ChatGPT has wrong information about your company? What to do
A price that no longer exists, a service you never offered or a company the AI thinks has closed. Why it happens, what can be fixed, through which channel and what is outside the company’s control.
Published 21 October 2026
Short answer
When ChatGPT or another AI assistant says something false or outdated about a company, the mistake comes from one of two places: what the model learned in training, or the pages it reads when it searches the web live. The second can be fixed, by correcting the company’s own site and the third-party sources the assistant consults. The first cannot be changed by the company: it only goes away when the provider trains a new model. So the first job is to work out which of the two the mistake comes from.
What is an AI hallucination about a company
A hallucination is an answer from an AI model that sounds confident and is false: an invented, mixed-up or outdated detail presented as fact. About a company, it is usually a price, a service, an address or the state of the company itself.
It is called a hallucination because the model is not lying on purpose: it produces the most likely answer from what it has, and when it lacks the right detail, it fills the gap. Google says so in its help page for AI Overviews: «AI responses may include mistakes.»
Two kinds of mistake are worth separating. Outdated information was true at some point, like last year’s pricing or an office that has since closed. Invented information was never true, like a service the company has never offered. In practice both are handled the same way: find where the mistake comes from.
Why ChatGPT gives false or outdated information about a company
ChatGPT gives false or outdated information about a company for two reasons: it answers from memory with training data that has a cutoff date, or it searches the web and reads a page that is wrong or out of date.
A model’s memory freezes at its knowledge cutoff. In OpenAI’s developer documentation, checked on 21 October 2026, its flagship model, GPT-6 Astra, has a knowledge cutoff of 30 April 2026. A price or name change made after that date does not exist for the model when it answers without searching.
Answering without searching is common. In a DoGeo study of 1,230 Gemini runs (October 2026), with questions about AI visibility asked in Spanish, 70% of runs did not issue a single web search; with the same questions in English, 46%. When each assistant decides to search is covered in the study of the sub-queries Gemini runs.
When the assistant does search, the mistake is usually in the source: a business listing nobody updated, a directory with the old address or a comparison written two years ago. Which pages each assistant reads before answering is explained in how AI models choose their sources.
How often AI assistants get facts wrong: the data
AI assistants often get checkable facts wrong. In a study coordinated by the European Broadcasting Union (EBU) and led by the BBC, 45% of answers about the news had at least one significant issue.
The EBU and BBC study, published on 22 October 2025, involved 22 public service media organisations in 18 countries working in 14 languages. Their journalists evaluated more than 3,000 answers from four assistants, including ChatGPT, Gemini and Perplexity. 20% of the answers contained major accuracy issues, including hallucinated details and outdated information.
The study measures answers about the news, not about companies, so it gives no error rate for questions about a brand. What it does show is that the two failures that hurt a company most, the invented detail and the outdated one, appear in assistants people use every day.
A mistake like that reaches real customers, because people already choose suppliers by asking AI. HeyTalent gets between 2 and 3 demos a week from people who arrived by asking ChatGPT, according to HeyTalent’s data, and Kulturo around 3 qualified form submissions a month from people who say they found it by asking ChatGPT or Claude, according to Kulturo’s data.
What can be fixed in AI answers and through which channel: comparison
A company can fix what the assistant reads, but not what the model learned. Each channel fixes a different part of the problem and depends on someone different.
| Channel | What it fixes | Who controls it | When it shows |
|---|---|---|---|
| The company’s own site | What the assistant reads when it searches live | The company | When the assistant’s crawler reads the page again |
| Google Business Profile | Address, hours, contact details and photos | The company, once the profile is verified | When Google publishes the edit |
| Third-party sources (directories, reviews, comparisons, Wikipedia) | What the assistant finds about the company outside its site | Whoever publishes each source | When the source is corrected and the assistant reads it again |
| Feedback to the assistant’s provider | Flags one specific answer as wrong | The provider (Google, OpenAI, Anthropic…) | No set time: the provider decides |
| Model training | What the model answers from memory | Only the provider | With a new model with a later cutoff |
What to do when ChatGPT says something false about your company, step by step
First save the answer and work out whether the assistant searched the web or answered from memory. Then fix the company’s own site, then the third-party sources, and finally tell the provider.
