Why AI Product Reviews Get Specs Wrong (and the Data That Fixes It)
Ask a general chatbot to review a popular cordless drill and it will give you a confident paragraph with a torque figure, a battery capacity and a weight.
Check them against the listing and usually one is wrong: the torque is from the brushless model, the battery is from last year's kit, the weight is for the bare tool and the page says "with battery". The sentence reads fine. The numbers aren't.
This post explains where those errors come from, why they cluster on the same few specs, what changes when the model is given the listing data instead of being asked to remember, and a checklist of the five specs to verify on every guide whatever wrote it. The fact-check post covers the full ten-minute check; this one is the why behind it.
The short answer
Wrong specs come from a model writing from memory: it recalls a product line, not your exact variant, and fills gaps with plausible numbers. The fix is to write from the listing's structured data and nothing else, which removes the guessing.
Five specs still need a human check because they vary between variants in ways the data can hide: battery and runtime, weight, capacity and dimensions, box contents, and compatibility.
Where the errors come from
Writing from memory
A language model doesn't look a product up. Asked about a product, it produces text that resembles what it read about that product and its relatives during training.
For a famous product that's a lot of text; for a mid-range one it's a few mentions. Either way, the model has the shape of the facts and not the facts, and when it needs a number it produces one that fits the shape.
Product-line confusion
Most products exist in families: the same name with a brushless version, a Plus, a 2024 revision, a kit with two batteries and a bare tool. The specs differ between them and the names barely do. A model writing from memory blends the family into one product, which is why the torque from one model lands next to the weight of another.
The exact variant
Even with the right product, a listing has variants: the 500 ml and the 750 ml, the 18 V and the 20 V, with battery or without. Your link goes to one of them. A model that was told the product name and not the variant describes whichever one it remembers best.
Invented precision
Models produce specific numbers because specific numbers read as authoritative. "Up to 12 hours of battery life" is the kind of sentence a review contains, so the model writes one, whether or not twelve is the number. The confidence of the sentence is unrelated to the accuracy of the figure.
What listing data changes
| Writing from memory | Writing from listing data | |
|---|---|---|
| Source of specs | Training text about the product family | The listing for the exact ASIN |
| Variant | Whichever the model remembers | The one you're linking |
| Missing numbers | Filled with plausible ones | Left out, or marked as unlisted |
| Discontinued products | Still described as current | Not in the data, so not in the guide |
| Typical error rate | One or two wrong specs per product | Close to the listing's own error rate |
| What still needs checking | Everything | The five specs below |
The important change is the third row. Given data and told to use only it, a model stops filling gaps, because the instruction is to leave them. The product sections lose the invented sentence about battery life and gain "the listing doesn't state runtime", which is less impressive and true.
The five specs that still go wrong
Listing data isn't perfect either. Sellers copy descriptions across variants, leave old specs in place after a revision, and list the kit's weight on the bare tool. These five are where the remaining errors concentrate, and they're the ones to check by hand.
- Battery and runtime. Varies with battery size, mode and what the manufacturer means by "up to". Check the number against the listing's spec table, not the bullet points, and against one owner review.
- Weight. With or without battery, with or without packaging, kit vs bare. Check which one the listing means and say it.
- Capacity and dimensions. Litres vs quarts, internal vs external, folded vs open. The commonest variant mix-up.
- What's in the box. Kits and bare versions share a product name. Check the variant you're linking actually includes what the section says.
- Compatibility. Which batteries, which phones, which sizes. The listing is often vague and the model fills the vagueness with a confident list.
Five specs, ten products, about one minute each: the ChatGPT post has the same check for guides drafted in a chatbot, where the first four rows of the table apply in full.
What to do when a spec isn't in the data
Leave it out. A product section that doesn't state the runtime is a slightly thinner section; one that invents the runtime is a wrong page.
If the spec is the one that decides the purchase, find it in the manual (manufacturer sites publish them) and cite that; the product descriptions post covers writing the section around what you have. What you don't do is let the model, or yourself, write a number because the sentence needed one.
Doing it with AffSERP
AffSERP's X Best Review tool is built on the right-hand column: product sections are written from the listing data for the exact ASIN being linked, including its variant, with the model instructed to use only the facts supplied and to leave out what the listing doesn't state.
Discontinued products never enter the guide because they aren't in the live data, and the Refresh option re-pulls the data later so a revised listing updates the page. The five-spec check above still runs in the editor before publishing; it's the last minute of a ten-minute job rather than the whole afternoon.
FAQ
Why does ChatGPT get product specs wrong?
Because it's writing from memory. A language model has read about a product line at some point in its training, remembers the general shape, and fills in the numbers it doesn't have with plausible ones. Without the listing in front of it, the battery life, weight, and capacity it writes are guesses dressed as facts.
How do I stop AI from making up product specs?
Give it the specs. A model that's handed the listing's structured data (title, variant, dimensions, capacity, what's in the box) and told to use only that stops guessing, because it has nothing to guess about. The error rate falls to the rate of the data, which is low. Then fact-check the handful of specs that still go wrong.
Which product specs do AI reviews get wrong most often?
Battery life and runtime, weight, capacity and dimensions, what's included in the box, and compatibility. All five have a common shape: a number or a list that varies between models in the same line, so a model writing from memory picks the wrong one.
Is an AI review with one wrong spec a problem?
Yes, more than it looks. A reader who knows the product spots it, and from then on doesn't trust the rest of the page. Google's review guidance asks for accurate, first-hand content, and a page with invented numbers is the opposite. One wrong spec per guide is the threshold to aim under.
Keep reading
