Math / WORKED EXAMPLE
Rounding and the half-star cliff: how 0.01 can change the display
When a platform rounds to the nearest half star, the thresholds fall at .25 and .75. An average of 4.24 shows four stars and 4.25 shows four and a half. Studies of Yelp found that crossing such a threshold changed restaurant revenue and bookings, even though the businesses on either side were essentially the same.
A star display is a rounded number, and rounding creates cliffs. Two businesses a hundredth of a star apart can be drawn with different numbers of stars. The cliffs have been measured, and they move money.
Where the cliffs are
Rounding to the nearest half star sends each average to whichever half-step is closest.
| True average | Displayed |
|---|---|
| 3.75 to 4.24 | 4 stars |
| 4.25 to 4.74 | 4.5 stars |
| 4.75 to 5.00 | 5 stars |
Trustpilot states its rule: the TrustScore is shown as stars "including half stars", using "standard mathematical rounding rules".11 Trustpilot Help Center, TrustScore and star rating explained Academic work on Yelp relied on the same feature; Luca's study notes that Yelp "prominently displays a restaurant's rounded average rating".22 Michael Luca, Harvard Business School Working Paper 12-016, Reviews, Reputation, and Revenue: The Case of Yelp.com
A display that prints one decimal place has smaller and more numerous cliffs: at 4.65, 4.75, 4.85 and so on.
What the cliffs are worth
Researchers love rounding because it creates a natural experiment. A restaurant at 4.24 and one at 4.26 are, for practical purposes, equally liked. One displays four stars and the other four and a half. Any difference in their fortunes is caused by the display, not by quality.
Luca used this to estimate that a one-star increase in displayed Yelp rating "leads to a 5-9 percent increase in revenue", an effect driven by independent restaurants.22 Michael Luca, Harvard Business School Working Paper 12-016, Reviews, Reputation, and Revenue: The Case of Yelp.com
Anderson and Magruder found that "an extra half-star rating causes restaurants to sell out 19 percentage points (49%) more frequently".33 Anderson and Magruder, The Economic Journal 122(563), 2012, Learning from the Crowd: Regression Discontinuity Estimates of the Effects of an Online Review Database They also checked whether restaurants were gaming their way just over the thresholds and reported that restaurants "do not manipulate ratings in a confounding, discontinuous manner".
Both papers date from the early 2010s and concern restaurants specifically. The mechanism, a display that jumps while quality does not, applies wherever ratings are rounded.
How much effort crosses a cliff?
Suppose a business sits just under a threshold and receives only five-star reviews from now on. How many does it need?
For a business at average a with n reviews to reach target t using k five-star reviews:
(a × n + 5k) ÷ (n + k) ≥ t, which rearranges to k ≥ n × (t − a) ÷ (5 − t)
| Current reviews | From 4.20 to 4.25 | From 4.70 to 4.75 | From 4.44 to 4.45 (one decimal) |
|---|---|---|---|
| 20 | 2 | 4 | 1 |
| 50 | 4 | 10 | 1 |
| 100 | 7 | 20 | 2 |
| 200 | 14 | 40 | 4 |
| 500 | 34 | 100 | 10 |
Three observations.
Cost scales with history. A business with 500 reviews needs 34 perfect reviews in a row to climb five hundredths. One with 20 needs two.
Higher cliffs are steeper. Moving from 4.70 to 4.75 takes about three times as many reviews as moving from 4.20 to 4.25, because each five-star review is closer to the current average and lifts it less.
One-decimal cliffs are cheap. On a one-decimal display, moving from 4.44 to 4.45, which changes the printed figure from 4.4 to 4.5, takes a handful of reviews even for a large profile.
The incentive this creates
A business sitting at 4.23 on a half-star display is two or three good reviews from a visibly better rating. The Yelp research found that restaurants were more likely to commit review fraud when their reputation was weak, with few reviews or recent bad ones.44 Luca and Zervas, Management Science 62(12), 2016, Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud Rounding thresholds sharpen that temptation at specific, predictable points.
It also works in reverse. A business at 4.26 is one bad review from dropping half a star in the display. For a 20-review business, a single one-star review takes the average from 4.26 to 4.10.
What it means for reading
Half-star displays hide a lot. Everything from 4.25 to 4.74 looks identical. If the platform also prints the decimal, read the decimal. If it does not, a four-and-a-half-star business could be nearly a five or barely above a four.
Do not over-read one-decimal differences either. A 4.5 and a 4.4 may be 4.45 and 4.44. And either figure carries a margin of error larger than 0.1 unless the review count is in the hundreds; see how many reviews are enough.
Watch for profiles parked just above a threshold. A business at exactly 4.8 with many reviews is unremarkable. We would not read anything into a single case, and the Anderson and Magruder check found no systematic bunching in their data. It is one more reason to read reviews and dates instead of leaning on the digit.
A note on truncation
Some displays truncate instead of rounding, turning 4.79 into 4.7. Others fill a fraction of the fifth star in proportion to the average. Unless the platform says which it uses, you cannot tell from the display alone. Of the platforms covered on this site, Trustpilot states its rounding rule; for Google's star graphic and Yelp's current display, the help pages we read do not specify.
Questions readers ask
Is 4.25 really displayed differently from 4.24?
Under rounding to the nearest half star, yes: 4.24 rounds down to 4 and 4.25 rounds up to 4.5.
Do rounding thresholds affect sales?
Two studies of Yelp restaurants found they did, with effects on revenue and on how often restaurants sold out.22 Michael Luca, Harvard Business School Working Paper 12-016, Reviews, Reputation, and Revenue: The Case of Yelp.com33 Anderson and Magruder, The Economic Journal 122(563), 2012, Learning from the Crowd: Regression Discontinuity Estimates of the Effects of an Online Review Database
How many five-star reviews does it take to raise a rating by 0.1?
It depends on the current count and level. From 4.6 to 4.7 with 100 reviews takes 34 consecutive five-star reviews; with 20 reviews it takes 7.
Method The threshold arithmetic is ours. Evidence on the effects of rounding comes from two published regression-discontinuity studies of Yelp. Trustpilot's rounding rule is quoted from its help center. 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.
- Trustpilot Help Center. TrustScore and star rating explained
- 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
- Luca and Zervas, Management Science 62(12), 2016. Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud