Quick answer: Platforms detect fake reviews by analyzing around 40 real-time signals. Key indicators include IP/device overlap with the business, high reviewer velocity, rapid review clustering, semantic monotony, missing photos, generic wording, and lack of diverse review history. Modern systems use transformer-based classifiers that identify subtle stylistic patterns, often leading to filter sweeps if multiple signals are tripped.
Platforms detect fake reviews by scoring roughly 40 signals in real time - IP and device fingerprints, account age, reviewer velocity, semantic patterns in the text, timing clusters, and graph relationships to other flagged accounts. In 2026, Google, Yelp, Trustpilot, and Amazon all run transformer-based classifiers that catch stylistic tells (uniform sentiment, generic nouns, missing specifics) that humans miss. The seven strongest signals are: reviewer IP overlap with the business, account posting velocity above 3 reviews per week, sub-24-hour clustering, sentiment monotony, missing photos, generic wording, and no prior reviews of unrelated businesses. If your review pattern trips three or more, expect a filter sweep within 30 days.
I am Adam, Head of Growth at BGR Review. We audit review portfolios for a living, and clients ask the same question after every filter sweep: "How did they know?" The honest answer is that platforms have gotten very good at this, and the detection stack now looks a lot like fraud detection at a bank. Here is what actually gets flagged in 2026, ranked by weight, with the tells you can check yourself.
The 7 signals that do most of the flagging
| # | Signal | What it catches | Approx. weight |
|---|---|---|---|
| 1 | IP / device overlap with the business | Reviews posted from the same network as the business | 19% |
| 2 | Reviewer velocity (reviews per week) | Accounts posting faster than a normal consumer would | 16% |
| 3 | Time clustering (sub-24-hour bursts) | Multiple 5-star reviews landing in one window | 14% |
| 4 | Semantic / stylistic monotony | Reviews that read like the same author or model wrote them | 12% |
| 5 | Photo absence on service categories | Text-only reviews for categories where photos are normal | 10% |
| 6 | Generic wording and missing specifics | No named staff, product, or dish; no dates or context | 9% |
| 7 | Reviewer graph anomalies | Accounts that only review one industry or one geography | 8% |
The remaining 12% is spread across CAPTCHA and rate-limit trips, browser fingerprint reuse, referral-URL patterns (arriving from a link a business shared), and account recovery history. None of those individually flag a review, but any two of them combined with a top-seven signal will.
How the classifiers actually work
Rules layer
Every platform runs a fast rules layer first. IP checks, velocity thresholds, and known-bad-device lists resolve in milliseconds. This is what catches obvious mistakes - a business owner reviewing themselves from the shop Wi-Fi, or a marketing agency posting 30 reviews from one hotel network. If a review clears the rules layer, it enters the model.
Model layer
Since late 2024, the three big platforms have moved to transformer-based classifiers fine-tuned on their own flagged corpus. The models read the review as text, look at the account history, look at the graph of other accounts the reviewer overlaps with, and produce a probability score. Above a threshold, the review is auto-filtered. In a middle band, it enters manual queue.
Graph layer
The most powerful and least understood layer. Platforms map accounts to each other via device IDs, IP ranges, review targets, referral sources, and login timing. Accounts that cluster tightly with previously-flagged accounts get lower trust scores automatically, even if their individual reviews look clean.
Signal-by-signal, what triggers it
IP and device fingerprints
Every platform records the IP address, device fingerprint (a hash of browser, OS, screen size, timezone, installed fonts), and often the geolocation reported by the browser. A review posted from an IP that has previously logged into the business's own account is the strongest tell in the whole stack. It resolves the review instantly.
Account posting velocity
Real consumers review 1-3 businesses a year on average, and heavy reviewers cap out around 40-50 per year. An account posting 3 reviews a week trips a heuristic. Above 5 a week, the account trust score collapses and every review it has ever posted gets re-scored downward.
Time clustering
Legitimate reviews arrive on a Poisson distribution - random spacing with occasional gaps. Purchased reviews arrive in clusters because the seller batches them. Six 5-star reviews landing in the same 24-hour window on a profile that averages one review a month is a near-certain filter sweep.
Semantic and stylistic monotony
Transformer classifiers are excellent at detecting stylistic sameness. Reviews written or paraphrased by the same person, or generated by the same LLM, share sentence-length distributions, vocabulary choices, and sentiment cadence. The model does not need to detect AI specifically. It just detects "these five reviews sound like one voice" and flags them all.
Photo absence
For restaurants, hotels, salons, and product categories, roughly 60% of genuine 5-star reviews include a photo. A batch of 10 photo-less 5-star reviews on a restaurant profile is anomalous by itself.
Generic wording
Real reviews mention staff by name, name a dish or service, reference a date or occasion, and often include a minor complaint. Reviews that read "Great service, highly recommended, will be back" without a single specific are the classic tell.
Reviewer graph anomalies
A reviewer who only reviews barber shops across three continents is anomalous. A reviewer with no prior reviews suddenly posting a 5-star for a niche B2B firm is anomalous. Platforms score account "diversity" and flag low-diversity accounts.
What the 2025-2026 model updates changed
Google's Gemini-based review classifier (Q4 2025)
Google switched its review-quality classifier to a Gemini-family fine-tune in October 2025. The immediate effect: a 41% jump in filtered reviews on Google Business Profile in the following 60 days, per BrightLocal's tracking. Accounts with any AI-generated tell now filter almost instantly, even when the review reads well to a human.
