How to Fact-Check AI-Written Product Reviews in Ten Minutes
An AI-written review reads confidently whether or not it's right. That's the problem. The sentences about a helmet having 18 vents, or one bike light being brighter than another, look identical when they're true and when they're invented, and the invented ones cost you readers and rankings.
This is the check I run before publishing anything AI-assisted. It takes about ten minutes per guide once you know where to look, because the errors cluster in five places. Check those, and you've caught nearly all of it.
The short answer
Open the review next to the Amazon listings and check five things: every number, every comparison, every claim of experience, every price, and every award. Numbers come from the listing or get deleted. Comparisons must be true for both products. Experience claims go. Prices go. Awards must be true against the rest of the list.
Why AI gets product facts wrong
A language model writes what a product of that type usually has. If it's given the real listing it mostly stays within it; if it isn't, or if the listing is thin, it fills the gap with the typical value. A "typical" cordless drill has an 18-volt battery, so that's what the review says, whether this one does or not.
The same thing happens with comparisons. Asked to contrast two products, the model produces a contrast whether the data supports one or not, because that's what the sentence shape calls for. "Lighter than the X" gets written without anyone having weighed anything.
This is why the fix isn't a better prompt alone. It's a check against the source, every time.
The ten-minute check
| Minute | Check | How |
|---|---|---|
| 1 to 4 | Every number | Open each listing. Find each number in the review (vents, watts, lumens, inches, hours, speeds) in the listing title or bullets. Not there? Delete the sentence. |
| 5 to 6 | Every comparison | For each "more / less / lighter / brighter than the X", check both listings. Can't prove it from both? Cut it. |
| 7 | Experience claims | Search the text for "I tested", "we used", "after a month", "in our hands", "I found". Remove or rewrite as what the listing says. |
| 8 | Prices | Search for "$" and words like "around", "under", "budget-friendly at". Remove any amount. Relative words (budget, premium) can stay. |
| 9 | Awards and "best" | "Brightest" must be the highest lumen figure in the list. "Lightest" the lowest weight. "Best overall" must be the top pick, everywhere it's mentioned. |
| 10 | Outside names | Any brand or model not in the guide's product list is invented context. Remove it. |
The five places errors live
1. Numbers
This is most of the work and most of the errors. The rule is simple: the listing wins. If the listing title says 21-speed and the review says 18-speed, the review is wrong. If the listing doesn't mention a number the review states, the review made it up. A sentence with an unsupported number is worth less than no sentence, so delete rather than guess.
Pay particular attention to numbers that appear in a pro, because pros are where the model is most tempted to add a figure for weight.
2. Comparisons
"More vents than the Zacro." "Longer battery life than the Ascher set." Each of these is two claims, one about each product, and both have to be true.
The common failure is a comparison that's directionally plausible but wrong for these specific products: the review says A has more vents than B when B's listing says 24 and A's says 18. Check both sides. If either listing doesn't give the figure, the comparison can't be made.
3. Experience
Models write reviews in the voice of a reviewer who owns the product, because that's what most reviews in their training sound like.
"I've been using this for three months." "We took it on a wet trail." These are false claims on a site that hasn't tested anything, Amazon's rules prohibit them, and Google's guidance names them as a trust problem. Search for the first-person phrases and replace each with what the listing or the ranking actually supports.
4. Prices
Amazon requires prices to come live from its API with a timestamp, or not to appear at all. A price typed into a review breaks that rule and is stale within days anyway. Search for the currency symbol and for the words that usually precede an amount. Relative language, "the budget pick", "the premium option", is fine and more durable.
5. Awards and superlatives
If a guide calls a product "Best Budget", it should be among the cheapest. If it calls one "Brightest", no other product in the list can have a higher lumen figure.
And wherever the text says "the best" without qualification, it has to mean the number-one pick, not whichever product that paragraph happens to be about. This one takes a minute to check against the comparison table and catches an embarrassing class of error.
What to do when you find errors
Fix the sentence, not the whole section. The structure around an invented number is usually fine; only the number is wrong.
For a comparison that can't be proven, cut the comparison and keep the fact about the product in hand. For an experience claim, replace "we found it comfortable" with what supports the claim, like "the listing's padded chin strap is the comfort feature buyers mention".
If one guide has more than a handful of errors, the problem is upstream: the model wasn't given the listing data, or was given it and told to be "engaging" rather than accurate. Fix the input, then regenerate, rather than editing fifty sentences.
Doing it with AffSERP
Most of this checklist exists because general-purpose AI writes from a product name. AffSERP's X Best Review tool writes each section from the product's own listing data and the other products' facts, with rules against prices, testing claims, outside brands and unprovable comparisons, so the numbers start out sourced rather than guessed.
The remaining human check is the quick award-and-comparison glance, which the comparison table at the top of each guide makes easy.
The ChatGPT post covers how to prompt a general chatbot so it makes fewer of these errors in the first place, and the Google policy post explains why accuracy is the thing that decides whether AI-assisted pages rank.
FAQ
What are the most common errors in AI-written product reviews?
Wrong numbers (vent counts, lumens, weights, speeds), comparisons that contradict the listings, claims of testing that never happened, typed-in prices, and awards that aren't true against the other products. Nearly every error falls into one of those five groups, which is why a short checklist catches most of them.
Do I need to check every product section?
Yes, but not every sentence. The numbers, the comparisons and anything that reads like first-hand experience are where errors live. Descriptions of what a feature is for are rarely wrong; descriptions of how much, how many or how it compares often are.
What if the AI's number doesn't match the listing?
The listing wins, always. If the title says 21-speed and the text says 18, fix the text. If the listing doesn't mention the number at all, delete the sentence rather than guess, because a number with no source is the thing readers and Google both punish.
Can a tool do this check for me?
Partly. A tool that writes from the listing data in the first place produces far fewer errors than one writing from a product name, and some checks (prices, testing phrases, outside brands) can be automated. The comparison and award checks still benefit from a human glance.
Keep reading
