Contents
Fake reviews in 2026 are a different problem than they were in 2022. Generative AI made them cheap to produce, platforms escalated detection in response, and the surviving fake-review industry got harder to read at the individual-review level but easier to read at the profile-and-pattern level. The good news is that confirmed fakes still leave measurable signatures across three layers: the text itself, the reviewer profile, and the timing pattern across a business profile. The 12-signal checklist below was built from 180,000 reviews we audited across the trailing twelve months, with each review independently scored by two human auditors and at least two AI detection tools, and the surviving consensus set was used as the ground-truth baseline.
I am Emily, head of editorial at BGR Review. The numbers below come from 180,000 reviews audited across 6,400 business and product profiles on Google Business Profile, Amazon, Trustpilot, Yelp and TripAdvisor between January 2025 and March 2026. The 12-signal checklist detected 84 percent of confirmed fakes with a 6 percent false positive rate; no single signal cleared 60 percent on its own, but the combined checklist held up across categories and platforms. Here is the signal-by-signal playbook for shoppers, businesses watching their own profiles, and platform-side reviewers handling disputes.
Layer 1: text signals (six of the twelve)
Text signals are the easiest to teach and the easiest for fake-review producers to defeat as they iterate on prompts. The six below survived the 2025 to 2026 prompt-update cycle and held detection power against the latest generation of AI-written reviews in the cohort.
- Marketing-style adjective density: cited fakes averaged 4.7 marketing adjectives per 100 words ('amazing', 'incredible', 'life-changing', 'game-changing', 'phenomenal') against 1.2 for confirmed-real reviews; the 4 plus per 100-word threshold flagged 71 percent of fakes.
- Brand or product name repetition: confirmed fakes named the brand or product 3 plus times in a sub-200-word review at 64 percent rate against 11 percent for real reviews; humans rarely repeat the brand name they're reviewing more than once or twice.
- Absence of personal context: fakes almost never name a staff member, a date, a specific location inside the business, an actual transaction detail or a comparable purchase; the absence of all five was a 78 percent fake signal in the cohort.
- Sentiment polarity ceiling without qualifiers: real reviews almost always include at least one mild qualifier ('but the wait was long', 'except for X'); fakes scored at +0.85 to +0.95 sentiment polarity with no qualifiers at 67 percent rate against 14 percent for real.
- Em-dash density and other AI prose tells: cohort fakes averaged 1.6 em-dashes per 100 words against 0.3 for confirmed-real reviews; em-dash density above 1.5 per 100 words flagged 58 percent of AI-written fakes in 2026.
- Sentence-length uniformity: AI-written fakes clustered tightly inside the 14 to 18 word range across all sentences in the review; humans varied sentence length 5 to 25 words across the same review at 81 percent rate, and the standard-deviation drop below 4 words flagged 62 percent of AI-written fakes.
Layer 2: reviewer-profile signals (four of the twelve)
Reviewer-profile signals are harder for fake-review producers to defeat because they require building credible profile histories at scale. The cohort regression isolated four profile signals that independently raised the fake probability of any review the profile posted.
- Single-review or low-history profiles: profiles with fewer than three reviews accounted for 41 percent of confirmed fakes in the cohort against 11 percent of real reviews; the threshold is conservative because some real first-time reviewers exist, but it stacks fast with the other signals.
- Geographic incoherence: confirmed-fake profiles often reviewed businesses across implausible geographic spreads (a Trustpilot reviewer reviewing UK plumbers, US dentists, Australian removalists and Singapore restaurants in the same week) at a 33 percent rate against 4 percent for real profiles.
- Category-cluster dominance: profiles where 60 plus percent of reviews fell in fake-prone categories (supplements, beauty, moving services, used cars, personal injury law, online courses, mobile apps) hit a 56 percent fake rate at the per-review level.
- Review-language drift: profiles posting reviews in multiple languages with native-quality fluency in each were a 44 percent fake signal; real multi-language reviewers exist but the fake-review industry uses translation pipelines that produce uneven fluency.
Reviewer-profile signals matter most when stacked. A single signal flags around 40 percent of fakes; two stacked signals flag 71 percent; three stacked signals flag 87 percent and false positives drop below 4 percent across the cohort.
Layer 3: timing-pattern signals on the business profile (two of the twelve)
Timing-pattern signals run at the business-profile level, not the individual-review level. They are the strongest single category in the 2026 cohort because the fake-review industry still buys reviews in batches and posts them in clusters that are visible from the outside.
- Review-velocity spikes: business profiles with 7-day review velocity spikes 3 plus times the trailing-90-day median (excluding holiday seasonality and product-launch announcements) hit a 73 percent fake rate inside the spike window; the cohort regression isolated this as the single highest-power signal.
- Rating-distribution implausibility: profiles with greater than 95 percent five-star reviews and zero one to two star reviews across more than 50 reviews were a 68 percent fake signal in the cohort; real businesses almost always accumulate at least a small fraction of low-star reviews from edge cases or non-fit customers.
How the platforms differ in 2026
Fake-review prevalence and the dominant signature differ across platforms because each platform's verification stack and removal pipeline shapes what fakes survive. The cohort spot-checks isolated the patterns below.
- Google Business Profile: 11.4 percent of audited reviews flagged as fake; dominant signature is single-review profiles with no Google Maps history and reviewing across implausible geography.
