Trust in Reviews: How Many Online Reviews Are Fake or AI-Generated in 2026?
Independent 2026 measurements put AI-generated reviews at 3% of front-page Amazon reviews, 19% of Google reviews and over 26% on some B2B software sites. Google alone removed 292 million policy-violating reviews in 2025. In controlled tests, people identify machine-written reviews 50.8% of the time — no better than chance.
- People do not reliably identify AI-written reviews by reading them. Humans scored 50.8% when separating authentic reviews from machine-written ones, while large language models performed no better.[2] In contrast, 51% of US shoppers believe they can identify AI-written reviews.[3]
- A "Verified Purchase" badge does not establish authenticity. In a May 2026 scan, 93% of the Amazon reviews identified as AI-generated carried the badge.[1]
- AI-generated reviews appear at both ends of the rating scale. They were rated five stars 74% of the time, compared with 59% for human reviews, and one star 22% of the time, compared with 10%.[1] Manipulated negative reviews can also be used against competing products.
- Regulators have expanded enforcement, although approaches differ. The FTC's Reviews Rule allows penalties of up to $53,088 per violation and was cited in a $4 million judgment in April 2026;[9] the UK opened its first five fake-review investigations in March 2026.[12]
- The widely cited $152 billion estimate was published in 2021. No newer credible global estimate has been published.[16]
Every quantitative claim on this page is linked to a named, dated source. Figures published before 2025 are identified as historical rather than presented as current. Several frequently repeated statistics can be traced to studies from 2021 or 2024; where no reliable measurement is available, the limitation is stated directly.
How many online reviews are fake or AI-generated?
There is no audited figure covering all platforms. Available estimates measure different samples, time periods and forms of manipulation, so they should not be combined into a single industry-wide percentage.
The most directly relevant measurement for Amazon shoppers comes from Pangram Labs. In May 2026, the company scanned 30,000 front-page reviews across 500 Amazon best-selling products in ten categories and classified 3.0% (909 reviews) as AI-generated with high confidence.[1] The study measured machine-written text only. It did not count paid, incentivised or employee-written reviews created by people.
Share of reviews found to be AI-generated, by platform
Most recent published measurement for each platform. Percentages are not directly comparable — samples, dates and detection thresholds differ.
| Platform | Share AI-generated |
|---|---|
| Capterra (2025) | 33.6% |
| G2 (2025) | 26% |
| Zillow agent reviews (2025) | 23.7% |
| Google reviews (2024) | 19% |
| Luxury hotels, Google Maps | 15% |
| Tripadvisor (2024) | 10.7% |
| Amazon front page (2026) | 3.0% |
| Yelp, filtered as AI (2025) | ≈2.3% |
Platform totals for all fake reviews — not just AI-written ones — run higher, because they include brokered and incentivised human reviews:
| Platform | Period | Removed | Share of submissions |
|---|---|---|---|
| Trustpilot[5] | 2024 | 4.5M | ≈7% |
| Trustpilot[5] | 2025 | 7.8M | not published |
| Tripadvisor[6] | 2024 | 2.7M | 8.7% |
| Yelp[7] | 2025 | ~2.4M | 11% (all moderation removals, not fakes only) |
Amazon stopped publishing a hard number. It reported blocking more than 250 million suspected fake reviews in 2023 and more than 275 million in 2024, but its 2025 report — published April 2026 — says only "hundreds of millions".[8] No 2025 figure, no share-of-submissions denominator, and no independent audit of any platform's claims exists. Every percentage on this page is therefore a measurement of one sample, by one method, at one point in time.
Can anyone actually tell the difference?
