Review Manipulation Flags

Why the restraint is the product

Every rule here can accuse an honest business, and most of the design is about not doing that.

The duplicate threshold is high on purpose. Every genuine review of a hotel says room, staff, breakfast, location. A lower bar flags the most ordinary sample there is — and a tool that accuses honest businesses gets switched off, taking the real detection with it.

The burst rule refuses to fire when it cannot tell. Only the newest page is ever sampled, and a popular venue's newest ten reviews are ALWAYS within days of each other. The naive "half in one week" test therefore fires hardest on the most-trusted places you have. So it only applies when a normally long-spanning sample suddenly clusters — and never for a venue big enough that a cluster is just a busy week.

Missing data is never evidence. Author history is optional; a source that does not expose it cannot trip the throwaway tell, rather than tripping it for everyone.

A flagged review is halved, not deleted. The score stays inside the range of ratings actually present. It re-reads the sample; it does not model a hidden population of "real" reviews, which is a claim the input cannot support.

Thin samples get no score. Below six reviews `trueScore` is undefined and only the plain average is returned, because a re-weighted mean of four reviews is arithmetic, not evidence.

The tells

| flag | fires on | |---|---| | `template` | near-identical review text across accounts — the paid-farm signature | | `throwaway` | 1★/5★ from accounts with ≤2 lifetime reviews, at scale | | `gush` | very short superlative-only 5★ with no specifics | | `burst` | a long-spanning sample that suddenly clusters into one week | | `bimodal` | almost nothing but 1s and 5s — a fought-over or brigaded listing | | `crosssite` | a large gap between two platforms' averages for the same place |

What it does not do

It does not fetch reviews — sourcing is yours, which means this runs over data you already hold without a single outbound request. It does not decide a review is fake; it says a review matches a known tell and counts it for less. And it does not model the reviews it cannot see.

Verified

20 tests, 10 deliberate defects, all 10 caught. The suite is written from both directions — each tell must fire on the shape it names AND stay silent on the honest shape that resembles it — because a false flag on a real business is the expensive failure here, not a missed fake. Caught defects include a collapsed duplicate threshold, absent author history read as suspicious, the burst rule losing its can-we-even-tell guard, and a flagged review being deleted from the score rather than halved.

Delivery

Source delivered as a private repository invite within 24 hours of purchase. Single-product commercial license: use and modify in any number of products; no redistribution or resale of the source.

Interface

What you call, and what comes back. Types and signatures only — the implementation ships with the source.

  export function analyzeReviews(reviews: SampledReview[], crowdCount?: number): ReviewAnalysis;
  export function crossSiteFlag(google?: number, yelp?: number): TrueFlag | null;

  export type MapsReview = SampledReview;