Find the exact source ChatGPT is citing by asking it again with web search on, correct the fact there first, then make your own site the clearest version of it. Fixing your own page alone is not enough if a directory, old review, or Wikidata item disagrees with you; recheck the same question in two weeks rather than assuming one edit is instantly everywhere.
This is the action side of a problem brand intelligence is built to track: once you know ChatGPT, Claude or Perplexity is stating something wrong about you, the next question is what to actually do about it. The mechanics of why engines hallucinate about brands at all are covered in When AI makes up facts about your brand; this guide is the fix path, in order.
Ask the same question again with web search turned on, and open every source it cites. ChatGPT's search mode, Claude's web search and Perplexity all show numbered links to the pages an answer drew from. Read each one. If one of those pages states the wrong fact, you have found the source to fix. If no source is shown at all, the engine is answering from its training data, and the fix is different.
Why it matters: fixing your own website does nothing if the engine is actually quoting a directory listing, an old press release, or a review site you have never touched.
Edit it where it lives, not where you wish it lived. If the source is your own page, update it directly. If it is your Wikidata item, sign in at wikidata.org and edit the statement with a source attached; Wikidata's own notability policy expects a verifiable citation for every claim, not just a correct-sounding one. If it is a directory profile, a Crunchbase entry, or a review site listing, most of those platforms let the business claim and edit its own profile.
Proof: Wikidata — Notability policy.
| Where the wrong fact lives | How you correct it |
|---|---|
| Your own page | Edit the visible copy directly, today |
| Your Wikidata item | Edit the statement at wikidata.org with a cited source |
| A directory or review profile | Claim the listing and correct it there |
| No source shown (training data only) | Cannot be edited directly; publish the correct fact clearly so the next training refresh picks it up |
Rule of thumb: a model has no edit button of its own. Every fix happens at a source it reads, not inside the chat window.
That usually means an engine with no live search is repeating an old training snapshot, and the honest answer is that you cannot edit that snapshot directly. OpenAI's own research describes hallucination as a structural result of how models are trained and scored, not a rare bug tied to one weak source, so there is no single page to patch in that case. What you can do is publish the correct fact clearly and consistently now, which is exactly what the next knowledge update or re-crawl has to find instead of the old version.
Proof: OpenAI — Why language models hallucinate.
Do this: treat it the same way you would treat a missing fact, not a bug to report. There is nobody on the other end to file a correction ticket with.
One Organization block with consistent sameAs links, an About page that states the key facts plainly, and a pricing page with exact figures give an engine one obvious place to default to. Structured data does not make a fact more persuasive on its own, Google is explicit that it helps machines understand content rather than acting as a trust signal, but it removes ambiguity when the JSON-LD and the visible copy agree. A pricing page is the single most common place this breaks down; the full fix for that specific fact lives in Why does AI get my pricing wrong, and how do I fix it?
Proof: Google — Understand how structured data works.
Bottom line: the fewer places your facts disagree, the less room an engine has to pick the wrong one.
Often within days on an engine that fetches pages live, and much longer on one that only answers from training data, so the honest timeline depends on which engine you check. Publish the correction, then re-ask the exact buyer question on ChatGPT, Claude and Perplexity over the following two weeks rather than checking once and assuming it is done everywhere. A fact can be fixed on one engine and still wrong on another for some time.
Good news: unlike a fact with no source anywhere, a fact you can trace to a specific page is one of the faster corrections to make, because the problem is concrete rather than diffuse.
None of this requires buying anything, and each step takes a few minutes:
Bottom line: the fix is rarely one edit. It is finding the right source, correcting it, and giving the engines time to catch up, not giving up after the first recheck shows no change.
Ask ChatGPT the same question again with web search turned on and open the sources it cites. That tells you whether it is quoting a page you can edit, a directory listing, or just an old training snapshot with no live source at all.
Edit it where it lives: your own page if it is yours, your Wikidata item if you have one, or the directory or review profile if that is where the wrong fact sits. A model has no edit button of its own, so the correction has to happen at the source it reads.
That usually means an engine without live search is answering from an old training snapshot. You cannot edit that snapshot directly, but publishing a clear, consistent, well-corroborated correct fact now is what gets picked up the next time that training data is refreshed.
An engine with live web search can pick up a correction within days of its next fetch. One answering only from training data can keep repeating the old fact for weeks or months, until the next knowledge update. Recheck the same question in two weeks rather than assuming one fix is instantly everywhere.
Not immune, but more resistant: one Organization block with consistent sameAs links, an About page that states the key facts plainly, and a pricing page with exact figures give engines one clear source to default to instead of whichever page happens to agree most often.
The free check reads your pages the way ChatGPT, Claude and Perplexity read them, and flags where your own schema and copy disagree. 3 minutes, evidence included.