A 2025 study by Meng and colleagues ran three experiments in which participants separated authentic Amazon product reviews from reviews generated by large language models. Average accuracy was 50.8% — "essentially the same that would be expected by chance alone". When large language models were given the same classification task, they performed "equivalently bad or even worse than humans".[2]
A separate 2025 study of AI-generated restaurant reviews found accuracy "clustering around chance" across 800 evaluations, and that experienced review readers were no more accurate — only more confident.[14]
Set that against what shoppers believe. In a Bizrate Insights survey of 1,006 US online shoppers fielded in November 2025, 51% said they can identify AI-written reviews, rising to 73% among 18–29 year-olds and falling to 30% among 55–64 year-olds. In the same survey, 88% said they at least occasionally question whether reviews are real, and 91% said a "verified buyer" label would increase their trust.[3]
The findings show a clear gap between confidence and measured performance. Human accuracy was 50.8%, while 51% of survey respondents believed they could identify AI-written reviews. At the same time, 93% of the AI-generated Amazon reviews in one study carried the badge that 91% of shoppers said would increase their trust. The badge therefore offers much less assurance than many shoppers assume.
Automated methods can outperform unaided human reading, but performance varies by task. A 2025 benchmark of 18 detectors over 788,984 review pairs found that a purpose-built context-aware method detected 59.6%–63.5% of AI text at a 0.1% false-positive rate.[15] A separate graph-based detector reached 89–99% precision against LLM-written spam, but 33.7% precision against human-written spam on Yelp data.[15] These results support analysing patterns across an entire review set rather than treating a single sentence as conclusive evidence.
Paste a product URL to review the full rating and review distribution rather than relying on a single suspicious paragraph.
Analyze a product for freeWhich signals still work?
Because sentence-level judgement is unreliable, distribution-level patterns and metadata provide more useful context. Recent research repeatedly identifies the following signals.
Rating shape, not rating average
AI-written reviews were more concentrated at both extremes: 74% were five-star, compared with 59% of human reviews, while 22% were one-star, compared with 10%.[1] A distribution dominated by five- and one-star reviews, with few ratings in between, can indicate polarisation or coordinated activity and warrants closer examination. On G2, four- and five-star reviews were 1.7 times more likely to be classified as AI-generated than lower-rated reviews.[4]
Text that is too clean
An analysis of 714,016 reviews published in the Journal of Retailing and Consumer Services found that AI-generated fake reviews scored higher on comprehensibility and "mechanicalness" and lower on specificity and empathy than authentic reviews. Unlike human-written fakes, they lacked the same psychological markers of deception and instead reflected algorithmic patterns.[13] Meng and colleagues reported a related difference: authentic reviews averaged 1.12 writing mistakes per review, compared with 0.76 for AI-generated reviews, and showed higher sentiment polarity (0.22 versus 0.10).[2] In aggregate, authentic reviews therefore tend to show greater variation in wording, errors and sentiment.
Specificity about this unit
Fabricated reviews often describe a product category in general terms, while first-hand reviews are more likely to mention specific experiences: performance after repeated use, fit, inaccurate indicators or a component failing over time. A persistent lack of concrete usage and failure details across a large review set can be more informative than one unusually worded review.
Timing
Clusters of reviews arriving within a narrow window — especially long after launch, or immediately after a listing changes — indicate coordination. Tripadvisor attributes 54% of the fraudulent submissions it removed to "review boosting", and removed 360,000 reviews tied to employee incentive programmes in a single year.[6]
What does not work
- The Verified Purchase badge. 93% of AI-generated Amazon reviews had one.[1] A refunded purchase produces a verified badge on a bought opinion.
- Perfect grammar as a fake-tell. The relationship runs the other way: AI text has fewer errors than authentic text.[2]
- Reading the top reviews. The front page is exactly where the 3% AI figure was measured — it is the surface with the most incentive to manipulate.[1]
Is it illegal to post a fake review?
As of 2026, major markets including the United States and the United Kingdom prohibit important forms of fake-review creation and distribution. Penalties and documented enforcement actions now accompany these rules.
