Amazon Fake Reviews: Detection Signals, Enforcement Actions and Impact Metrics

Assume that roughly 40% of the product reviews on Amazon are fake. You are a product manager at Amazon and have been asked to tackle the problem. Structure your answer in three parts: how you would recognize that a review is fake, what action you would take, and what impact you would expect if the number of fake reviews went down. ### Clarifying Questions - How was the 40% figure produced: a hand-labeled audit sample or a model estimate, and is it a share of all reviews or of the reviews shoppers actually see? - Are the fake reviews mostly positive ones that inflate a seller's own listings, negative ones aimed at competitors, or a mix? - Can anyone with an account post a review today, or only customers who bought the item? - Which outcome matters most to leadership: shopper trust and conversion, fairness to honest sellers, or legal and reputational exposure? ### Part 1 — Recognizing a fake review How would you decide whether a particular review is fake? ```hint Follow the trail Consider what a genuine reviewer leaves behind before, during and after writing a review that a fake reviewer usually cannot reproduce. ``` #### What This Part Should Cover - Signal families beyond the review text: purchase history, account behavior, timing, and links between accounts and sellers. - How the signals are combined into a score, and where trustworthy labels come from. - How accuracy is measured, separating precision on flagged reviews from coverage of all fakes. ### Part 2 — Deciding what action to take Once you can identify fake reviews, what would you do about them? ```hint Match the response to the certainty A wrong call costs different people different amounts. Think about how sure you are, who wrote the review, and who benefited from it. ``` #### What This Part Should Cover - Responses graded by confidence and severity, aimed at the review, the reviewer and the seller who benefits. - Prevention at the point of posting as well as cleanup of the existing backlog. - Appeals, false positives, and how fraudsters will adapt to the response. ### Part 3 — The impact of fewer fake reviews If the number of fake reviews drops, what impact would you expect? ```hint Some numbers fall for good reasons Removing fakes changes how shoppers, honest sellers and dishonest sellers behave. Think about which metrics could get worse even though the marketplace got healthier. ``` #### What This Part Should Cover - Expected effects on shoppers, honest sellers, dishonest sellers and the platform, in the short and the long term. - The metrics that would show the effect, including fake-review prevalence itself. - How to attribute changes to the effort rather than to seasonality or other launches. ### What a Strong Answer Covers - An understanding of who writes fake reviews and why, used to drive both detection and the choice of actions. - Explicit treatment of false positives against false negatives, and who bears the cost of each. - An adversarial mindset: the system has to keep learning, and prevalence has to be re-measured over time. - A clear thread through the three parts: detection quality determines which actions are safe, and the actions determine which impact metrics move. ### Follow-up Questions - Your model catches seller-run accounts well but misses real buyers who were refunded in exchange for five-star reviews. How would you find those? - After the cleanup, average star ratings fall and conversion drops in some categories. How would you explain this to leadership? - How would you know prevalence is really falling rather than fraudsters getting better at evading detection?