How do you find reviews nobody paid for?

For a decade the standard answer to fake reviews was detection after the fact: paste a product link, get a letter grade on the reviews. That category has collapsed. Mozilla retired Fakespot, its fake-review detector, on 1 July 2025, citing economics. My read of the deeper cause: scoring text after publication was losing its arms race, because generated reviews now pass for tired human ones.

Admission works where detection failed. Instead of asking “is this review fake?”, ask a question a machine can actually answer: “did this sentence come from a place, a speaker, and a form that paid speech uses?” Kill everything that matches, and read what is left.

Gate 1 · Community sources only. The pipeline reads Reddit, Hacker News, X, Trustpilot reviews, YouTube comments, Quora, Bluesky, niche forums. Blogs, press releases, and official pages are refused as evidence. The single exception is deliberate: the company’s own page is fetched to anchor the PROMISED side. The company gets to state the promise. Testimony decides what happened to it.

Gate 2 · The marketing kill. Every candidate quote from every source is screened for the fingerprints of paid speech: brand voice, self-description, calls to action, affiliate links, emoji-studded promos, discount codes. One hit and the quote dies. This gate runs on community platforms too, because marketing long ago learned to dress casual.

Gate 3 · Relevance and entity match. A quote must demonstrably concern the claim’s subject. Zero hits on the claim’s terms: rejected. The entity token missing where the platform is noisy: rejected. This is the gate that keeps a record about one company from quietly filling with another company’s complaints.

Gate 4 · Recency, weighed and disclosed. Fresh testimony rises. A 2019 wound sinks below a 2026 one, but it stays in the record, dated, because it happened. the Trustpilot guide shows what review platforms do with old pain: it gets deleted or drowned. Here it stays, dated, and sinks only as far as its age.

Gate 5 · Grounding. Every number in a finding must exist in the claim, the official page, or an admitted quote. A headline with an ungrounded number gets rebuilt. Fewer than two admissible quotes and the report says so in those words instead of performing certainty (the thin-evidence guide is about exactly that case).

Stated once: read any sentence on the internet while asking who profits if I believe this, and the five gates fall out by themselves. Ranked and hosted speech can be bought, so sourcing goes community-only. Paid speech dresses casual; the marketing kill screens for the costume. Confusion between two companies always benefits one of them, hence entity matching. Reputations get sold on their best year, which is why testimony is weighed by date. And an invented number costs nothing to type, so nothing ships unless it traces to a quote.

On 9 July 2026, building the record for the Trustpilot guide, we ran the claim Revolut is rated 4.7 Excellent on Trustpilot. The pipeline returned 19 quotes; our audit found 8 of them were about the wrong companies, because “4.7” is a number every rated business on the internet shares. A second phrasing came back worse: it had drifted into stating manipulation as fact where testimony only alleges it. Both checks were killed and logged, about 16 Serper credits burned, and the claim was rebuilt on Revolut’s own promise, relax, your money is safe, which produced 38 clean reports.