United States
The FTC's Rule on the Use of Consumer Reviews and Testimonials (16 CFR Part 465) took effect on 21 October 2024. It bans writing, buying, selling or disseminating fake or false consumer and celebrity testimonials — explicitly including AI-generated ones — as well as compensation conditioned on a review's sentiment, undisclosed insider reviews, review suppression through threats, and buying fake indicators of social-media influence.[9] Civil penalties run up to $53,088 per violation; that level was set in January 2025 and carries into 2026, because the annual inflation adjustment was cancelled by OMB memorandum M-26-11 in April 2026.[10]
Recent US enforcement examples include:
- 22 December 2025 — the FTC sent warning letters to ten companies, its first wave under the Reviews Rule.[9]
- 13 April 2026 — the FTC charged the marketers of TruHeight under both the FTC Act and the Reviews Rule over thousands of fake five-star reviews written by employees, free product exchanged for five-star reviews, and bot-run social profiles. Judgment: $4 million, partially suspended, with $750,000 payable. Final order approved 15 July 2026.[11]
- 11 May 2026 — the FTC and the Illinois Attorney General sued a home-services operator over thousands of fake business listings and fabricated five-star reviews used to dilute negative ones.[11]
On the same day it issued those warning letters, the FTC reopened and set aside its 2024 order against Rytr, an AI writing tool whose service had allegedly been used to generate tens of thousands of reviews. The agency said the original complaint had not satisfied the FTC Act's requirements and had placed an undue burden on AI innovation under the administration's AI Action Plan.[11] The decision illustrates a distinction in current US enforcement between businesses that publish deceptive reviews and general-purpose tools that may be used to create them.
United Kingdom
Fake and misleading reviews became a banned commercial practice under the Digital Markets, Competition and Consumers Act 2024, in force from 6 April 2025, with fines up to 10% of global turnover.[12] The CMA published dedicated fake-review guidance in April 2025 requiring publishers to run detection controls, investigate, remove and sanction — "reasonable and proportionate" is the standard.[12] In January 2025 it secured undertakings from Google covering detection, repeat-offender bans and warning labels on offending business profiles. On 27 March 2026 it opened its first five direct enforcement investigations into review practices, naming Autotrader, Feefo, Dignity, Just Eat and Pasta Evangelists.[12] No fake-review fine has yet been imposed; the cases are open.
European Union and elsewhere
The EU has no fake-review-specific Digital Services Act decision to date. Review authenticity has appeared inside broader consumer-protection actions — the CPC network's action against Temu in November 2024 explicitly cited inadequate information about review authenticity, and its May 2025 action against Shein included a request for information on how rankings, reviews and ratings are presented.[12] The proposed Digital Fairness Act was still unpublished as of June 2026, expected in Q4 2026. In Australia, the ACCC collected A$39,600 in infringement penalties in March 2026 from a retailer that failed to disclose paid influencer reviews and edited a negative video review.[12]
What are the platforms doing about it?
| Platform | Year | Enforcement reported |
|---|---|---|
| Google Maps[17] | 2025 | 292M+ policy-violating reviews blocked or removed; 13M+ fake Business Profiles removed; 782,000+ accounts restricted |
| Amazon[8] | 2025 | "Hundreds of millions" of suspected fake reviews blocked; 40+ fake-review brokers ceased activity; 100+ websites shut down via legal action |
| Yelp[7] | 2025 | 11% of ~22M contributed reviews removed; ~500,000 suspected AI-generated reviews filtered; 1.3M+ accounts closed (+138% year on year) |
| Trustpilot[5] | 2025 | 7.8M reviews detected as fake and removed; 100% of submissions screened automatically |
| Tripadvisor[6] | 2024 | 2.7M fraudulent reviews removed (8.7% of submissions); 214,000 AI-generated reviews flagged; ~9,000 businesses warned |
These figures measure detected or removed content, not the total number of fake reviews submitted. Without that denominator, they cannot show what proportion of manipulation remains online. Amazon and Tripadvisor also did not publish directly comparable current-year totals in their latest reporting cycle. Platform enforcement data is therefore useful for understanding the scale of moderation activity, but it cannot establish that a particular listing is free from manipulation.
What do fake reviews cost shoppers?
No current estimate measures worldwide consumer losses with high confidence. Three figures are cited frequently, but each measures a different scope and outcome.
