polarity of Online Reviews: Prevalence, Drivers and Implications
- Chevalier and Mayzlin, Journal of Marketing Research 43(3), 2006. The Effect of Word of Mouth on Sales: Online Book Reviews
- Zervas, Proserpio and Byers, Marketing Letters 32, 2021. A first look at online reputation on Airbnb, where every stay is above average
- Bolton, Greiner and Ockenfels, Management Science 59(2), 2013. Engineering Trust: Reciprocity in the Production of Reputation Information
- Michael Luca, Harvard Business School Working Paper 12-016. Reviews, Reputation, and Revenue: The Case of Yelp.com
- Anderson and Magruder, The Economic Journal 122(563), 2012. Learning from the Crowd: Regression Discontinuity Estimates of the Effects of an Online Review Database
- Filippas, Horton and Golden, Marketing Science 41(4), 2022. Reputation Inflation