A customer calls and says, "ChatGPT said you were open on Sunday." But you are closed on Sundays. Or ChatGPT describes a service you stopped offering years ago as if you still offer it. This is different from not being visible. You are being mentioned, but you are being mentioned wrongly.
In this post we explain why AI assistants give wrong information and the order to follow when you see a wrong answer about your brand.
Why does ChatGPT give wrong information?
OpenAI says this openly itself. According to its help page, ChatGPT can produce wrong or misleading answers and can sound confident even when it is wrong. This is called "hallucination."
The same page lists three limits that matter for brands:
- Knowledge cutoff: the model is trained on data up to a certain date. Unless tools are used, it does not know what came after.
- Sounding confident is not the same as being right: the model can give a wrong answer with high confidence.
- Lack of access: because of technical problems, paid content or robots.txt settings, it may not be able to reach the information on a site.
OpenAI's research on hallucinations adds one more detail. Regular patterns, such as spelling rules, are learned over time. But rare, unpredictable facts cannot be worked out from patterns alone, and this leads to errors.
The opening hours, address or service list of a small or mid-sized business is exactly this kind of fact. If there is little, scattered information about you on the web, the engine may fill the gap with a guess.
Gemini is no different. Google's Gemini help page says Gemini can present wrong information as if it were true.
Where does wrong information usually come from?
Wrong answers about brands usually rest on one of three sources:
- Outdated information: information that was right when the model was trained but changed later. A moved address, a changed price, a closed branch.
- A wrong source: an error on a page the engine reads at that moment. An outdated directory listing, an old news story, a forgotten page on your own site.
- Made-up information: information that is in no source, produced by the model to fill a gap.
Knowing which one it is decides what you do. That is why the first step is to keep a record.
Step 1: Record the wrong answer
AI answers change even for the same question. A wrong answer you saw once may not show up on the next try. Or the other way around. So:
- Write down the exact question where you saw the error.
- Note which engine and on which date you saw it.
- Take a screenshot of the answer.
- Ask the same question several times and on different days. Does the error show up every time, or only now and then?
- Ask the same question to ChatGPT, Gemini and Perplexity separately. Is the error in one engine or in all of them?
This difference gives you a clue. If the error shows up in every engine and every time, there may be a source on the web that carries that information. If it shows up now and then and in one engine, the model filling the gap with a guess is also a possibility.
Step 2: Find the source
If the answer has a source link, start there. ChatGPT's search help page says sources can be shown in answers that use search, and it suggests opening the source to check whether it supports the answer.
If there is no source, search for the wrong information on Google in quotes. Search with your old address, your old phone number or the name of the service you no longer offer. You will see which pages still carry that information.
We covered which pages the engines read in general in where does ChatGPT get its information.
Step 3: Fix the source
You cannot fix the engine directly. But you can fix the source it reads.
- If the error is on your own site, update the page. Instead of deleting an old page, redirect it to the new page.
- If it is in a directory, map or guide listing, update the listing or write to the site owner. Also check your Google Business Profile separately; according to Gemini's maps help page, Gemini can take address, website and opening hours information from Google Maps.
- If the right information is not written anywhere, write it on your site in a clear sentence. "We are closed on Sundays." "We have not offered service X since 2024." The engine should be able to find this sentence when it looks for an answer to the question.
Finally, check that your site is open to bots. Even if the right information is on your site, the engine may not reach it because of a robots.txt setting or a technical problem, as OpenAI's help page also notes.
Step 4: Report it to the engine itself
Both engines have a way to report a wrong answer:
- In ChatGPT, you can give feedback with the thumbs-down button under the answer. For harmful or illegal content, OpenAI has a content report form.
- In Gemini, you can send feedback with the "Bad response" button under the answer and state a reason. According to Google's feedback page, the web version also has a "Report legal issue" option.
These help pages give no timeline or outcome for when and how a report will affect the answer. So do not treat reporting as the only fix. Fixing the source is the real work.
If the wrong information harms your brand and is defamatory, talk to your legal advisor.
Step 5: Track the fix
It takes time for a fix to show up in answers. The engine has to read the page again, and sometimes old information on other sites has to be cleaned up too.
Ask the question you recorded regularly over a few weeks. Did the error disappear, did it become less frequent, does it continue in another engine? Add the general questions your buyers ask to the same list. That way you track not only the error but also your overall visibility. We explained how to do this in how to measure AI visibility.
AIShortlist does this tracking for ChatGPT, Gemini and Perplexity separately. On paid plans, the same approved question list is asked every week. The Sources screen lists the pages the engines show; you can start looking there for the page the wrong information comes from.
Is it possible to prevent wrong information completely?
No. Both OpenAI and Google say openly that their models can make mistakes. No one can guarantee that an engine will always talk about your brand correctly. What you can do is write the right information clearly and consistently everywhere, fix errors at their source, and track the results regularly.
If you want to see how your brand is mentioned today and in which questions, you can run a 10-question measurement with the free trial. The trial includes 1 brand and 1 measurement, and no card is required.