- 01 / RECORD
Save the wrong AI answer along with its question
Take a screenshot of the answer, copy the exact question and note the date and the assistant. Ask the question a couple more times: if the mistake does not come back, it was a one-off; if it does, it has a cause worth finding.
- 02 / DIAGNOSE
Check whether the AI answer cites sources
If the answer links to pages, open each one: the false detail is almost always in one of them. If it links to nothing, the assistant probably answered from memory, and the mistake comes from training.
- 03 / OWN SITE
Fix the detail on the company’s site and let crawlers read it
Write the correct detail on a clear page of the company’s own site (pricing, services, contact) and remove or date the old pages. Also check that the assistants’ crawlers can get in: according to OpenAI’s documentation, sites that opt out of OAI-SearchBot are not shown in ChatGPT search answers. You can check whether AI crawlers can read your site for free.
- 04 / THIRD PARTIES
Fix the third-party sources that repeat the mistake
Update the company’s Google Business Profile, which can be edited once verified, and the sector’s directories. For reviews, comparisons and media, write to whoever publishes them with the correct detail and a source that proves it. Which outside sources carry most weight in assistants’ answers is explained in the article on brand mentions in AI search.
How to report a mistake about your company to the AI provider
AI providers offer two routes: the feedback button on each answer, for any mistake, and privacy requests, only when the mistake involves someone’s personal data.
In Google’s AI Overviews, each answer has thumbs up and thumbs down icons at the bottom. According to Google’s help page, anyone who finds it inaccurate can click thumbs down, then «Report a problem», choose a category and add details. Google says it uses that feedback and human reviews to evaluate and improve its results, but it does not promise to fix a specific answer.
The privacy route applies when the mistake affects a person, for example the founder. Anthropic, in its privacy centre, asks for the name, email address, screenshots of the answer, the prompt used and the URLs referenced, and warns that it may decline the request if it has a lawful reason, such as a public interest in the information staying available.
Neither route changes the model. The feedback button is a signal to the provider; a privacy request, at most, filters what the assistant says about that person.
How long it takes for AI answers to change and what the company cannot control
Nobody can give a fixed time. What the assistant reads when it searches changes once it reads the corrected sources again; what it learned in training only changes with a new model, and that is not up to the company.
The part that depends on live search is the fastest, because the assistant reads the page at the moment it answers. If the site and the listings already say the right thing and the crawlers can get in, the corrected answer appears as soon as the assistant consults them again.
The part that depends on training cannot be touched by anyone outside the provider. In April 2024, the European privacy organisation noyb filed a complaint against OpenAI with the Austrian data protection authority because ChatGPT gave a wrong date of birth for a public figure. According to noyb, OpenAI replied that correcting the data was not possible and offered to filter or block answers about that person.
That is why the company’s site and third-party sources have to say the right thing even while the current model is still wrong: they are the material the next one will be trained on.
How to check whether AI still gives the false detail about your company
It is checked with two measurements: a fixed set of questions put to the assistants every month, and a log of AI crawlers in the site’s firewall.
What an assistant answers can only be seen by asking it. The questions have to stay the same, including the one that produced the mistake, and be asked several times in each assistant: that shows whether the false detail goes away and which source the new answer cites. How to set it up is in the guide to tracking AI visibility.
To know whether the assistants have read the corrected page, a rule is created in the site’s firewall or CDN that logs, without blocking, the requests from each AI crawler, identified by its user agent and, where the provider publishes its IP ranges, verified by IP. Those logs show which pages each crawler visits and when.
Search crawlers, such as OpenAI’s OAI-SearchBot, Anthropic’s Claude-SearchBot or PerplexityBot, show which pages enter each assistant’s index. The ones that open a page when someone asks a question, such as ChatGPT-User, Claude-User or Perplexity-User, are the closest signal that the assistant has read the corrected version in a conversation. This is what DoGeo sets up for each client, alongside the fixed questions.
Keep reading
What does AI say about your company today?
DoGeo reviews what assistants answer about your company, which source each detail comes from and what can be fixed. Leave us your name, company, website and work email.