ChatGPT wrong information about my product: how to fix
ChatGPT wrong information about my product? Trace a wrong price, spec or feature to its source, fix what you own and file the rest.
If ChatGPT or another AI assistant gives wrong information about your product, the error usually sits in a source the assistant read: your own page, a product feed, a retailer listing, a review or a comparison page. Find that source, correct it where you control it, ask for a correction where you do not, and use the assistant's report button last, because a report does not change the source.
This article is about product-level facts: the price, the plan or version, the specifications, the availability, and which features belong to which product. It is not about facts concerning the company behind the product or the people who run it. Those problems have different sources and different routes. A product fact has an advantage: it is checkable, it usually has an owner, and it usually lives on a page someone can edit.
- Wrong product claims come from sources, not from nowhere: your pages, feeds, structured data, retailer and marketplace listings, reviews and comparison pages.
- Trace first. Capture the exact answer, open every cited source, and find the page that states the wrong fact before you change anything.
- Fix what you own in one pass: visible page text, structured data and any feed must state the same facts, with clear version and identifier fields.
- For pages you do not own, send a short correction request with proof. Use the assistant's own report tool as a supplement, not the main route.
- Expect days for owned pages and longer for third parties. No route forces an assistant to change an answer.
Where AI assistants get product facts
AI assistants take product facts from retrieved web pages, from product data supplied by merchants and data providers, and from what the model absorbed in training. The first two can be corrected at the source. The third cannot be edited directly.
For ChatGPT's shopping results, OpenAI's help page is specific about part of this. As of the page read, it says product selection draws on structured metadata from providers, prices come from third-party providers, and titles and descriptions may be simplified by ChatGPT. Labels such as "budget-friendly" are generated by ChatGPT and are not promises, and review summaries are generated from public reviews that OpenAI does not verify. It also says price and shipping updates can take time to appear. Check the current OpenAI shopping help page before relying on any detail.
Related readAI Answers About Job Candidates: Audit and FixThat one page shows four origins for four kinds of fact. A price is a data fact, a label is a model-generated judgment, a rating summary is a digest of other people's words, and a description is a rewrite of your own text. Each needs a different fix.
Access matters too. OpenAI's crawler documentation says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links.
For Google's AI Overviews and AI Mode, Google's documentation says a page must be indexed and eligible to be shown in Search with a snippet, that there are no additional technical requirements, and that no special markup or AI text file is needed. Its advice on structured data is that it should match the visible text on the page.
Seven shapes a wrong product claim takes
Most wrong product answers fall into seven patterns, and the pattern points to the likely source. Naming the pattern before you search saves hours.
| Wrong claim | What it looks like | Usual source | Who can fix it |
|---|---|---|---|
| Wrong price | An old price, a sale price shown as regular, a wrong currency | Feed, retailer page, price in an old article | You for feed and own pages; the retailer for theirs |
| Discontinued feature | A removed feature described as current | Old blog posts, reviews, comparison tables | You for own pages; each publisher for the rest |
| Wrong specification | Wrong size, capacity, compatibility or certification | Reseller copy of an old spec sheet | You supply the data; the reseller updates |
| Wrong availability | Out of stock, not sold in a country, or still on sale after withdrawal | Stale feed or retailer listing | You for feed; the retailer for theirs |
| Blended versions | Version 2 features credited to version 3, or the reverse | Pages that do not state a version | Mostly you, by labeling versions |
| Competitor feature attributed to you | Another product's capability or limit appears under your name | Comparison pages, listicles, forum threads | The publisher, after your request |
| Invented detail | A claim with no page behind it | The assistant's own synthesis | Nobody directly; fix the surrounding sources |
The last row is the one people treat as the whole problem, and it is the smallest. Before accepting that an answer was invented, check whether a source exists. It often does, and it is often a page you did not know was indexed.
Trace a wrong claim to its source before you change anything
Tracing means capturing the answer exactly, listing every source the assistant showed, and finding the first page that states the wrong fact. Do it before editing, because an early edit destroys the evidence of what the assistant read.
- Capture the answerSave the full prompt, the full response, the date and the assistant. Screenshot any product card, since cards and prose can disagree inside one answer.
- Split it into claimsWrite each claim on its own line: price, version, feature, availability, specification. Mark the wrong ones and write the correct value beside each.
- Open every cited sourceUse the sources button or links the assistant shows and search each page for the wrong value. A citation can sit next to a claim the page does not make.
- Search for the wrong valueIf no cited page states it, search the exact wrong figure or phrase in quotes with your product name. Old pages that never appeared as citations often turn up here.
- Label each source by ownerOwn page, own feed, marketplace listing, retailer, publisher, review platform. The owner decides the route.
- Test your own pages against each otherTwo of your pages with two prices are the cause more often than any outside site.
Run the same prompt in more than one assistant and in two or three natural wordings, since the wrong claim may only appear in one. Treat citations as leads. An assistant can cite a page and still state something that page does not say, or combine two pages into one claim. A claim that no cited page supports is a finding in itself: fixing the cited pages may not be enough, and reporting is worth doing.
