ChatGPT Has Wrong Information About My Business: What Causes It and How to Fix It

ChatGPT has two ways to answer a question about your company, and they fail differently.

The first is trained knowledge. The model learned from a snapshot of the web that stops at a knowledge cutoff. OpenAI's model documentation lists a February 16, 2026 cutoff for the GPT-5.6 family and an August 31, 2025 cutoff for the faster Instant path (as of OpenAI's docs checked in August 2026). If you changed your pricing, name, address or product line after that date, the model's memory still holds the old version.

The second is live search. As of October 2026, ChatGPT decides per question whether to search the web. When it does, OpenAI says it rewrites your question into one or more targeted queries and sends them to search providers. It then reads the pages that come back and writes an answer from them.

So a wrong answer is either an old memory or a bad read of what search returned. You can't see which mode produced a given answer unless citations appear. Visible source links usually mean search ran. No links often means the model answered from memory. For a deeper split of the two, see how trained knowledge and search errors differ in ChatGPT.

Before you fix anything, classify the error. An outdated fact and an invented one need different work. The guide on telling a hallucination from an outdated answer covers this in detail. Here is the short version.

Run each test question logged out, or with memory turned off. OpenAI notes that saved memories can shape how ChatGPT rewrites a search query, so a logged-in test can differ from what a prospect sees. Ask the same question a few times. Answers vary, and one bad answer is a signal, not a verdict.

Be clear about the limits before you plan the work.

Not in your control

• The model's trained memory. Nothing you publish today reaches an answer given without search. It changes only when OpenAI ships a model with a later cutoff.
• Whether ChatGPT searches at all. Search is model-discretionary per question.
• Which queries it runs. ChatGPT rewrites the user's question, and you can't see the rewritten queries.
• Editing the output. There's no form that rewrites what ChatGPT says about you. See whether you can ask OpenAI or Google to correct your company's facts for what reporting can and can't do.

In your control

• Whether OAI-SearchBot is allowed in robots.txt.
• Whether your host, CDN or firewall lets OpenAI's crawlers through.
• Whether your key facts are in the HTML before JavaScript runs.
• Whether your own pages agree with each other.
• Whether you've asked third-party sites to fix wrong facts about you.

Every fix below works by removing a block or a contradiction. None of them guarantees a correct answer, because the model still decides what to say.

1. Stale training data

Compare the date you changed the fact with the cutoffs above. If the change came later, the model's memory can't know it. You can't fix this directly. What you can do is make sure the search path finds the current fact, because a searched answer can override an old memory. For realistic timelines, read how long it takes AI to update what it says about your brand.

2. OAI-SearchBot is blocked

OpenAI runs separate crawlers with separate jobs. OAI-SearchBot fetches pages for ChatGPT search. GPTBot fetches content that may be used for training. OpenAI states that each setting is independent, and that sites opted out of OAI-SearchBot won't be shown in ChatGPT search answers. A blanket rule written to stop AI training can remove you from search too.

A minimal rule that allows search looks like this:

User-agent: OAI-SearchBot
Allow: /

robots.txt is only half the gate. OpenAI's help centre says eligibility also requires that your host or CDN allows traffic from its published search bot IP addresses. Bot protection can return a 403 or a challenge page to a crawler your robots.txt allows, and nothing in your dashboards will flag it. Cloudflare changed its defaults on September 15, 2026 to block Training and Agent crawlers on ad-displaying pages for new domains and free-tier customers, while leaving Search allowed by default. A site can lose crawler access without anyone editing anything. 99Visibility's free AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away.

3. Your facts only appear after JavaScript runs

Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so they can't read client-side rendered content. If your pricing table, opening hours or plan names load in the browser after the page arrives, ChatGPT search may see an empty shell. To check, view the raw page source (not the rendered page in developer tools) and search for the fact. If it isn't in the source, render it on the server.

4. Conflicting facts on your site or elsewhere

If your pricing page says one number, an old blog post says another and the footer says a third, ChatGPT has three candidates. A likely cause of wrong answers is that the engine picked the one you forgot about. The same applies off-site. When ChatGPT cites a directory, review platform or old press article, that page is the source of the error. Fixing your own site doesn't change it. Pricing errors are the most common version of this; the page on why ChatGPT keeps getting your pricing wrong walks through it.

Work from the gates outward. A perfect page that ChatGPT can't fetch changes nothing.

1. Record the errors. Write down the exact questions, the wrong answers, the date, and any cited URLs. Test logged out. The free checker on this site gives a snapshot of how AI engines answer up to three buyer questions about you, which is a quick baseline.
2. Check robots.txt. Confirm OAI-SearchBot is allowed. Decide separately whether you want GPTBot for training. Neither choice affects the other.
3. Check network access. Confirm your CDN or firewall serves a normal 200 response to OpenAI's crawlers, not a 403 or a challenge.
4. Check rendering. Make sure every fact you care about appears in the raw HTML.
5. Check indexing controls. Make sure the pages with your key facts don't carry noindex or nosnippet by accident.
6. Make your own site agree. Pick one canonical statement for each fact (price, founding year, location, product names). Update or remove every page that contradicts it.
7. State facts plainly. Put the specific fact in a clear sentence on a fetchable page, such as "The Growth plan costs $99 a month." A plain, specific sentence gives the engine something it can quote.
8. Fix third-party sources. Go through the URLs ChatGPT cited. Ask each site to correct its page; the guide on asking third-party sites to fix wrong facts has templates and priorities.
9. Retest on a schedule. Ask the same questions again in a few weeks and compare against your record.

If you're dealing with many errors at once, the AI hallucination triage playbook helps you rank them by business damage.

Much of the advice in circulation doesn't hold up against primary sources.

• Adding schema or FAQ schema to correct ChatGPT. Google says structured data isn't required for its generative AI features. Ahrefs' matched-control study of 1,885 pages found adding schema produced no major uplift in AI citations on any platform. Existing markup does no harm, but it isn't a fix for wrong answers.
• Publishing an llms.txt file. Google says Search ignores it, and no other engine documents reading it.
• Rewriting pages in an "AI-friendly" style or chunking content. Google says you don't need to write in a specific way for generative AI search, and there's no ideal page length.
• Allowing GPTBot so ChatGPT can find you. GPTBot governs training only. Search depends on OAI-SearchBot.
• Blocking GPTBot and assuming you've left ChatGPT. You haven't. Search access is a separate setting.
• Hiding text for AI crawlers. Hidden instructions, cloaking or injected content aimed at AI systems is manipulation. It risks search penalties and your reputation, and you should never use it.
• Trusting one test. Answers vary by question, by session and by whether search ran. Judge by a recorded set of questions over time.
• Fixing your site while ignoring the cited source. If ChatGPT links to a third-party page, that page is the thing to fix.