- "$152 billion" — the most-quoted figure on the internet. It comes from a study by CHEQ with the University of Baltimore, published June 2021 and popularised by a World Economic Forum post in August 2021 written by CHEQ's own communications director. It estimates global online spending influenced by fake reviews, not money lost, by multiplying ~$3.8 trillion of e-commerce by an assumed ~4% fake rate. It is five years old and frequently mis-attributed to Cambridge.[16]
- "$300 billion" — from The Transparency Company, December 2024: an estimated $273–334 billion in annual US consumer harm, or $2,385.43 per household, derived from 73 million reviews across home services, legal and medical categories in 100 US cities. US-only, three sectors, not a global or all-category figure.[16]
- "£23 billion" — the UK CMA's estimate of annual UK consumer spending influenced by online reviews, January 2025. Influenced, again — not lost.[12]
No new global estimate has been published in 2025 or 2026. If you see a 2026 article citing a precise worldwide loss figure, it is recycling one of the three above.
What happened to Fakespot?
Fakespot, the best-known consumer review-grading service, was acquired by Mozilla in 2023 and shut down on 1 July 2025. The Firefox Review Checker feature built on it was retired on 10 June 2025.[18] Mozilla said it was refocusing on core browser work.
The closure reduced the number of widely known consumer review-checking tools. It also occurred before the May 2026 measurement of AI-generated reviews on Amazon front pages, during a period when review manipulation was becoming harder to assess through manual reading alone.
How to check a listing in five minutes
Ordered by how much signal each step carries, based on the evidence above:
- Look at the rating histogram before reading individual reviews. A gradual distribution generally provides more context than one concentrated at five and one stars with very few ratings in between. An unusual shape is a reason to investigate, not proof of manipulation.
- Sort by most recent and check the dates. Bunching — dozens of reviews across a few days, particularly months after launch — is coordination.
- Read the three-star reviews. They often contain more balanced descriptions and may surface limitations that are missing from reviews at the extremes.
- Search for specific defect and usage terms — such as "broke", "returned", "smaller" or "after a month". Concrete details can reveal recurring problems that are difficult to see in general praise.
- Check the product identity. If reviews describe a different item, the listing was repurposed and its rating history is borrowed, not earned.
- Cross-check outside the platform. Independent discussion on Reddit, owner forums or specialist communities can provide context that is absent from the marketplace listing.
- Treat the Verified Purchase badge as a limited signal. It confirms that a purchase was recorded through the platform, but does not establish that the opinion was independent. [1]
What Review Detector does — and what it can't
Review Detector automates steps 1 through 6. Paste an Amazon, Best Buy or Target product URL and it analyses the full review set rather than individual reviews — rating distribution, posting velocity, reviewer-account patterns, linguistic markers of machine generation — and returns a Trust Score with the reasoning shown. It cross-references the product against Reddit discussion, surfaces lower-priced alternatives with cleaner review profiles, and requires no signup or browser extension.
| Capability | Why it matters | Review Detector |
|---|---|---|
| Whole-distribution analysis | Single-review detection performs near chance; listing-level analysis does not[15] | Yes |
| AI-generated text detection | 3%+ of front-page Amazon reviews and rising[1] | Yes |
| Review hijacking / listing reuse | Transfers a rating history the current product never earned | Yes |
| Off-platform corroboration (Reddit) | Independent of the incentives acting on the listing itself | Yes |
| Transparent, source-shown scoring | A score you cannot interrogate is a second thing to trust blindly | Yes |
| Alternative product suggestions | Identifying a bad listing is only useful if you can replace it | Yes |
| No account or extension required | No review-history data collected to analyse a public listing | Yes |
No tool, including Review Detector, can certify that an individual review is fake. The 2025 benchmark literature shows that single-review classification remains unreliable at low false-positive rates.[15] A Trust Score is a probabilistic assessment of a review set and can produce errors in both directions: a genuinely well-received launch may resemble coordinated activity, while a gradual manipulation campaign may appear organic. The score should guide further review rather than serve as a final verdict.
Check a listing before you buy — free, no signup, results in seconds.
Analyze a product for freeFrequently asked questions
What percentage of Amazon reviews are fake in 2026?
