Find the page the wrong fact comes from, fix it there, and make sure your own site states the right fact plainly. If the answer cites sources, the wrong page is usually among them. If it answers from memory, there is no page to fix, so you repeat the right fact clearly in the places models learn from and check again after newer models arrive.
There is no form that edits what ChatGPT says about a company. OpenAI offers privacy requests for personal data about people, and Google lets you report a problem with an AI Overview, but we found no official route at OpenAI, Google or Anthropic for correcting a claim about a business. The fix sits upstream, in the pages.
That sounds slower than it is. A wrong fact is often an old fact: a retired price, a former positioning, a product that no longer exists. They come from pages someone forgot to update.
Where do wrong facts come from?
From an old page, a stranger’s page, or the model’s memory.
Three places.
- An old page that is still live. A directory profile with your price from two years ago. A partner page describing a product you retired. Your own blog post announcing a plan you no longer sell.
- A third party who got it wrong. A roundup that confuses you with a company of a similar name, or a review that describes a feature you never had.
- The model’s memory. With search off, the assistant answers from what it learned in training. If the web described you a certain way back then, the model may repeat it now.
The first two show up in search answers and can be traced. The third cannot be traced to a page, which is why it is the slowest to fix.
A wrong fact is often an old fact.
How do you trace a wrong fact to its source?
Ask the same question with search on, then follow the citations.
Fifteen minutes per wrong fact.
- Write down the exact question that produced the wrong fact, and the wrong sentence word for word.
- Ask the question again in a fresh chat with web search on. If the wrong fact appears, open every cited page and search each one for it.
- Ask it with search off, or in a temporary chat with personalisation off. If the wrong fact appears there and not with search, it is most likely coming from memory.
- Search the web yourself for the wrong fact in quotes, plus your company name. Old pages that no assistant cited today may still be what a model learned from.
- Record the page, the owner, the wrong text and the right text.
Do it in Claude, Gemini and Perplexity too if the fact matters. Each reads a different mix of pages, and a wrong fact that shows up in two assistants usually has a source both of them read.
How do you fix it at the source?
Fix your own pages today, ask others to fix theirs, then say it the same way everywhere.
| Source | What to do | How fast it can show |
|---|---|---|
| Your own site | Correct the page. Put the right fact in text, near the top. Remove or redirect old pages that contradict it | Once the page is crawled again |
| A profile you control (directory, review site, LinkedIn) | Log in and update it. Check price, category, location and description | Once the profile is crawled again |
| A page someone else owns | Email the author or editor with the wrong text, the right text and a page that proves it | When they update it, if they do |
| The model’s memory | State the right fact plainly on your site and in the places you are written about | When a newer model is trained |
Then make the fact consistent. If your site says one price, your directory profile another and a partner page a third, the assistant has three stories and may pick the wrong one. Write one short fact sheet: what you do, who for, where, how pricing works. Use those words everywhere.
Make sure the search crawlers can read the corrected page. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, and Google says a page must be indexed and eligible to show with a snippet to appear in its AI features.
Can you report it to OpenAI, Google or Anthropic?
Use the feedback buttons, but do not wait on them.
Assistants have a thumbs-down or feedback button on their answers. Google’s help pages describe giving a thumbs-down on an AI Overview and then choosing to report a problem. Use them. They tell the company the answer was wrong.
They do not edit the answer. And the formal routes we found are about personal data: OpenAI and Anthropic describe privacy requests for information about people. If the wrong fact is about a person, such as a founder’s name or role, a privacy request may apply. If it is about the company, the fix is the source.
How do you know the fix worked?
Same question, same wording, once a month.
Keep a correction log: the question, the wrong sentence, the source, what you changed and the date. Ask the same question with search on and off each month and note what it says now.
Search answers often change first, because they read the pages as they are today. Memory answers change when a newer model arrives. Answers also vary between runs, so look for the wrong fact across two or three runs before calling it gone or calling it stuck.
You cannot argue with an answer. You can change what it reads.
See the use case for defense and dual-use software: How does a defense or dual-use software company get named when a program office asks an assistant what to pilot?
See the check page for ChatGPT: Does ChatGPT name you when buyers ask?
Questions people ask next
How do I correct false information ChatGPT says about my company?
Find the source. Ask the question with web search on and open the cited pages. Fix your own pages and profiles, ask the owners of other pages to update theirs, and state the right fact the same way everywhere. Then ask again monthly.
Can I ask OpenAI to change an answer about my business?
Not directly. You can flag an answer with the feedback button. OpenAI’s formal requests are about personal data about people. For facts about a company, the fix is in the pages ChatGPT reads.
Why does ChatGPT still say our old price?
Either an old page with that price is still live and being cited, or the model learned it in training. Search answers can change once the page is fixed and crawled again; memory changes when a newer model is trained.
Should I block AI crawlers to stop wrong answers?
Blocking a search crawler removes your corrected pages from what it reads too. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers. Usually the better move is to fix the pages.
Can Throughline find what AI assistants get wrong about us?
Yes. Request an AI visibility check. A growth strategist reviews how AI assistants see your company and replies by email within two business days.