RATINGS/DECODED

Math / WORKED EXAMPLE

Bayesian averages: what seven phantom reviews do to a rating

A Bayesian average mixes a business's real reviews with a fixed number of imaginary neutral ones. Trustpilot uses seven at 3.5 stars. With 7 real reviews the phantoms carry half the weight; with 63 they carry a tenth; with 700 they are 1%. The effect is to distrust small samples automatically.

Trustpilot's help center says its weighted average "includes 7 neutral" reviews valued at 3.5 stars, to "ensure all companies start off with a balanced TrustScore", and that this "has less impact as more reviews come in".11 Trustpilot Help Center, TrustScore and star rating explained

That one sentence describes a classic statistical device. Here is how it behaves.

The formula

With n real reviews averaging a, a prior of m imaginary reviews at value p gives

adjusted = (m × p + n × a) ÷ (m + n)

For Trustpilot's stated prior, m = 7 and p = 3.5, so m × p = 24.5:

adjusted = (24.5 + n × a) ÷ (7 + n)

A business that only gets five-star reviews

Real five-star reviewsPlain averageAdjustedWeight on real reviews
0none3.500%
15.003.6912%
25.003.8322%
55.004.1342%
105.004.3859%
205.004.6174%
505.004.8288%
1005.004.9093%
3005.004.9798%
1,0005.004.9999%
Adjusted score for a business with only five-star reviews3.43.84.24.653.503.6914.1354.38104.61204.82504.901004.97300Number of real five-star reviewsAdjusted score
Our calculation of (24.5 + 5n) ÷ (7 + n). A flawless record needs 20 reviews to pass 4.6 and 100 to reach 4.9.

A perfect business does not reach a displayed 4.5 until its 14th review, and cannot round to 5.0 at one decimal until it has 203. The prior demands evidence in proportion to the claim.

It protects in both directions

The same arithmetic softens a bad start.

Real one-star reviewsPlain averageAdjusted
11.003.19
21.002.94
51.002.46
101.002.03
501.001.31

A new business hit by one angry customer shows 3.19, not 1.0. That is a fair reflection of how little a single review proves.

The weight rule

The share of the score that comes from real reviews is n ÷ (n + 7).

  • 7 reviews: half.
  • 21 reviews: three quarters.
  • 63 reviews: nine tenths.
  • 700 reviews: 99%.

So a prior of seven behaves like a statement: "until you have a few dozen reviews, I am mostly guessing". That matches the sampling arithmetic. With typical spread, the margin of error on an average at 20 reviews is about 0.4 stars, which is the range over which the prior is still doing visible work.

Why 3.5 and why seven?

Trustpilot calls 3.5 "neutral". It is half a star above the midpoint of a 1 to 5 scale. The company does not explain the choice of seven, and we will not guess at its reasoning.

What can be said is how the two parameters behave.

The prior value sets where newcomers start. At 3.5, a new business looks worse than most established ones, since review averages tend to bunch near the top of the scale.22 Schoenmueller, Netzer and Stahl, Journal of Marketing Research 57(5), 2020, The Polarity of Online Reviews: Prevalence, Drivers and Implications A prior nearer the platform-wide average would be gentler on newcomers. A lower one would be harsher.

The prior count sets how long the cushion lasts. Seven is light. A prior of 25 would keep a 10-review business near the middle: ten fives would display as 3.93 instead of 4.38. A heavier prior is safer for readers and slower for honest new businesses.

There is no correct answer. It is a design decision about whose risk to favor: the reader who might trust a tiny sample, or the newcomer who has earned its early praise.

Using it on platforms that do not

Google describes its score as a plain average of published ratings, with no prior.33 Google Business Profile Help, Understand review scores for local places and businesses You can add one in your head. For a rough version:

  1. Multiply the displayed average by the number of reviews.
  2. Add 24.5.
  3. Divide by the number of reviews plus 7.
DisplayedReviewsAdjusted
5.044.05
4.984.25
4.8124.32
4.7404.52
4.63004.57
4.52,0004.50

Ranked by displayed average, the order runs top to bottom. Ranked by adjusted score, it nearly reverses: the 300-review 4.6 leads, then the 40-review 4.7, then the 2,000-review 4.5, with the three tiny profiles last. Neither ranking is the truth. The adjusted one is the ranking you can defend with the evidence available.

What a prior does not do

It does not weight recency. Trustpilot handles that separately, giving newer reviews more weight.11 Trustpilot Help Center, TrustScore and star rating explained Our tables ignore that step, so they will not match a live TrustScore exactly.

It does not detect fakes. Twenty fabricated five-star reviews overwhelm seven phantoms as easily as twenty real ones.

It does not fix selection. If only pleased customers are invited, the prior delays the inflated score. It does not prevent it.

Relation to confidence intervals

A confidence interval says "the average is 4.8, and I am unsure by this much". A Bayesian average says "given how little I know, my best single guess is 4.3". They are two responses to the same shortage of data. The interval keeps the uncertainty visible. The adjusted average folds it into the number, which is convenient for sorting and hides how unsure you are. For the interval approach see confidence intervals for star ratings.

Questions readers ask

Does Trustpilot really add fake reviews to a profile?

No reviews are added to the page. The seven neutral reviews exist only inside the calculation of the TrustScore.11 Trustpilot Help Center, TrustScore and star rating explained

When do the seven neutral reviews stop mattering?

They never reach exactly zero weight, but by 63 real reviews they are a tenth of the calculation and by 700 about 1%.

Is a Bayesian average more accurate than a plain one?

For small samples it is a better guess, on average, because extreme early results are usually partly luck. For large samples the two are nearly identical.

Method The prior of seven reviews at 3.5 stars is Trustpilot's published figure. All other numbers are our calculations of the prior's effect alone, leaving out Trustpilot's recency weighting, whose parameters are not published. 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.

  1. Trustpilot Help Center. TrustScore and star rating explained
  2. Schoenmueller, Netzer and Stahl, Journal of Marketing Research 57(5), 2020. The Polarity of Online Reviews: Prevalence, Drivers and Implications
  3. Google Business Profile Help. Understand review scores for local places and businesses