There is no audited figure for Amazon as a whole. The best independent measurement available is from Pangram Labs, which scanned 30,000 front-page reviews across 500 Amazon best-sellers in May 2026 and identified 3% as AI-generated with high confidence. That covers machine-written text only, not incentivised or brokered reviews written by humans, so it is a floor rather than a total. Platforms that do publish totals report far higher rates: Trustpilot removed 4.5 million fake reviews in 2024, about 7% of everything submitted, and Tripadvisor removed 2.7 million, about 8.7% of submissions.
Are AI-written fake reviews illegal?
In the United States, yes. The FTC's Rule on the Use of Consumer Reviews and Testimonials (16 CFR Part 465) took effect on 21 October 2024 and bans creating, buying, selling or disseminating fake reviews, including reviews generated by artificial intelligence, with civil penalties of up to $53,088 per violation. In the United Kingdom, the DMCC Act 2024 made fake and misleading reviews a banned practice from 6 April 2025, with fines of up to 10% of global turnover. Writing an honest review with AI assistance about a product you actually used is not what these rules target; fabricating experience is.
Can you tell if a review was written by AI?
Not reliably by reading it. In a 2025 study by Meng and colleagues at the University of Nottingham, people asked to separate authentic product reviews from machine-generated ones were right 50.8% of the time — statistically indistinguishable from guessing. Large language models asked to do the same job performed equally badly or worse. Meanwhile 51% of US shoppers surveyed by Bizrate Insights believe they can spot an AI-written review. That gap between confidence and measured ability is the core problem.
Does the Verified Purchase badge mean a review is real?
No. It means an order was placed through the platform, not that the reviewer's opinion is authentic. In Pangram Labs' May 2026 scan of Amazon front-page reviews, 93% of the reviews identified as AI-generated carried the Verified Purchase badge. Brokered review schemes routinely reimburse a genuine purchase in exchange for a five-star review, which produces a verified badge on a bought opinion. Treat the badge as a weak positive signal, never as proof.
What happened to Fakespot?
Fakespot was acquired by Mozilla in 2023 and shut down on 1 July 2025. The related Firefox Review Checker feature was retired on 10 June 2025. Mozilla said it was concentrating resources on its core browser work. Its closure removed the best-known consumer-facing review-grading tool from the market at exactly the point when AI-generated reviews were scaling, which is why comparable checks now have to be run elsewhere.
How can I spot a fake review myself?
Read the distribution rather than individual reviews. Four patterns carry real signal: a rating curve that is bimodal, with a wall of five-star and a spike of one-star reviews and almost nothing in between; clusters of reviews posted within a few days of each other, especially long after launch; reviews that praise the category rather than the specific unit, with no mention of size, fit, failure mode or use over time; and mismatched product identity, where reviews describe a different item because the listing was repurposed. Individual-review reading is the weakest method available, because that is precisely where humans score at chance.
Do platforms actually remove fake reviews?
At very large scale, yes, but with an unknown residual. In 2025 Google blocked or removed more than 292 million policy-violating reviews on Maps and took down over 13 million fake Business Profiles. Amazon reported blocking hundreds of millions of suspected fake reviews in 2025, after 275 million in 2024 and 250 million in 2023. Yelp closed more than 1.3 million user accounts in 2025 and filtered nearly 500,000 suspected AI-generated reviews. What none of these figures tell you is the removal rate — how many fake reviews were submitted in total — so they measure enforcement effort, not the share of fakes that survive.
How much money do fake reviews cost consumers?
The honest answer is that nobody has a current global figure. The widely quoted $152 billion number comes from a June 2021 study by CHEQ with the University of Baltimore and is five years old; it estimates spending influenced by fake reviews, not money lost. The most recent substantive estimate is from The Transparency Company in December 2024: roughly $300 billion a year in US consumer harm, or $2,385 per household, but only across home services, legal and medical categories. The UK's competition regulator estimates that £23 billion of annual UK consumer spending is influenced by online reviews. Any 2026 article citing a precise global loss figure is recycling one of these.
Sources
Every figure above maps to one of these. Publication dates are given so you can judge freshness yourself; figures from before 2025 are labelled as such in the text.
