The review was five stars, detailed, and glowing — and it praised the waterproofing on a tent that is not waterproof, in sentences that somehow also described a coffee maker. Product reviews now live in a strange era: some are written by customers, some by marketers, some by the seller’s cousin, and a growing share are generated by AI tools that have never touched the product. No single test exposes them all, and anyone promising you a foolproof detector is selling you something. What does work is a checklist of signs that, in combination, separate most machine-written reviews from the real thing.
The checklist: eight signs to run through
1. It praises features the product doesn’t have
The strongest tell of all. Generated reviews are often written from the product category rather than the product, so they confidently mention a timer the appliance lacks, a pocket the jacket doesn’t have, a battery life from a different model. A human reviewer can be wrong about a spec; a review that reads like a brochure for a similar-but-different product is usually a sign nobody involved ever held this one.
2. It is full of enthusiasm and empty of incidents
Real reviews contain events: “the zipper jammed at the airport,” “assembly took me two evenings because the diagram is upside down.” Generated text deals in adjectives instead of incidents — amazing, incredible, game-changer, exceeded all expectations — because the software has no evening with the diagram to report. Ask of any review: did anything happen in it? If the answer is no, be suspicious.
3. The complaints are suspiciously balanced and weightless
AI asked to sound “balanced” adds a token flaw of no consequence: “the only downside is I didn’t buy it sooner.” Genuine mixed reviews name trade-offs that cost something — it’s heavy, the app is annoying, size up if you’re between sizes. A drawback that is secretly a compliment is a marketing voice, whether a person or a program produced it.
4. The same phrases repeat across reviews
Scroll the full review list, not just the top ones. Generated batches share bones: the same sentence openings, the same structure of praise, the same odd word choices appearing under different names. One “delve into” is nothing; fourteen reviews that all “exceed expectations in every way” are a production line clearing its throat.
5. A burst of reviews arrived at once
Check the dates. Real products accumulate reviews in a messy trickle tied to sales. A product that collected forty reviews in three days after months of silence — especially generic ones — has usually been seeded. Platforms’ own rules treat solicited fake engagement as fraud; Google’s review policies on fake engagement spell out what counts as fake on their platform, and the burst pattern is part of what enforcement looks for.
6. The reviewer’s history is one product deep
Where profiles are visible, look. An account created to review this single item, with no history of mundane opinions about restaurants and shoes, tells its own story. Real reviewers are boringly consistent across years of purchases; generated campaigns cannot afford that much patience. The reverse also holds: a reviewer whose history shows picky, specific, sometimes grumpy opinions across dozens of unrelated products is carrying years of credibility that no campaign would bother to build, and their four stars are worth more than a stranger’s five.
7. Perfect grammar wrapped around zero specifics
Price paid, room size, model number, how it compares to the thing it replaced, what broke and when — human reviewers leak specifics without trying. Generated reviews stay at cruising altitude where specifics can’t be checked. Ironically, the small typo or odd detail (“fits my weird narrow hallway”) is evidence of life that flawless prose is not.
8. It never mentions a single competitor or alternative
People who actually shop say “narrower than the Brand X I returned.” Machines writing praise in a vacuum rarely risk the comparison, because the prompt was about this product, not the market. A review page where dozens of buyers all discovered the category through this one item is a page that has been composed, not collected.
How to use the checklist without becoming paranoid
No single sign convicts. Humans write gushing, detail-free reviews all the time — as gifts, in a hurry, out of genuine undiscriminating delight. The checklist works on accumulation: one sign means nothing, four signs together mean the review section is better treated as advertising than as evidence. Weight the verified-purchase reviews, the mid-star reviews (three-star reviews are the hardest to fake convincingly and the most informative when real), and the reviews with photos that show the product in an actual home rather than a studio void.
There is a seller’s side to this, too, and it matters if you ever run a shop: the same signs that expose bought reviews to shoppers expose them to platforms, and the penalties land on the business. What small businesses must know about AI-generated customer reviews covers the legal and platform risk in detail — the short version is that the fake-review economy is a bad bet from both chairs.
The harder middle case: real people, bought opinions
The checklist catches pure fabrications well. It struggles more with the scheme that has largely replaced them: refund-after-review clubs and rebate marketplaces, where genuine customers buy the product with their own money, receive a full refund through a side channel for leaving five stars, and write a review that is technically from a verified purchase. These reviews trip fewer wires — the account is real, the product was in the house — but they carry their own tells. The enthusiasm is uniformly total, because a four-star review does not earn the refund. The text often reads like it is written to the seller rather than to you: thanking the company, praising the packaging and customer service above the product, the telltale rhythm of someone fulfilling an agreement. And a cluster of them shares a subtle sameness of purpose even when the words differ — nobody in the group found anything to criticize at all. Platforms prohibit incentivized reviews they do not control, and regulators treat undisclosed paid praise as deceptive advertising regardless of whether the payer used cash or a refund, which is why this economy hides in private groups rather than operating in the open.
A word on photos and videos
Reviews with photos deserve their reputation as more trustworthy — a picture of the tent sagging in someone’s actual backyard is evidence in a way prose is not. But photos are not proof either. Sellers lift images from other listings, buyers in refund clubs take perfectly genuine photos of products they never chose freely, and unboxing-style videos are a paid genre of their own. Judge photos for what they show rather than the fact that they exist: a photo that reveals a flaw, a fit problem, or a size comparison against a hand or a doorway is worth ten that look like the manufacturer’s own gallery. The most useful review content, photo or text, is the content that costs the poster something to admit.
Why detectors won’t save you
It is tempting to paste a suspicious review into an AI-detector tool and let it judge. Resist the temptation to trust the verdict. These detectors disagree with each other, flag human writing regularly, and miss lightly edited generated text just as regularly — and the review ecosystem now includes humans imitating AI style and AI imitating human messiness, which collapses the whole signal. The checklist above works precisely because it does not ask “does this sound like a robot?” It asks “does this sound like a purchase?” That second question has a much better track record.
A practical routine pulls all of this together in about two minutes. Sort the reviews by most recent rather than most helpful, because campaigns age badly while honest recent reviews reflect the product as it is sold now. Read the newest ten and run the checklist mentally as you go — features, incidents, bursts, sameness. Then search the review text for the words that matter to your situation: “a year later,” “second one,” “returned,” the name of the specific problem you are worried about. Long-ownership mentions and repeat purchases are the closest thing reviews have to a gold standard, because a campaign can fake a first impression cheaply but almost never fakes a two-year relationship. If the newest ten pass and the long-term owners sound like people, you have done what can reasonably be done.
The realistic goal is not certainty. It is calibration: read reviews the way you’d listen to a stranger’s strong opinion in a store — useful for leads, worthless as proof, and always better in a crowd that disagrees a little. A product whose buyers argue in the comments about sizing is a product real people bought. Treasure the argument. It is the sound of the genuine article — imperfect, specific, contradictory, and written by people who demonstrably bought something and lived with it afterward.