Related readAI answers about doctors and lawyers: what to fixFix what you own in one pass
Fixing your own sources means making the visible page, the structured data and any feed state the same facts in the same words. Do it in one pass, because a fix that updates the page but not the feed leaves two sources disagreeing, and disagreement is what lets a wrong value survive.
Product pages and documentation
Put the current facts in plain text where a crawler can read them without running scripts. Google's merchant listing documentation warns that dynamically generated markup can make Shopping crawls less frequent and less reliable. Price and availability change fastest, so they are the facts at risk.
State the version, edition or plan wherever a feature is described. "Includes single sign-on" is ambiguous. "Single sign-on is included in the Business plan and not in the Starter plan" is a sentence an assistant can lift without guessing. Update or label old release notes and pricing posts so they say which version they describe.
Structured data
Product structured data is markup that tells search systems a product's name, price, availability and identifiers. Google's documentation separates product snippets, for pages where people cannot directly purchase the product, from merchant listings, for pages where customers can buy from you, and says merchant listings require name, image and offers. It also says some experiences combine structured data with Merchant Center feeds, and that product snippets may use feed pricing when the page markup has none, which is another reason to keep both in step.
Use the identifier fields of schema.org's Product type: gtin for the trade item number, mpn for the manufacturer part number, sku for your own code, and model. They separate your product from similarly named ones, which is the cure for the borrowed-feature pattern when the confusion comes from a shared name. For versions, Google's variants documentation describes ProductGroup, which groups variants of one parent, each variant carrying a unique identifier such as sku or gtin. Mark each version as its own entity instead of describing three versions on one page.
Related readWrong company information in ChatGPT: how to fixFeeds
A feed is a file of product data sent to a platform on a schedule. For ChatGPT, OpenAI's commerce documentation says product feeds are available to approved partners who apply at chatgpt.com/merchants, supplied as CSV or JSON, with a full snapshot recommended at least once a day carrying up-to-date titles, descriptions, images, price and availability. The help page adds that Shopify merchants are integrated through Shopify Catalog. The pages read for this article describe no process for correcting a single listing, so the practical fix for a wrong price in this channel is to correct the data you send.
For Google Shopping surfaces, the Merchant Center product data specification says price must match the landing page, structured data and checkout, and availability must match too. It warns that conflicting data between your data source and your website can cause disapprovals, limited eligibility or incorrect displays, and that a product with an incorrect GTIN will be disapproved. Read the full page for your country.
Merchant Center also has an automations setting that uses landing page data to update price, availability and condition. Google's page says it is enabled by default, is not a replacement for regular data updates, and may not suit items whose price changes more than once a day. If your feed and page differ for a good reason, such as a member price, decide deliberately whether to leave it on.
Crawler access
Check robots.txt for OAI-SearchBot and the crawlers of the search engines you care about. OpenAI's documentation says each setting is independent, so you can allow OAI-SearchBot while disallowing GPTBot, the crawler for content that may be used to train its models, and that a robots.txt change can take about 24 hours to affect search.
Related readGoogle AI Mode reputation: audit what it saysPages you do not control need a correction request
When the wrong fact sits on a page you do not own, the route is a correction request to the owner, backed by proof. The table sorts the routes by source.
| Source | Route | Speed | Limit |
|---|---|---|---|
| Your page or feed | Edit, then request a recrawl | Days | Answer may lag the source |
| Marketplace listing you manage | Edit through your seller account, using the marketplace's own help pages | Varies | The marketplace sets its own rules; its procedure was not read for this article |
| Retailer or reseller page | Send corrected data to the partner contact | Days to weeks | They may not prioritize it |
| Review or comparison article | Correction request to the author or editor | Weeks, often no reply | Opinion is not an error |
| Review platform profile | Update the vendor profile; ask the platform about factual text | Weeks | User reviews usually stay |
| Forum or social thread | Reply with the correction; report only if rules are broken | Variable | Rarely removed for being wrong |
| Assistant's own synthesis | Report in the product, then fix surrounding sources | Unpredictable | No assured change |
The last column is the honest part. An article that calls your product "slow" states an opinion. An article that says it "has no mobile app" when it has had one for two years states an error. Ask only about the second.
What a correction request contains
Keep it to facts. Name the page, quote the exact sentence, state the correct value, give one link that proves it, and say what you would like changed. Offer a current spec sheet or screenshot. Send one polite follow-up after a reasonable gap and stop.
Expect a mixed response. In a published outreach test by the SEO software company Ahrefs, 26 authors were contacted, 14 replied, and 6 made verified corrections. That is one company's account of a small sample, still under way when written, so read it as a sense of the range, not a rate. Never offer payment, free products or links in exchange for a correction. It turns a factual fix into a commercial deal.
