How to Spot Fake Reviews: The 8 Signals That Actually Work
"How do I know if these reviews are real?" is one of the oldest questions in online shopping, and it's gotten more urgent, not less. Amazon says it blocked more than 275 million suspected fake reviews in 2024 alone, and a 2026 BrightLocal survey found 49% of consumers now trust online reviews as much as a personal recommendation from a friend — meaning the stakes of a manipulated rating are higher than ever. This page lays out the actual methodology behind a working fake-review detector: not vague advice like "watch out for reviews that sound too positive," but eight concrete, threshold-based signals you can apply to any product listing.
Why a checklist beats a gut feeling
Most advice about spotting fake reviews is qualitative — "look for generic language," "be suspicious of a flood of 5-star reviews." That's not wrong, but it's not actionable either. A real detection system converts those instincts into measurable thresholds applied against the actual review data: how many reviews arrived in a short window, what fraction are short and generic, how many pairs of reviews are near-duplicates of each other, and so on. Below is the real 8-signal table our own review-analysis engine runs — the same class of deterministic checks that Fakespot itself used before its July 2025 shutdown, made transparent instead of hidden inside a black-box score.
The 8 signals, with real thresholds
| Signal | Threshold | What it catches |
|---|---|---|
| Date clustering | More than 25% of reviews arrive within a single one-week window | Reviews bunched into a narrow burst — a strong indicator of a coordinated review campaign rather than organic purchases spread over time |
| Short reviews | More than 40% of reviews are under 15 characters | Generic, low-effort text like "Great!" or "5 stars" posted in bulk, often by paid reviewers who never actually describe the product |
| Near-duplicates | More than 5% of review pairs have Jaccard text similarity above 0.8 | Reviews that are near-identical to each other, suggesting the same template was reused across many "different" reviewers |
| Rating anomaly | More than 70% five-star combined with less than 5% one- or two-star | A rating distribution too clean to be organic — real products almost always accumulate some negative reviews over time |
| Verified purchase rate | Less than 30% of reviews are marked as verified purchases | A low verified rate means a large share of reviewers may never have actually bought the item, which is a common pattern in incentivized or fake review schemes |
| Photo/video presence | Informational, not a red flag on its own | Reviews with attached photos or video tend to correlate with more trustworthy, effort-invested feedback — its absence isn't proof of anything by itself |
| Seller response rate | More than 80% of reviews receive a seller response | A very high response rate, especially with templated wording, can indicate mass automated seller replies rather than genuine customer service engagement |
| Average useful-vote count | Informational, not a red flag on its own | Low community engagement (few people marking a review "helpful") can be a soft signal of astroturfing, but isn't conclusive alone |
Two of these thresholds get refined further in practice. A rating-anomaly check that fires just because a sample of reviews looks skewed can be a false alarm if the full population of reviews is also skewed the same way — so a working detector suppresses the anomaly signal when the sample and population five-star percentages are within about 5 percentage points of each other. Similarly, date clustering gets disabled entirely when the review sample was pulled in "most recent first" order, since a recency-sorted sample will always look clustered even on a perfectly normal product.
Why deterministic checks come before AI
You might expect a "fake review detector" to just throw everything at a language model and ask "does this look fake?" That approach is fast, but it's also unreliable and expensive to run at scale, and it gives you a confidence-sounding answer with no way to verify the reasoning. The stronger design runs the deterministic checks above first — pure math, zero AI, completely reproducible — and only then has a language model read the surviving review text to summarize genuine pros, cons, and an overall verdict reason. This two-stage approach means the headline verdict is always traceable back to specific, checkable numbers, not just a model's impression.
In our own pipeline, after the deterministic pass identifies which signals fired, a language model (Claude Sonnet, running at a low, deterministic temperature setting) reads the reviews that weren't already flagged as unreliable and produces the human-readable summary — trust score, verdict, and the two or three most useful pros and cons. The model doesn't get to override what the deterministic math found; it just explains and summarizes on top of it.
