AI Hallucinations in Buyer Guides: The Seven Things Generators Invent and the Check for Each
A generated guide says the light has a brake sensor, that it won the 2024 design award, that it works with any handlebar, that it beats the brand's "Metro Ultra", that a reviewer called it "the best light I've owned", and that it launched in 2023.
Each sentence is fluent, specific and plausible, and none of it is true: the light has no sensor, there's no award, the mount is 31.8 mm only, the Ultra doesn't exist, nobody said that, and it launched in 2021. That's hallucination, and in product content it has seven shapes.
This post covers the seven things generators invent that aren't numbers (the specs post covers the numbers), why each one happens, a thirty-second check for each, and the line in the brief that reduces each. It's the companion to the fact-check post, which is the ten-minute pass; this is the field guide to what you're looking for.
The short answer
Generators invent features, model variants, awards, compatibility, dates, rival products and quotes, because each is a pattern in product writing that the model completes whether or not the fact exists.
Write from live listing data, ban awards, quotes and release dates in the brief, and check each of the seven against the listing before publishing. Thirty seconds each, three and a half minutes a guide.
The seven, at a glance
| Invention | What it looks like | Why it happens | The check | Brief line |
|---|---|---|---|---|
| 1. Features | "Has a brake-sensing mode" | Category pattern: most lights in the training data have one | Listing's feature list and manual: is it there? | "Only features present in the supplied data" |
| 2. Model numbers and variants | "The Metro Pro 1100 Plus" | Suffix blending across a product line | Exact name on the listing; the ASIN's variant | "Use the product name exactly as supplied" |
| 3. Awards and badges | "Winner of a 2024 design award", "Editor's Choice" | Marketing copy pattern; also Amazon-badge leakage | Search the award name plus the product; almost always nothing | "No awards, badges or 'choice' labels" |
| 4. Compatibility | "Fits any handlebar", "works with all apps" | Generalising from "fits most" | Listing's compatibility section; the manual's fitting range | "Compatibility only as stated in the data" |
| 5. Dates | "Released in 2023", "updated for 2025" | Pattern completion; the model has no calendar | "Date first available" on the listing; the manufacturer's page | "No release or update dates" |
| 6. Rival products | "Beats the Metro Ultra and the Lezyne Macro 1200" | Plausible-name generation within a brand's naming scheme | Search each named rival; discontinued or nonexistent is common | "Compare only against the other products in this guide" |
| 7. Quotes | "One reviewer called it 'bombproof'", "experts say" | Review-summary pattern; attribution without a source | Search the quoted phrase; if it's not findable, it's invented | "No quotations, no 'experts say', no named sources" |
Why each happens
A language model writes the most probable next words given what it's seen. Asked for a bike light description with nothing but the name, it produces the average bike light description, and the average one has a sensor mode, an award, a launch year and a rival. None of those are lies in the model's terms; they're the shape of the genre.
The invention rate falls sharply when the model is given the actual listing data and told to stay inside it, because the most probable next words are now the ones in the data. It falls further when the categories that are pure pattern (awards, quotes, dates) are banned outright, since there's no data they could come from.
The checks, in order of damage
- Quotes first. An invented quote attributed to a reviewer or expert is the one a reader can catch in a search and the one with legal edges. Search every phrase in quotation marks; delete any that isn't findable.
- Awards and badges. Amazon's badges ("Amazon's Choice", "Best Seller") are both an invention risk and a rule violation. Any award named: search it with the product. Nothing found means it goes.
- Compatibility. The one that causes returns. "Fits any" becomes the listing's actual range.
- Features. Read the product section against the listing's bullet list. Anything in the section and not on the listing is out, unless the manual confirms it.
- Rival products. Each named competitor gets a search. Nonexistent and discontinued models are replaced with ones from the guide.
- Model variants. The name in the guide matches the listing character for character.
- Dates. Delete them unless the listing's "date first available" confirms, and even then ask whether the reader needed it.
Three and a half minutes a guide once it's habit, and most guides written from live data pass every check with nothing to change.
What the brief does
The right column of the table is seven lines for the brief, and together they do more than the checks: they stop the inventions being written. "Only features in the supplied data; product names exactly as supplied; no awards, badges, quotes, sources or dates; compatibility as stated; compare only against products in this guide."
A tool that takes a banned-claims field, which the brief post covers, takes these as a block. The checks then confirm what the brief prevented.
Doing it with AffSERP
AffSERP's X Best Review tool writes product sections from the listing's live data through Amazon's API (the feature list, the exact product name, the stated compatibility), which removes most of rows one, two and four at the source.
Its defaults ban awards, badges, quotations and release dates and restrict comparisons to the products in the guide, which covers rows three, five, six and seven.
The seven checks are still worth running on the first batch from any new category, to confirm the data the listing supplied was itself right; after that they're a spot check.
FAQ
Why does AI make up product features?
Because it's completing a pattern, not reading a spec sheet. Asked to describe a bike light, it produces what bike-light descriptions usually contain (a mode, a sensor, an app), whether or not this light has them. The fix is to give it the actual data for the actual product and to constrain it to that; a tool writing from live listing data invents far less than one writing from a product name.
How do I fact-check AI content for product reviews?
Check the seven places inventions cluster: features, model numbers and variants, awards and badges, compatibility claims, dates, the rival products it compares against, and anything in quotation marks. Each has a thirty-second check against the listing or the manufacturer's page. Numbers (specs) are an eighth category with their own post; this one is about the things that aren't numbers.
What are the most common AI hallucinations in affiliate content?
In rough order: a feature the product doesn't have, a compatibility claim ('works with any mount'), a comparison to a model that doesn't exist or was discontinued, an invented release year, an award or 'editor's choice' nobody gave, a quoted reviewer or expert, and a wrong model suffix (the Pro where the listing is the Plus). The invented feature is the most common; the invented quote is the most damaging.
Can AI hallucinations be prevented in buyer guides?
Reduced to near zero, not eliminated: write from live product data rather than from memory, ban the categories that are pure invention (awards, quotes, release dates) in the brief, and run the seven checks on the output. A guide written from the listing's data with those lines in the brief typically has nothing to fix; one written from a product name has something in every section.
AffSERP builds product sections from live Amazon data and bans awards, quotes, dates and invented rivals in its defaults, so the seven checks are a spot check rather than a rescue.
Start with free wordsKeep reading
