What the Platforms Remove

Fabricated third-party endorsement is old enough to be uninteresting as a moral matter and recent enough, in its industrialised form, to be badly measured. The commentary on it is dominated by a small number of large figures whose origins are difficult to establish. The enforcement disclosures published by the platforms themselves are less quoted, more specific, and considerably more useful, provided one is careful about what they actually count.

The disclosed numbers

Yelp’s 2024 Trust and Safety Report, published in February 2025, states that of roughly 21 million reviews contributed to the platform in 2024, 18% were not recommended by Yelp’s automated filter, with a further 3% removed by staff and more than 551,200 accounts closed.

Tripadvisor’s 2025 Transparency Report, covering 2024, states that of approximately 31.1 million submitted reviews, 2.7 million were blocked as fraudulent, and 54% of the fraud identified was “review boosting” — owners, employees or affiliates reviewing their own establishments. A further 214,000 machine-generated reviews were removed.

Amazon’s 2025 Trustworthy Shopping Experience Report says the company blocked hundreds of millions of suspected fake reviews during the year and shut down more than 40 fake-review brokers. The imprecision is Amazon’s own, and it is more honest than the precise figures that circulate in its place.

What these figures do and do not establish

Each of these is an enforcement statistic. It records what a detection system flagged, not what existed. Two inferences are therefore unavailable.

The first is a prevalence rate. Nothing in Yelp’s 18% tells us what share of submitted reviews were in fact fraudulent, because the filter has both false positives and false negatives and neither is disclosed. A commonly repeated claim that a fixed 8% of Tripadvisor reviews are fake is a ratio computed by outside readers from the company’s totals; Tripadvisor does not state it, and it should not be quoted as though it did.

The second is a trend in fraud. A platform that improves its detection reports more removals while the underlying problem is shrinking, and a platform that lets detection lapse reports fewer while it grows. Enforcement counts move with enforcement effort, and the two cannot be separated from outside.

What the figures do establish is a floor and a shape. The floor is that manufactured endorsement occurs at industrial volume on every platform that publishes numbers at all. The shape is more interesting, and it comes from Tripadvisor’s breakdown: the largest single category of detected fraud was not a broker network or a competitor attack but the business writing about itself through accounts that did not look like its own. The commonest form of fabricated endorsement is the subject praising itself under another account.

The regulatory line

The United States now has a rule that draws the boundary in usable terms. The Federal Trade Commission’s rule on consumer reviews and testimonials, 16 CFR Part 465, effective 21 October 2024, prohibits fake or misattributed reviews, the purchase of positive or negative reviews, undisclosed insider reviews, the suppression of unfavourable ones, and the sale or purchase of fake indicators of social-media influence. It also, at §465.6, prohibits a company from operating a review site that presents itself as independent of that company.

That last provision is the one worth dwelling on, because it identifies the structural offence rather than the false statement. A company-run site that reviews the company’s own products favourably need contain no factual inaccuracy. The misrepresentation is about provenance. What is being sold to the reader is the appearance of an independent second reading, and the reason that appearance has value is precisely that it is supposed to be unavailable to the subject.

The number that travels

A figure of $152 billion in global spending influenced by fake reviews appears widely, frequently credited to the World Economic Forum. The attribution is wrong. It is a modelled 2021 estimate produced by CHEQ with the University of Baltimore, derived by applying an assumed 4% fake-review rate to roughly $4.3 trillion of e-commerce activity. The assumed rate is itself drawn from platform self-reporting, which is the thing the model was meant to illuminate.

The figure is not fraudulent and the modellers stated their method. What happened to it afterwards is the instructive part. Each retelling shortened the provenance and raised the authority of the source, until a 2021 vendor-academic estimate was being cited as a finding of an intergovernmental body. No one falsified anything. The chain simply got longer, and every link dropped a qualifier.

This is the same failure the review rule addresses, performed on a statistic instead of a testimonial. In both cases an assertion acquires weight it was never given, by appearing to have come from somewhere other than where it came from.

Reading a body of favourable material

For anyone assessing a set of positive third-party notices, the useful questions concern origin rather than quantity, and they can be answered from outside.

Was each notice arrived at independently, or does the set share an author, a template, or an arrival date? Does a given notice say what it looked at and against what standard, so that its verdict can be tested instead of accepted? Is anything in the set unfavourable, hedged, or merely dull, and if not, what process removed it?

The B2B market has already priced in the doubt. TrustRadius and Pavilion’s 2024 study of 2,164 technology buyers found that 73% believe they regularly or sometimes encounter fake reviews, and that vendor-supplied references rank as a top information source for only 27% of them. Manufactured endorsement does not merely fail to convince; it devalues the genuine article alongside it, which is a cost borne by every organisation in the category rather than by the one that manufactured the evidence.