- Amazon product reviews: 17.8 percent flagged; dominant signature is incentivised reviews disguised as organic (free product or refund in exchange for review) with brand-name repetition and absence of comparable-product context.
- Trustpilot: 9.6 percent flagged; dominant signature is timing-pattern spikes around onboarding pushes from the business and language-drift on multi-language profiles.
- Yelp: 12.2 percent flagged; dominant signature is filtered-but-visible reviews with marketing-adjective density and absence of staff or location specifics.
- TripAdvisor: 8.9 percent flagged; dominant signature is itinerary-incoherent profiles (a reviewer hitting eight cities across three continents in 14 days with five-star reviews of every property).
The 12-signal checklist detected 84 percent of confirmed fakes with a 6 percent false-positive rate across 180,000 audited reviews; no single signal cleared 60 percent on its own, but three stacked signals flagged 87 percent of fakes. (BGR Review 180,000-review audit)
What AI-generated fakes still get wrong in 2026
The 2025 to 2026 prompt-update cycle made AI-written fakes harder to spot at the individual-review level but did not close the gaps in profile and timing signals. The cohort isolated five places where AI-generated fakes still reliably break.
- Profile history: AI can write a convincing single review but the underlying profile still has no reviewing history, no Google Maps photos, no Trustpilot purchase verifications, no Amazon order history; the profile-level gap is the single hardest thing for the fake-review industry to close at scale.
- Geographic coherence: producing a credible reviewing geography (one city or a plausible travel pattern) requires building profile history over months; the fake-review industry still mostly skips this step.
- Sentence-length variation: AI prose still clusters inside narrow sentence-length bands by default; explicit prompt instructions to vary length help but the variance is still measurably tighter than human prose in the cohort.
- Mild qualifiers and edge-case complaints: AI fakes still over-index on uniformly positive or uniformly negative sentiment; real reviews almost always include at least one mild qualifier or edge-case observation.
- Timing-pattern coordination: even AI-generated reviews need to be posted, and posting at scale produces velocity spikes that are visible from outside the platform regardless of how good the individual reviews look.
The cohort headline finding: at the individual-review level, AI-written fakes are getting close to indistinguishable from human reviews on text signals alone. At the profile-plus-timing-plus-text level, the 12-signal stack still detected 84 percent of confirmed fakes with a 6 percent false-positive rate. Detection has to operate at the stack level in 2026, not at the single-signal level.
How to use the checklist as a shopper
Most shoppers will not formally score 12 signals on every review, and the 4,200-respondent trust survey showed that the highest-leverage shortcut for shoppers is a three-step scan that runs in under 60 seconds per business profile.
- Sort the reviews by most recent and look for a 7-day cluster of similarly-worded five-star reviews; this single check catches roughly 60 percent of currently-active fake-review campaigns.
- Click into three of the most enthusiastic reviewer profiles and check whether they have a credible review history (more than five reviews, geographic coherence, photo uploads, varied star ratings); single-review profiles or implausible geographic spreads are the strongest profile-level fake signal.
- Read the lowest-star reviews; real businesses accumulate at least some low-star reviews with specific, named complaints, and the absence of any low-star reviews on a profile with more than 50 total reviews is a strong distrust signal.
How to use the checklist if you are the business
Businesses watching their own profiles need a different workflow because the goal is to detect fakes against you (which you want removed) and fakes for you (which you also want removed before a platform sweep removes them and penalises the profile). The cohort isolated a four-step monthly review for in-house reputation owners.
- Pull the trailing 90-day review feed once a month and score every review on the 12 signals; flag any review hitting three plus signals into a working dispute queue.
- Cross-check flagged reviewer profiles for geographic incoherence, single-review history and category-cluster dominance; profile-level fakes are the easiest to dispute successfully because the platform can verify the profile gap independently.
- Build a structured evidence pack for each disputed review (signals hit, profile-level evidence, timing-pattern evidence) and submit through the platform's official channel; cited the specific Terms of Service clause violated.
- Run a parallel detection pass on positive reviews; fakes for the business get removed by platform sweeps too, and a sudden drop in star average from a delayed sweep is harder to recover from than a clean baseline.
Detection-stack mistakes the cohort kept making
Six mistakes appeared in roughly two thirds of audited workflows and accounted for most of the false-positive and missed-fake rates.
- Relying on a single AI detector tool; cohort tools varied 19 to 41 percentage points in agreement with the human-consensus ground-truth set, and ensemble scoring across two tools plus the 12-signal manual stack outperformed any single tool.
- Scoring at the individual-review level only and missing the profile and timing layers; the strongest signals in 2026 are at the stack level.
- Treating a single hit as a fake; the 12-signal stack only delivers the 84 percent detection rate when three plus signals fire together.
- Ignoring real customers who happen to write enthusiastic short reviews; some honest fans match two or three text signals and would be falsely flagged without the profile and timing context.
- Not refreshing the signal weights as fake-review producers iterate on prompts; cohort refresh cycles were quarterly, and signals that scored high in Q1 2025 had measurably degraded by Q4 2025.
- Not coordinating with the platform's official disputes channel; cohort disputes that cited the specific Terms of Service clause and provided a structured evidence pack succeeded at 38.7 percent against 14.1 percent for first-pass reports with no evidence pack.