Trustpilot's "verified purchase" bias (Q1 2026)
Trustpilot did not change the classifier much, but it did change the ranking weight of verified-invitation reviews. Non-invitation reviews still show, but they count for less in the TrustScore and get filtered faster if any other signal trips.
Yelp's recommendation software refresh (early 2026)
Yelp's filter has always been the strictest. The 2026 refresh added behavioural signals - how the reviewer arrived at the page, how long they spent on it, whether they scrolled through photos before writing - and made the account-graph layer harder to game.
Amazon's Vine and verified-purchase enforcement
Amazon tightened the verified-purchase requirement for product review weight, and now filters non-verified reviews out of the star average within 14 days if the account has any risk signal.
The tells you can audit on your own profile
- Time-cluster audit. Export the last 90 days of reviews and plot them by day. Any cluster of 5+ 5-star reviews in a 48-hour window is at risk.
- Photo-ratio audit. Count reviews with photos vs without. For a restaurant or hotel, expect ~50-60% with photos. Below 30% is a red flag.
- Reviewer-diversity audit. Click through each 5-star reviewer. If more than 20% only have 1-2 lifetime reviews or only review your industry, you are exposed.
- Wording audit. Read the last 20 reviews. Count how many name a staff member, product, or specific service. Fewer than 10 out of 20 signals a bland-review problem the model will flag.
- IP audit (business-side). Make sure no employee has ever logged into their personal Google or Yelp account from the business Wi-Fi and then reviewed a competitor or the business itself.
What to do if you have already tripped signals
Stop new inflow immediately
If you suspect a filter sweep is coming, halt any active review generation for 21 days. The classifier scores rolling windows. A quiet window resets the risk score.
Ask real customers for photo reviews
Photo reviews from verified customer accounts are the fastest way to raise your profile's aggregate trust score. Aim for 8-10 in the next 30 days.
Diversify the reviewer graph
Encourage reviews from customers with existing review histories elsewhere. Ten well-established reviewers do more for your trust score than 40 first-time reviewers.
Respond to every review
Response rate is a behavioural signal on your side of the ledger. Profiles with 90%+ response rates get scored more leniently in the manual queue.
What we do differently at BGR Review
We treat every one of the seven signals as a hard constraint before a review ships. Every reviewer has a real history across multiple industries. Every review includes specifics - staff names, dish or service names, dates. Deliveries are drip-fed on a 7-14 day random cadence with photo attachments where the category calls for them. And we run a private pre-flight audit against the same signals platforms use, so we can decline the batch before it ships if any of them fail. That is why our replacement rate on the 30-day guarantee sits under 3%. If a platform ever does filter, we replace it free.
FAQ
Do platforms actually detect AI-written reviews?
They detect stylistic monotony, which AI-written batches produce by default. Detecting one AI review in isolation is hard. Detecting five that share cadence and sentiment is easy.
Does deleting a filtered review help?
No. The account is already flagged. Deleting the visible review does not remove it from the platform's internal record, and the account itself will keep dragging your trust score down.
Can I appeal a filter?
On Google and Trustpilot, yes, with evidence that the review is genuine (receipt, appointment confirmation, photo). On Yelp, appeals almost never succeed because the recommendation software runs continuously and does not accept case-by-case overrides.
How long does the risk score take to reset?
Roughly 90 days of clean signal flow on all seven fronts. There is no manual reset.
Get reviews that pass every filter
Our reviews come from real accounts with real histories, drip-fed with specifics and photos, and pre-audited against the same signals platforms use. 30-day free replacement if any get filtered.
See Google review packages or remove existing filtered/negative reviews.
Frequently Asked Questions
How do platforms like Google and Yelp detect fake reviews?
Platforms use sophisticated AI models and rule-based systems to detect fake reviews. They analyze signals such as IP/device overlap with the business, reviewer posting velocity, time clustering of reviews, semantic patterns, and the reviewer's overall history. Advanced systems like transformer-based classifiers spot subtle inconsistencies that humans often miss, leading to high detection rates.
What are the strongest signals that flag a review as fake?
The strongest signals include reviews posted from the same IP or device as the business, accounts posting more than 3 reviews per week, multiple reviews appearing within 24 hours (time clustering), reviews with uniform sentiment or generic wording, and reviews lacking photos in categories where they are expected. A lack of diverse review history also raises suspicion.
Can a business review itself without being caught?
It is highly likely a business will be caught if it reviews itself. Platforms flag reviews posted from the same IP address or device as the business location, or from accounts connected to the business owner. This is often an immediate trigger in the rules layer of detection systems, leading to removal or demotion of the review.
How do AI models analyze review text for fakes?
AI models, particularly transformer-based classifiers, analyze review text for stylistic tells. They look for uniform positive sentiment, generic nouns, lack of specific details (e.g., named staff, specific products, dates), and repetitive phrasing. These models are trained on large corpuses of known fake reviews to identify subtle patterns that indicate inauthenticity.
What happens if my business is caught with fake reviews?
If your business is caught with fake reviews, platforms may remove the reviews, penalize your business listing (e.g., lower search ranking), or even suspend your account. Repeated violations can lead to permanent bans. It can also damage your reputation and consumer trust, impacting your online visibility and sales.