The two shutdowns that made this checklist scarce
If this methodology feels unusually detailed compared to what you're used to seeing, that's not an accident — it's partly because the two most popular free consumer tools that used to run checks like these both disappeared in 2025. Firefox's built-in "Review Checker" was discontinued on June 10, 2025, and Fakespot's full suite of extensions, apps, and its website shut down on July 1, 2025, with Mozilla citing no sustainable business model rather than any flaw in the underlying detection approach. A separate, older tool called ReviewMeta had also gone quiet before that. The result is that millions of shoppers who used to lean on a one-click review check now have no default tool to reach for — which is exactly why writing out the actual methodology, rather than just selling a black-box score, is worth doing.
What this checklist can't do
No detection system — ours or anyone else's — catches everything. A well-funded, patient review-manipulation operation can stagger fake reviews over months to avoid date clustering, write longer and more varied fake text to avoid the duplicate-text and short-review signals, and even solicit real (if unrepresentative) purchases to inflate the verified-purchase rate. Treat a clean result from any of these 8 signals as a positive sign, not an ironclad guarantee. The value of a structured checklist isn't perfection — it's catching the large share of low-effort, high-volume fake-review campaigns that don't bother covering every angle, which based on Amazon's own 275-million-blocked-reviews figure for 2024, is still the overwhelming majority of what's out there.
Running these signals yourself, today
Six of the eight signals above need marketplace-side metadata — verified-purchase flags, timestamps, seller-response records — that only exist if you're pulling data directly from a live product page. But two of them, short-review detection and near-duplicate text matching, are pure text signals that work on any block of review text you have in front of you, from any marketplace, including Amazon.
You can run those two checks right now by pasting review text into @vustReviewBot — the result screen will tell you plainly which of the 8 signals it was able to evaluate and which ones it couldn't, rather than presenting a score that quietly assumes data it doesn't have. Direct link-based analysis for Amazon and other global marketplaces — which is what unlocks full 8-signal coverage — is in development, tracked through waitlist demand rather than promised on a fixed date.
Either way, understanding the actual 8 signals — not just trusting a single opaque number — puts you in a stronger position the next time you're staring at a wall of reviews trying to decide whether they're real.
Reading the signals together, not in isolation
A single signal firing on its own is rarely conclusive — the real value of this checklist comes from reading multiple signals together. A product with a high short-review percentage but a normal, gradually-accumulated rating distribution and a healthy verified-purchase rate is probably just a product whose genuine buyers don't write much — plenty of real shoppers leave one-line reviews. The same short-review pattern combined with heavy date clustering, a suspiciously clean rating distribution, and a low verified-purchase rate is a much stronger case for manipulation, because now several independent signals are all pointing the same direction at once rather than one metric behaving oddly for an innocuous reason. This is exactly why a working detector doesn't stop at the first signal that fires — it evaluates all 8 and lets a language model weigh how they combine before writing a verdict, rather than triggering a "fake" label off a single threshold crossing.
Signals that mean something different depending on category
Context matters more than the raw numbers suggest. A handmade or niche product with only 40 total reviews will naturally show a higher percentage of five-star ratings than a mass-market item with 40,000 reviews, simply because a small, self-selected group of early buyers tends to skew positive — that's not automatically manipulation. Similarly, a product launched during a single flash sale will legitimately show heavy date clustering in its review timestamps, because that's when the purchases actually happened, not because reviews were coordinated. A responsible reading of these 8 signals treats them as flags worth investigating rather than automatic disqualifiers, and a well-built detector accounts for exactly this — suppressing the rating-anomaly signal when a sample's skew matches the broader population, and disabling the date-clustering check entirely when the review sample itself was pulled in a way (most-recent-first) that would make any product look artificially clustered.
A checklist you can apply without any tool at all
Even without running anything through an automated pipeline, you can apply a lightweight version of this checklist manually the next time you're evaluating a listing: scan the most recent 20-30 reviews for how many are under a sentence long, glance at whether the dates are spread out over months or bunched into a single week, skim for two or three reviews that read almost identically to each other, and check what fraction show a "verified purchase" badge. None of that takes more than a minute or two, and it captures a meaningful share of what the full 8-signal system checks automatically — the automated version just does it faster, more consistently, and across all 8 signals at once instead of the 3 or 4 you can realistically eyeball yourself.