Why single reviews defeat every detector
A typical product review runs 30 to 80 words. That is far below the roughly 150 words a statistical detector needs before its verdicts mean anything — for us and for every competitor, whatever their marketing says. A tool that stamps a confident percentage on a 40-word review is guessing. Cobalynx refuses to, and tells you so. That sounds like a dead end for review checking; it isn't — because fake-review operations almost never write just one.
The batch method: test the pattern, not the review
Mass-produced reviews are generated from the same prompts, the same templates, the same model — and that shared origin leaves statistical sameness a human crowd doesn't produce. So flip the unit of analysis:
- Open the reviews you distrust — a suspiciously glowing product page, or one prolific reviewer's history.
- Copy five to ten of them, from the same product or the same reviewer, into one paste. Together they clear the length floor.
- Run the free scan. You're now asking the meaningful question: does this batch read like many independent people, or like one generator with many names?
- Read the result as calibrated evidence — a probability with a published error rate — alongside the manual signals below, never instead of them.
The same move works on seller messages and outreach mail — five near-identical “personal” messages pasted together reveal the template a single message hides; the email guide covers that variant.
The manual checklist detectors can't replace
- Timing bursts. Twenty five-star reviews inside one week on a product listed for years is a campaign, whoever wrote it.
- Empty reviewers. Profiles with one review, no history, and no verified purchase — check a few authors, not just the text.
- Interchangeable praise. Reviews that never mention a concrete use, flaw, or comparison — text that could apply to any product usually was written for any product.
- The rating-review gap. A high star average built almost entirely from unreviewed ratings, or reviews for a different product entirely (a listing-swap tell).
The slop wave: listings, patterns, and product photos
Reviews are one front of 2026's marketplace problem. AI-generated listings — “crochet patterns” that can't be crocheted, seed varieties that don't exist, product photos of objects never manufactured — pair generated text with generated images. The images are the checkable half: save a listing photo and run it through the file check. C2PA credentials recording an AI generator are near-conclusive; no credentials means unknown, not authentic — the Content Credentials guide explains the asymmetry. A listing whose photos are provably synthetic and whose reviews batch-flag as uniform has told you everything you need before your money moves.
Report it where it counts
Platforms act on patterns, not vibes: report the specific evidence — identical phrasing across reviews, the timing burst, the generated listing photo. Amazon (“Report abuse” on a review), Google Maps (“Report review”), Etsy (report listing/shop), and app stores all take per-item reports; consumer-protection agencies take the bigger cases. One good pattern report outweighs ten “this looks fake” clicks.
What Cobalynx can and can't do here
One 40-word review is below every detector's reliable floor — ours refuses to guess on it, and asks you to paste a batch instead. And an AI-flagged batch is not automatically fake: real buyers polish honest reviews with AI, and sellers translate genuine feedback through it — just as a human-sounding review can be a paid lie. Detection informs a judgment; it isn't one. Our numbers and their limits are published on the evidence page, and the method behind them on the methodology page.
Common questions
How many reviews should I paste at once?
Enough to pass roughly 150 words — in practice five to ten reviews from the same product or the same reviewer, pasted together as one text. You are not testing any single review; you are testing whether the batch shows the statistical sameness that mass-generated text produces and independent humans do not.
Does an AI flag mean the reviews are fake?
Not by itself: real buyers polish honest reviews with AI, and sellers translate genuine feedback through it. The reverse also holds — a human-sounding review can be a paid lie. Detection tells you about the text's statistical origin; combine it with reviewer profiles, timing, and specificity before deciding what to trust.
Can I check the product photos too?
Yes — download a listing image and run it through the Cobalynx file check: a C2PA manifest recording an AI generator is near-conclusive evidence the “product photo” was synthesized. No credentials means unknown, not authentic — most real photos carry none.
How do I report suspicious reviews?
On the platform itself: Amazon and Google both take reports on individual reviews (“Report abuse” / “Report review”), Etsy takes reports on listings and shops, and app stores on individual ratings. Report the pattern you found — same phrasing across reviews, burst timing — rather than just “looks fake”; pattern reports are the ones moderators can act on.