- Pangram Labs, "Three percent of front-page Amazon reviews are now AI-generated". 4 May 2026. 30,000 reviews across 500 best-sellers. Vendor of AI-detection software.
- Meng, Harvey, Goulding, Carter, Lukinova, Smith, Frobisher, Forrest & Nica-Avram, "Large Language Models as 'Hidden Persuaders': Fake Product Reviews are Indistinguishable to Humans and Machines", arXiv:2506.13313. 16 June 2025. Preprint.
- Bizrate Insights, "How shoppers navigate AI and authenticity". 13 May 2026 (fielded November 2025, n=1,006 US shoppers).
- Originality.ai studies of Google reviews (11 Oct 2025), Zillow (14 Oct 2025), Tripadvisor (11 Dec 2025), luxury hotels (26 Jan 2026) and G2 / Capterra / TrustRadius (17 Oct & 19 Nov 2025). Vendor of AI-detection software.
- Trustpilot, Trust Report 2025 (29 May 2025, covering 2024) and Trust Centre for the 2025 figure. No full 2026 report published as of August 2026.
- Tripadvisor, 2025 Transparency Report. 18 March 2025, covering calendar 2024. No 2026 edition published as of August 2026.
- Yelp, 2025 Trust & Safety Report. 25 February 2026.
- Amazon, 2025 Trustworthy Shopping Experience Report (22 April 2026) and "Amazon's latest actions against fake review brokers" for the 2023 and 2024 figures.
- US Federal Trade Commission, "The Consumer Reviews and Testimonials Rule: Questions and Answers" (8 Nov 2024, updated 1 May 2025); warning letters to ten companies (22 Dec 2025).
- FTC, inflation-adjusted civil penalty amounts for 2025 (11 Feb 2025); OMB Memorandum M-26-11 cancelling the 2026 adjustment (17 April 2026).
- FTC, action against TruHeight (13 April 2026); FTC and Illinois v. Premium Home Service (11 May 2026); Rytr order set aside (22 Dec 2025).
- UK CMA, Fake reviews guidance (CMA208) (4 April 2025); first five investigations (27 March 2026); Google undertakings and the £23bn estimate (24 January 2025); European Commission CPC actions on Temu (8 Nov 2024) and Shein (26 May 2025); ACCC, PhotobookShop penalties (24 March 2026).
- Zhao, Tang, Zhang & Lyu, "AI vs. human: A large-scale analysis of fake reviews", Journal of Retailing and Consumer Services, vol. 87. 2025. 714,016 reviews analysed.
- Tuomi, Zainal Abidin, Tuominen & Ascenção, "Convincingness of AI-Generated Restaurant Reviews", Springer. 27 May 2025.
- Yu, Luo, Madasu, Lal & Howard, detector benchmark over 788,984 review pairs, arXiv:2502.19614; Liu et al., "FraudSquad", arXiv:2510.01801. 2025.
- World Economic Forum / CHEQ & University of Baltimore, "Fake online reviews are a $152 billion problem" (10 August 2021 — five years old); The Transparency Company, "The High Cost of Review Fraud" (5 December 2024).
- Google, "New ways we're protecting businesses on Maps". 16 April 2026.
- Mozilla, "Building what's next". 2025. Fakespot shutdown 1 July 2025; Firefox Review Checker retired 10 June 2025.
If you are quoting these figures, please link the original studies above. To cite this compilation:
Review Detector Operation Team. "Trust in Reviews: How Many Online Reviews Are Fake or AI-Generated in 2026?" Review Detector, 18 August 2026. https://reviewdetector.ai/en/legal/trust_in_reviews
Figures on this page were collected in August 2026 by reading each primary source directly rather than secondary summaries. Where a widely circulated statistic could not be traced to a primary source with a stated method — including the frequently quoted claim that "30% of online reviews are fake" — it was excluded rather than repeated. Vendor-produced measurements are labelled as such, because companies selling detection software have a commercial interest in the prevalence they report. Figures published before 2025 are dated in the text so that their age is visible. This page is reviewed quarterly; the "last updated" date changes only when a figure or conclusion changes.