Reporting inside the assistants is a supplement
Every major assistant has some way to flag an answer. None of the pages read for this article promises a correction, a reply or a timeline.
For ChatGPT product results, OpenAI's shopping help page says you can use the thumbs icon on a product, use the menu in the product to report incorrect information, or submit its content report form. For Google AI Overviews, the AI Overviews help page says responses can include mistakes and describes clicking thumbs down, choosing "Report a problem" and choosing the category that fits. It notes that your most recent search and its results are included, so report from the search that produced the error. For Perplexity, its help center says to open the three-dot menu below a response and select "Report".
Related readWhat Siri, Alexa and Google Assistant say about youA good report is one fact long: the product, the claim, the correct value and the page that proves it. If the wrong answer carries a safety, health or regulated claim, escalate to your legal or compliance contact the same day instead of waiting for a correction cycle.
A worked scenario: a blended version and a borrowed limit
This scenario is invented, with generic parties, to show the order of work. A company sells a scheduling tool. A buyer asks an assistant whether it supports team calendar sync, and the assistant says no, and adds a per-user fee that the company dropped two years ago.
The company captures the answer and splits it into two claims. Both are wrong: team sync launched in version 3, and pricing moved to flat team tiers. Tracing finds two sources: a comparison article that describes version 2, and the company's own old release-notes page, which still shows the per-user fee under a title with no version.
The company fixes what it holds. It labels the old release notes with the version they describe and a pointer to current pricing, adds plan and version wording to its sync documentation, and gives version 3 its own identifier and a ProductGroup relationship in its markup. It then sends the comparison site a four-line request quoting the sentence and linking the current pricing and documentation. Two weeks later it reruns the same prompts.
What to expect on timing
Timing depends on the source. Owned pages change as fast as you publish and as fast as systems re-read them. Third-party pages depend on someone else's calendar. The assistant's answer lags both. OpenAI recommends a full product feed snapshot at least once a day, says a robots.txt change can take about 24 hours to affect search, and Google's recrawl guidance says crawling can take anywhere from a few days to a few weeks. Those figures describe when a source can change, not when an answer will.
Related readWhat Microsoft Copilot says about you: check and correctIn our experience a corrected owned page usually shows up in answers within days to a few weeks, third-party corrections take longer when they happen at all, and some answers hold until a stale page stops being the strongest source. These are ranges from experience, not promises.
| Usual timing | What you do | What to check |
|---|---|---|
| Day 0 | Capture answers, trace sources, label owners | Evidence saved before any edit |
| Days 1 to 3 | Fix pages, markup and feeds in one pass; request recrawl | Page, markup and feed agree |
| Days 3 to 10 | Send correction requests; file assistant reports | Each names a sentence and a proof link |
| Weeks 2 to 4 | Rerun the same prompts; one follow-up to silent publishers | Same wording, same assistants, dated notes |
| Weeks 4 to 8 | Recheck monthly; log any new stale source | Wrong claim gone, or down to one source |
Keep a dated log of each prompt, answer, source found, action and result. Add a change checklist to every price change or launch: pages, markup, feed, partner spec sheets and old posts that quote the old value. Reputation monitoring covers the recurring check for teams that want it run for them.
If wrong product answers are costing you sales now
Start with the trace, not the report. If the same wrong claim appears across several assistants and several sources, or touches a safety or regulated claim, a private review of what each assistant says and where it comes from saves time. Reputation Partners can run that review.
Send us the product, the wrong claims and the answers you captured. We will map each one to its source and tell you which fixes are yours and which need a request. Start with the private audit.
Questions people ask
Why does ChatGPT show an old price for my product?
Usually a source still carries it: a feed that was not refreshed, a retailer page, or an old article. OpenAI's help page says prices come from third-party providers and updates can take time. Correct the price at its origin, keep your page and feed identical, then rerun the same prompt after the sources have been re-read.
Can I apply to make ChatGPT show correct product data?
OpenAI's documentation says feeds are available to approved partners and merchants apply through a form. That is an application to supply data, not a way to buy a correction. Its help page says product results are selected independently and are not ads. Accurate data gives the system better material, but it does not oblige an assistant to use it.
Do I need special markup for AI assistants?
Google's documentation says no special markup or AI text file is needed for AI Overviews and AI Mode. It does advise that structured data match the visible text. Standard Product markup with accurate price, availability and identifiers still helps search systems read a product page, so keep it correct instead of adding anything AI-specific.
What if the assistant mixes up two similar products?
Give each product and version a distinct name, a unique identifier such as a GTIN or SKU, and its own page. Use variant markup only where the products really are variants of one parent. State in plain text what each one is and is not. Confusion between near-identical names usually persists until the sources separate them.
How do I prove an AI answer was wrong?
Capture it when it appears: the full prompt, the full answer, the assistant, the date and any cited links, plus a screenshot of any product card. Keep the page of yours that states the correct fact, with its date. A dated pair of the wrong claim and the correct source is what a report or a lawyer needs.
