AI models hallucinate product details when they have to fill a gap. OpenAI's help pages say it plainly about ChatGPT: it will occasionally make up facts or "hallucinate" outputs. A model that cannot find a clear, consistent fact about your product does not stop and say so. It produces a confident answer from whatever fragments it has, and the fragments may be an old price, a reseller's copy, a similar product from another brand, or nothing at all. The wrong detail usually has a source, even when the model invented the wording. Your job is to find which source it came from and fix that source.
There are two kinds of wrong detail, and they need different fixes. A retrieval error happens when the answer carries citations or links to live pages, and the wrong fact is sitting on one of those pages. A training error happens when the answer has no citations and reads like remembered knowledge. Fixing a page can change a retrieval error after the next crawl. A training error waits for a newer model, so the fix there is to make the correct facts consistent and easy to retrieve everywhere, then recheck on a later run.
Why the wrong detail exists
- Conflicting facts across the web. Your site says one price, a reseller says another, and a directory lists a third. The model picks one, and the one it picks may not be yours.
- Too little published. When a product fact appears nowhere clear on the web, the model fills the gap, and the fill is often wrong. New products and small catalogs see this most, because there are fewer pages about them.
- Stale copy still live. An old product page, an old price, or a retired variant is still published and still indexed, so the model reads it as current.
- Schema that disagrees with the page. The JSON-LD on a product page says one price and the visible page shows another. Google's guidance for AI features asks that structured data match the visible text on the page. A mismatch creates two competing facts, and the model may use either one.
- A similar name. Another brand shares your product name, or your product shares a name with a different product, and the model merges the two. The wrong detail then belongs to a product you never sold.
- A label or summary the model generated itself. OpenAI says ChatGPT may generate simplified product titles and descriptions from third-party provider data, and that labels like "Budget-friendly" are generated by ChatGPT, not verified statements. Review summaries are model-generated from reviews on public websites, and OpenAI says reviews and ratings are not verified. Some wrong details are generated, not copied, which means no single source page holds the exact wrong sentence.
Step 1: Capture the wrong detail exactly
Save the prompt, the full answer, and every citation or link shown. Note which model gave it and whether it carried citations. Test the same prompt in other models too, since the same wrong fact often appears in more than one place, and each one needs its own report. Do not paraphrase the error when you log it. The exact wrong sentence is what you will search for in step 2.
Step 2: Work out where it came from
Search the web for the exact wrong detail, in quotes. If it appears on a page, you have found a retrieval error, and that page is the fix. Check your own pages first, then reseller listings, directories, review sites, and old press. If the wrong detail appears nowhere, it is likely a generated fill or a training error, and the fix is indirect: publish the correct fact clearly and consistently, as in steps 3 and 4, so the next retrieval has something better to read.
Step 3: Fix your own pages first
Correct the fact on every page you control: the product page, the category page, the homepage, the About page, pricing, and the FAQ. Make the corrected fact easy to find in plain text, not only in an image or a script. Then make the structured data match it. A price in your markup that still shows the old number will keep the wrong answer alive, because the model reads the markup as a declared fact. Generate the markup from the same data that renders the page, never from a separate feed, so the two cannot drift apart.
Step 4: Reconcile the copies you do not control
Check the resellers, marketplaces, and directories the model cited. Make the name, price, availability, and product facts identical everywhere. Where a listing belongs to someone else, ask the site owner to correct it and give them the link to your corrected page. You cannot rewrite old press, but you can fix the listings and directories the model actually cites, and those are the ones that matter. Where two products share a name, add the distinguishing detail to your own pages, such as the model number, the size range, or the material, so the model has something to separate yours from the other one.
Step 5: Publish one canonical facts page per product line
A model fills gaps where facts are thin. Give it a dense, plain source to read: a product page that states the name, price, currency, availability, key specs, and what it is for, in ordinary sentences, plus matching Product schema. Keep the brand name identical in the markup, in your llms.txt file, and on the page itself. Consistent identifiers reduce the chance that the model merges your product with another brand's product of the same name. For the full treatment of brand consistency, see our guide to writing a good llms.txt file.
Step 6: Report the answer, after the source is fixed
Use the feedback control under the wrong answer in each model. For ChatGPT, that is the thumbs-down button, which OpenAI's help pages describe as the way to report an incorrect answer. Include three things: the exact wrong sentence, the correct fact, and the link to the page that proves it. Report it once per model. A report about a fact that is still wrong on the web changes little, which is why this step comes after the fix. For product listings specifically, OpenAI's shopping help page also describes reporting products with incorrect information through the menu on the product.
Step 7: Retest the same prompts on a schedule
Ask the same questions again with the same wording, and log the answer and the citations each time. Search-based answers can move after a recrawl. Training-based answers will not move until a model update, so keep the log and stop expecting a change between model releases. If the same wrong detail keeps coming back with the same citation, the cited page is still wrong, or the model is still reading a cached copy. Go back to step 2 and check the cited page again.
Say you run a pet food brand with a new recipe
A customer asks a model for the ingredients in your grain-free salmon kibble, and it lists chicken meal, which your recipe never contained. You search the exact wrong phrase and find it on a reseller page that copied an old listing from a different product in your range. Your own product page is correct, but the reseller page is the one the model cited. You ask the reseller to correct the listing, you check that your own page states the ingredients in plain text with matching Product schema, and you report the wrong answer with the thumbs-down button, quoting the wrong ingredient, the correct recipe, and the link to your product page. You retest the same prompt and log each result. The search-based answer changes after the next crawl. Any training-based version waits for a newer model.
The fundamentals
These mechanics are shared across every engine and covered in full in their own guides:
- •llms.txt: the file format, a complete sample, and per-platform publish steps, see our guide to writing a good llms.txt file
- •Product schema: the JSON-LD markup models read to verify price, availability, and specs in structured form, see our product schema guide for AI search
- •Every AI crawler in one reference table, see our robots.txt guide for AI search
What this guide does not cover
This guide covers wrong product details. It does not cover removing a listing, a legal dispute over a false claim, or the monitoring work of catching the next wrong detail early. It also does not cover getting recommended in the first place, which is a separate guide per model. If an answer is defamatory, bring in your legal team rather than treating a feedback button as the remedy.
Check your own brand now
Run the free scan. It takes seconds, no signup, and shows whether AI crawlers can read your corrected pages today: aizii.ai/aeo-scan
Run the free scanSources
- OpenAI Help Center, What is ChatGPT (specific): https://help.openai.com/en/articles/6783457-what-is-chatgpt
- OpenAI Help Center, Shopping with ChatGPT Search (specific): https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search
- Google Search Central, AI features and your website (specific): https://developers.google.com/search/docs/appearance/ai-features
- Google Search Central, Product structured data (specific): https://developers.google.com/search/docs/appearance/structured-data/product
- Schema.org, Product type (general standard): https://schema.org/Product
Note: links marked specific point to the page that supports the claim in this guide. Links marked general docs point to the vendor's general docs section for that topic, not to a page that proves every step. Check the vendor docs for the current path before relying on a menu name or click path in this guide.
