ChatGPT has the wrong information about my business. How do I fix it?

First, work out whether the answer came from a search or from memory

Look for citations. If the reply carries links, it was assembled from pages fetched at the time you asked, and fixing those pages fixes the answer. If it carries no links, you are most likely reading the model's own memory of a web that has moved on, and no edit you make today can reach it directly.

The two paths run on different plumbing, and OpenAI documents them separately. GPTBot is the crawler that collects content which may be used to train the foundation models. OAI-SearchBot is the one that surfaces sites in ChatGPT's search features, and OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. ChatGPT-User is the agent that fetches a page because a person asked it to, and OpenAI says it is not used to decide whether content may appear in search.

That distinction is the whole of this page. Everything below fixes the retrieval path, which is the path you can actually reach.

Fix the source page, not the chatbot

Correct the page the wrong fact came from, and make the correct version the only version a machine can find.

• State the fact once, plainly, in text. Not in an image, not in a PDF, not only in a booking widget that loads after the page does.
• Remove or redirect the stale page. An old services page still returning 200 is a live source of the wrong answer. Redirect it to the current one.
• Put it in structured data too. A JSON-LD block that states your name, address, phone, hours and services leaves less room for a model to infer.
• Make sure it can be read. If robots.txt blocks the retrieval crawlers, the corrected page cannot reach the answer. See should I block AI crawlers.

Contradicting yourself on your own site is the most common own goal here: a homepage that says one thing, a footer that says another, and a blog post from three years ago that says a third.

Fix the same fact everywhere else it is published

Your website is one source among many, and it usually is not the one the model trusted. Work through everywhere else the fact appears.

• Google Business Profile, Apple Business Connect and Bing Places
• Industry directories and association member listings
• Review platforms with a business description field
• Social profiles, including the ones you stopped using
• Old press coverage and supplier or partner pages that describe you

The goal is not tidiness. It is that every source a model can reach says the same thing, so there is no older, more widely repeated version of the fact for it to prefer. If four sites say your old address and one says the new one, the wrong answer is the well supported one.

Use the in-product feedback control, and expect nothing from it

Send the thumbs down on the answer, add a sentence saying what is wrong and what is correct, and then carry on with the rest of this page as though you had not.

The feedback control is not a correction form. There is no published turnaround, no case number, no confirmation, and no mechanism we are aware of by which a single report edits a fact for everyone. It costs you thirty seconds and it is the only direct signal available, so it is worth sending. It is not a plan.

Be wary of anyone who offers to submit a correction to OpenAI or Google on your behalf for a fee. There is no such channel for an ordinary business.

Re-test on a schedule and record what changed

Set a date, ask the same questions in the same words, and write down what came back. Without a fixed question set you cannot tell a real fix from a lucky run.

Ask the wrong-fact question and your ordinary buyer questions together, keep the answers verbatim, and note whether citations were present. That last column is what tells you whether you are watching the retrieval path move or waiting on the training path, which is the difference between a fix that has landed and a fix that cannot land yet.

We will not tell you how long it takes, because nobody outside the model providers can. The reasoning is in how long does it take to show up in ChatGPT.

Where to start

Everything above you can do yourself with a spreadsheet and an afternoon. If you want the re-test run for you across both engines, with the transcripts kept so a change is provable, that is what our monitoring does. Start with the free scan.