How Our Fake-Review Score Works (and What It Can't Tell You)

The score is a 0-100 number built from a business's entire public Google review history. Four weighted categories feed it: clustered posting times, the share of reviews written by accounts with no history, text repeated across reviews, and star ratings that contradict the words next to them. Under 30 is Clean, 30 to 59 Suspicious, 60 and above Likely Manipulated. The verdict comes from the combination, never from one signal firing alone.

The point is not to judge any single review. It is to do in seconds what a careful reader would do across three thousand reviews if anyone had the afternoon. To run the checks yourself first, the field guide covers the manual version.

#What signals make up the score?

Four categories, weighted rather than summed, because some patterns carry more evidence than others.

Review bursts. Clusters of reviews, usually 5-star, arriving in a short window measured against the place's own pace. A sharp break from that baseline moves the score more than a slow drift upward. Burst markers sit on 12,259 flagged reviews in our data, second in frequency. The piece on burst days covers the detection in full.

Low-activity accounts. For each reviewer we read the visible history: how many reviews in total, and how scattered across unrelated businesses. One or two lifetime reviews weighs heavily, especially when several such profiles land on one business in the same fortnight. This marker leads our data at 37,061 flagged reviews, more than the other two together.

Duplicate text. We compare review text across the business's history, and at lower weight across other analyzed businesses, looking for exact and near-exact phrase matches. Reviews-for-hire operations recycle short phrases; genuine reviewers essentially never share sentences. 5,958 flagged reviews carry it.

Rating-text mismatch. Five stars with nothing specific in the text, or one star reading like a form letter with no incident in it. Small weight alone, considerably more when the other three appear alongside.

#Is every high-scoring pattern the same kind of manipulation?

No. A smaller set of listings score high for a specific reason: real customers, prompted by staff in a compressed window, to leave a five-star review naming an employee. That pattern gets its own Review Drive verdict rather than folding into Likely Manipulated, since the reviewers are typically genuine even though the solicitation itself breaks platform rules. What a review drive looks like in the data walks through real examples and how the scoring treats it differently.

#Why weight the signals instead of just counting them?

Because one thin reviewer and one vague rave happen constantly in honest review histories. Co-occurrence separates a Clean business from a Suspicious one: several markers on the same reviews, in the same window, at a rate that breaks the business's own baseline. Weighting lets a rare, strong marker like repeated text move the number more than a common, weak one, and it is the reason the score reads the whole history rather than the most recent page.

#What does the score distribution actually look like?

By July 2026 we had completed 242 analyses covering 241,041 reviews, average score 32.1 out of 100.

Most places score low, with 120 of 242 landing in the 10-19 band and a smaller group gathering in the 60-79 range. The full data release breaks the histogram down band by band.

#What can't the score tell you?

Everything it uses is public: review text, ratings, timestamps, reviewer profile metadata. IP addresses, device fingerprints, the money trail between a vendor and a business, how an account was created: none of it reaches us. Four consequences follow.

• It cannot identify who wrote a specific review, or prove any single review is fake. It reports patterns across the set.
• It cannot prove a business purchased reviews. A high number reflects patterns consistent with manipulation. Intent and transactions are beyond it, which is why we word our write-ups the way described below.
• It can be wrong in both directions. A new location with a loyal, first-time-reviewer following can score higher than it deserves. A well-funded campaign using aged accounts and varied text can score lower than it deserves.
• It ages. Reviews keep arriving, so a score is a snapshot of a moment, worth rerunning rather than treating as settled.

Academic work on fake-review detection lands in the same place. Behavioral and contextual signals, timing, reviewer metadata, text similarity, beat text analysis alone, and no published method claims certainty about an individual review. Every one of them talks in probabilities.

The rule prohibits fake reviews and testimonials, including AI-generated reviews, suppression of negative reviews, and paying for positive or negative reviews.

Even the FTC, framing its own 2024 rule, builds enforcement around conduct over time rather than one review pulled out of a listing. What the rule bans and how enforcement has gone is covered separately.

#How we write about individual places

Since the score cannot establish intent, we never state as fact that a named business bought or wrote reviews. Writing about a specific place, we say what the analysis flagged, a spike, a concentration of empty accounts, and link the public report so anyone can read the same evidence and decide for themselves.

The model reads Google Maps data only. How Google's moderation compares with Yelp's and Tripadvisor's is covered in the platform comparison, and how this approach sits against other free tools, Fakespot's closure included, in the checker roundup.

#Frequently asked questions

#Is a manipulation score the same as an accusation?

No. It describes statistical patterns in a public review history that are consistent with manipulation. It is not a legal finding, and it does not establish that a particular review is fake or that anyone paid for anything.

#Can a business's score change over time?

Yes. New reviews arrive constantly, so rerunning the analysis periodically makes sense, particularly for places sitting near the Clean/Suspicious line.

#Why is "low-activity account" the most common signal instead of duplicate text?

Because it is the hardest trace to avoid leaving. Wording varies for free; accounts with real history do not, so review sellers work with profiles that have almost nothing behind them.

#Does a low score guarantee a business has never had a fake review?

No. It means the history shows no clustering strong enough to cross the Suspicious threshold, which is a different claim from zero fakes.

#Can I see which specific reviews were flagged?

Each public report lists the evidence behind the score, the signal types found and their weight, without exposing reviewer identities.

Paste a Maps link into the analyzer and the same model runs on that listing, with the signal breakdown returned in seconds.