Somebody asks ChatGPT about your company and the answer is wrong. Old pricing, a product you retired, a founder who left, a category you were never in. The first instinct is to find the form that fixes it. There is no form.
What there is instead is two different mechanisms producing answers that look identical on screen, and only one of them responds to anything you do this week. Working out which one you are looking at takes about a minute, and it decides whether the rest of the work is worth starting.
Key Takeaways
- OpenAI's own crawler documentation separates training from search by name: GPTBot for training foundation models, OAI-SearchBot for ChatGPT's search feature.
- An answer with citations was retrieved. An answer without them was recalled from training, and your website has no direct route to it.
- In a complaint published by noyb on 29 April 2024, OpenAI's position was that it could filter or block outputs, not correct a fact inside the model.
- OpenAI's documentation states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, so the common defensive move removes your own correction path.
- The wrong fact usually has a specific published source. Finding it is the job, and it is rarely your own site.
The two places an answer comes from
OpenAI's crawler documentation names four agents and gives each a job. Two of them matter here.
GPTBot crawls pages to train foundation models. Whatever it collected is now inside a set of weights, and those weights do not change because you edited a page. OAI-SearchBot crawls pages so they can be surfaced in ChatGPT's search features, and OpenAI says plainly that content collected this way is used for live search results rather than for training.
That is the whole distinction, and it has a visible consequence. When ChatGPT searches, it shows the pages it used. When it does not search, it answers from memory and shows nothing.
So the diagnostic is: ask the question that produces the wrong answer, and look underneath it.
If there are citations, the answer was assembled from pages that exist right now, and you can read them. If there are none, you are looking at a recollection of the web as it was when the model was trained, and no edit you make today reaches it directly.
The two failure modes need different work, and most advice on this subject skips the diagnosis entirely, which is why so much of it does not work.
What OpenAI has said it can and cannot fix
The clearest public statement on this came out of a data protection case rather than a support page.
On 29 April 2024 the European privacy group noyb published a complaint it had filed with the Austrian data protection authority about ChatGPT stating a false date of birth for a public figure. The GDPR gives individuals a right to rectification, which is a right to have inaccurate personal data corrected. The complaint reports OpenAI's answer: it could filter or block data relating to certain prompts, but not correct it, and it described factual accuracy in large language models as an area of active research.
Two things follow, and the second is the uncomfortable one.
The first is that a wrong fact sitting inside model weights is not addressable by request, by OpenAI's own account. Blocking a topic is a different operation from correcting it, and blocking is what was on offer.
The second is that this was a case about a person, who has a legal right to demand rectification. Your company does not. Corporate information is not personal data, so the regulation that gave the complainant standing gives a business none. If the person with the strongest legal claim available was told the fix is unavailable, a business writing to support has less to work with, not more.
That is not a reason to give up. It is a reason to stop spending time on the channel that does not exist and spend it on the one that does.
The wrong fact has a source, and it is usually not your site
When the answer does carry citations, it was built from specific pages. Read them.
The useful surprise is how often the cited pages are not yours. A directory listing nobody has updated since the last rebrand. A comparison page written by somebody selling against you, describing your product as it was two versions ago. An old press release still ranking on a wire service domain. A conference bio from a talk in 2023. Your own site can be word perfect while every source an engine actually reached says something else.
This is worth doing carefully rather than quickly, because the citations tell you exactly which pages carry weight for this question, and that list is the actual work order. Correcting a page nobody cited changes nothing. Getting a correction into a page that is cited every time changes the answer.
A second check is worth the minute it takes: open the cited page and confirm it actually says what the answer claimed. Retrieval sometimes summarises a page into something the page does not support, and that is a different problem with a different fix. If the source is right and the summary is wrong, editing the source is not going to help. Adding an unambiguous statement of the fact, in plain sentences rather than implied by a table, sometimes does.
What moves a live answer
Once you know which pages are being cited, the levers are ordinary and mostly not exotic.
The page carrying the correct fact has to be reachable by the crawler that feeds the search step. OpenAI's documentation is explicit that OAI-SearchBot exists to surface sites in ChatGPT's search features and that sites opted out of it will not be shown in those answers, so a robots.txt rule you inherited three years ago can quietly disqualify the page you are counting on. Check that first, before writing anything.
The fact should be stated, not implied. An engine lifting a sentence needs a sentence to lift. "Founded in 2019 in Berlin" is liftable. A timeline graphic is not, and neither is a fact that exists only in your pricing table's fine print.
The correction needs to be in more than one place. One page saying the new thing against four saying the old thing is a fight the new page will usually lose, and the four are the ones the engine already found.
And the defensive reflex is worth naming, because it backfires. Blocking GPTBot is a reasonable choice about training data, and it does nothing about a wrong answer, which our post on whether blocking GPTBot removes you from ChatGPT search covers in detail. Blocking OAI-SearchBot is the one that hurts: OpenAI's documentation says sites opted out of it will not be shown in ChatGPT search answers, so you would be removing the only mechanism capable of carrying your correction into a future answer.
What nobody can tell you
Here is the part that vendors in this category tend to leave out, including the ones selling AI visibility.
Nobody can tell you that a specific change caused a specific answer to change. The engines do not report what they read, how they weighted it, or why one page won over another. You can observe that an answer said one thing in March and another in April, and that a page changed in between. You cannot observe the link between those two events, and anyone who tells you they can is describing a correlation and pricing it as a mechanism.
What you can do is measure the answer itself, repeatedly, and see which sources it names. That is a fact, it is checkable, and it changes.
What to do this week
- Ask the question that produces the wrong answer, and look for citations. No citations means a training recollection, and your website is not a direct route to it. Citations mean there is a work order underneath the answer.
- Open every cited page. Note which are yours, which are third party, and which say something you can get corrected by asking a human.
- Confirm the page carrying the correct fact is crawlable. A page the search step cannot reach cannot be cited, and our free robots.txt tester shows which rule an agent would follow.
- State the fact in a plain sentence on a page that already gets read, then get it into the third party sources that were cited. One corrected directory listing usually beats a new blog post.
- Check your robots.txt is not disallowing OAI-SearchBot. Removing yourself from the retrieval path is the one own goal available here.
- Then watch it. Run the free check on your domain and the topic you want to be found for. It needs no account and no card, puts a real buyer question to Google's AI Overview and to ChatGPT, and reports whether you were named and which sources were used instead. That list of sources is where the next hour of work goes.