Trustpilot removed 4.5 million fake reviews in 2024. It has never said how many were one star.

On 27 March the UK's Competition and Markets Authority opened its first five fake review investigations under the DMCC Act. Four are what you would expect: staff writing their own praise, discounts traded for five stars, a delivery app accused of inflating restaurant ratings. The fifth is Autotrader, and the allegation runs the other way. The CMA is examining whether one star reviews moderated by Feefo were kept off the platform and out of the star rating.

That is the case worth watching, because the rest of the economy has quietly stopped caring about fake five stars. It has started betting everything on the star rating itself.

A brand with no reviews shows up in 1% of AI answers

Research Trustpilot commissioned from Seer Interactive, published in May, ran more than 15,000 buying journey prompts across ChatGPT, Gemini, Perplexity and Google AI Mode: 800,000 responses, 1,926 brands. Brands with no active review profile appeared in 1% of AI answers. Brands with a high volume, actively managed profile appeared in 75.3%. The same release puts review and trust sites second among all cited source types, at 14% of citations.

Note who paid for that. Trustpilot funded research proving you need Trustpilot. The direction still matches independent work: SE Ranking found that 34.5% of AI Overviews across 30,000 commercial keywords cited at least one review platform, with Gartner Peer Insights, G2, Capterra, Software Advice and TrustRadius carrying 88% of those citations between them.

The platforms being quoted have lost their readers

From the same SE Ranking data, across 2024 and 2025: G2 lost 84.5% of its traffic, Gartner Peer Insights 76.5%, Capterra 89%, TrustRadius 92.2%. Nobody is reading them. Everybody is quoting them.

A review written today is written for a machine that cannot ask a follow up question. We have made a version of this argument before, when private equity bought the meeting count, when a free event registration turned out to be a manufactured number, and when Reddit's citation collapse turned out to be unverifiable by anyone outside the companies involved. When the number that proves a channel works is produced by the party selling the channel, it stops being evidence and starts being inventory.

7.4% of submissions were fake, and no breakdown

Trustpilot's own Trust Report is blunt about the scale. It published 61 million reviews in 2024 and removed 4.5 million as fake, which it puts at 7.4% of submissions, up from 6.1% the year before, with 90% of those removals made automatically.

What the report does not contain is a split by star rating. Every transparency number in this category counts brands inflating themselves. The cheaper attack, buying one stars against somebody else, does not appear in the disclosure at all. Trustpilot's own policy names competitor targeting as a violation. Its report does not say how often it happens, and no platform in the category publishes that number either.

So the brand team that spent $500,000 perfecting how the company looks now has its reputation set by a moderation queue it cannot see, audited by nobody, and quoted back by all four systems in that study.

Free to report, $799 to bury

Flagging a review is free on Trustpilot, on every tier including the free one, and that is the part critics usually get wrong. What costs money is volume. The free plan includes 50 automated review invitations a month. After that it is $99, $319 or $799 per month per domain.

Read that as a business under attack: the slow remedy is free and the fast one is a subscription. The lie costs nothing to report and the burial is billed annually.

None of this makes Trustpilot the villain of the story. It makes the star rating a bad primary source, which is a different and more serious problem, because the star rating is now a primary source for the four systems that increasingly decide which vendors a buyer ever hears about.

The correction is coming

The rise is real and it will be brief. Model makers will end up discounting review platform ratings the way search engines eventually discounted link farms, and for the same reason: a signal that anyone can manufacture at either end, on a platform that will not publish the shape of the manipulation, is not a signal. It is a number with a marketing budget attached.

Which leaves the same conclusion this publication keeps arriving at from different directions. The proof that survives the correction is the kind that cannot be submitted through a form. A named person, in a room, saying what actually happened to them, to a buyer who can ask a second question and watch the face that answers it. That has no API, no moderation queue and no price per domain, which is exactly why it is the last thing left that cannot be bought in bulk.