How to spot fake reviews
By ReputationKiln Editorial · Published · Updated
You cannot tell a fake review from a real one by reading it, and that is the single most important thing to accept before you start. Reviews written by current AI are, in one 2025 preprint (not yet peer-reviewed), hard to tell apart from genuine ones to both people and the detectors built to catch them. So the reliable signals are not in the words, they are in the pattern: when the reviews arrived, how they are spread, and who wrote them. No single one of these proves anything. Several together, across different categories, is when you slow down.
Run the checklist below, mark each green, amber or red, and treat a profile as genuinely suspicious only when multiple reds land across different categories at once.
The checklist, each with its limit
- Velocity. Sort newest-first and look at the rhythm. A long quiet stretch broken by a sudden cluster, especially a wall of five stars from a standing start, is a flag. Limit: a real promotion, a viral moment or a "please review us" email also cause a spike.
- Date clustering. Many reviews on the same day or in a narrow window, particularly if only the five-stars cluster. Limit: seasonal trades and email campaigns cluster genuinely too, and many platforms only show "months ago," which is imprecise.
- The spread. Compute roughly how many are three or four stars versus the total. Eighty or ninety per cent five-star with almost no middle, or a sharp five-and-one split, is the shape of a manufactured sample. Limit: genuinely polarising trades and small new samples skew.
- Templated language. The same openings, brochure-like repetition of the brand name, identical structure across reviews. Limit: real customers echo each other ("friendly, on time"), and you cannot reliably tell AI from human by feel.
- Thin reviewer profiles. New accounts, only ever one all-five-star review, or serial five-star reviewers hitting unrelated products fast. Limit: privacy-minded real customers keep bare profiles, and a small firm's earliest reviewers are first-timers.
- Content specificity. Do the positives name the actual job, the date, the staff member, or just say "amazing"? Are the negatives specific while the positives are vague? Limit: some happy customers simply say "great job."
The limit, and what to do with it
Cross-check the average against the spread, not the headline number, and check a second platform the business cannot curate. One red is a coincidence waiting for an explanation; four reds pointing the same way is a pattern. The hard limit is that detection is imperfect and tools that score a review as fake are an input to judgement, never a verdict. The most reliable test of all is not on this list: ask the business to show you proof it did not produce itself, and watch what happens.
Sources
- Reviews written by current AI are, in one 2025 preprint (not yet peer-reviewed), hard to tell apart from genuine ones to both people and detectors, so the reliable signal is the pattern over time, not the wording. — Large Language Models as Persuaders (preprint, 2025). https://arxiv.org/html/2506.13313v1 · checked 2026-06-04