How to Correct ChatGPT About Your Company: Trained Knowledge vs Search

ChatGPT answers a question about your company in one of two ways. It either answers from what the model learned in training, or it runs a web search and builds the answer from pages it retrieves.

Those two paths fail for different reasons and get fixed in different ways:

• Trained knowledge (retrieval off). The model answers from data frozen at a knowledge cutoff. When OpenAI's model pages were checked in August 2026, the GPT-5.6 family listed a Feb 16, 2026 cutoff, and the Instant path (chat-latest) listed Aug 31, 2025. Anything you changed after the cutoff does not exist for that answer.
• Search (retrieval on). OpenAI says ChatGPT search rewrites your question into one or more targeted queries and sends them to search providers. The answer then draws on whatever pages come back.

As of October 2026, ChatGPT decides per question whether to search. You don't control that switch. So the first job is diagnosis, not rewriting. If you're unsure whether the answer is stale or simply invented, start with the difference between a hallucination and an outdated answer.

Run the question yourself and read the answer closely.

1. Look for source links. If the wrong fact sits next to cited pages, the answer used search. Open each cited page and find the sentence ChatGPT paraphrased.
2. No links is not proof of training data. Without visible citations you do not know whether search ran. Treat it as probably parametric, and say "probably".
3. Test clean. OpenAI notes that ChatGPT may use saved memories when rewriting a search query. Test logged out or with memory off, or your own history may shape the result.
4. Date the wrong fact. If ChatGPT gives your old price or a product name you retired before the cutoff, trained knowledge is the likely source. If it quotes something that only ever appeared on one third-party page, search is the likely source.
5. Repeat a few times. Because search is discretionary, the same question can produce a searched answer one time and an unsearched one the next.

Record each result with the date, whether citations appeared, and which pages were cited. That record tells you which of the next sections applies.

If the error comes from trained knowledge, no edit you make today reaches that answer. The model's weights are fixed until OpenAI ships a new model with a later cutoff. That timing is OpenAI's, and it is not published in advance. Correcting the model inside a chat doesn't help either, as covered in whether correcting ChatGPT in a chat fixes it for everyone.

GPTBot is the crawler OpenAI uses to gather content that may be used to train future models. OpenAI's bots page says disallowing GPTBot indicates your content should not be used for training. It does not say allowing GPTBot improves how ChatGPT describes you. Nothing in the evidence measures whether training inclusion changes brand recall, in either direction. The honest label is unmeasured.

What you can do:

• Make the correct fact public, plain and consistent now, so it is available if a future model is trained on it.
• Fix the search path (below), because a searched answer can override stale training data when ChatGPT chooses to search.
• Set expectations internally. The playbook for facts past a knowledge cutoff covers what to publish and what to wait out, and how long AI takes to update brand information covers timing.

When search produced the error, the wrong fact came from a page. Often that page is yours.

Remove contradictions

Check every place a key fact appears: pricing page, about page, footer, help docs, old blog posts, press pages. If two pages state different prices or founding years, ChatGPT may retrieve the wrong one. Pick one current value and make every page agree. For prices specifically, see fixing AI answers that get your pricing wrong.

State the fact in a sentence

Write the fact as a short, specific statement on a page that can be fetched: "The Starter plan costs $40 a month." (Hypothetical numbers.) A plain sentence like this gives the engine something it can quote. It is the only on-page approach with a positive experimental signal, and it promises nothing.

Put it in the raw HTML

Vercel's crawler study found ChatGPT's crawlers fetch JavaScript files but don't execute them, so they can't read client-side rendered content. If your pricing table only appears after JavaScript runs, ChatGPT search may never see it. View the page source, not the rendered page, and confirm the fact is there.

Skip the shortcuts. Schema markup, llms.txt files and "AI-friendly" rewrites have no credible evidence behind them as a way into AI answers.

Sometimes ChatGPT cites a review site, a directory, a news story or a competitor comparison instead of you. In that case, your own pages are not the problem, and editing them alone won't change the answer.

Work through the cited pages:

• Directories and listings you control. Update the profile directly.
• Review and comparison sites. Most have a correction or update process. Send the current fact with a link to your page that states it.
• Old news coverage. Publishers rarely rewrite archives, but some add updates. Publishing a current announcement gives search something newer to find. See whether old news stories shape what AI says today.

The evidence shows a strong association between being named accurately on the sites engines already cite and appearing correctly in answers. It is an association, not a guaranteed effect. Also note that OpenAI's API documentation says the pages ChatGPT consulted often outnumber the pages it cites, so a wrong fact can come from a page you never see in the links.

None of the fixes above matter if OpenAI's search crawler can't fetch your pages. OpenAI's bots page says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. It also says each crawler setting is independent of the others.

That means you can block training and still allow search:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

robots.txt is only half the gate. OpenAI's help centre says eligibility also requires your host or CDN to allow traffic from OpenAI's published searchbot IP addresses. A firewall can return 403 to a crawler that robots.txt allows, and nothing in your dashboard will say so. Cloudflare changed its AI crawler defaults on 2026-09-15, blocking Training and Agent categories by default on ad-displaying pages for some accounts while leaving Search allowed. Check your settings rather than assuming.

The AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away. Once access is confirmed, the free checker on this site shows whether ChatGPT names you for three buyer questions. For the wider repair order across platforms, see the fix plan for AI getting your brand wrong.

Does blocking GPTBot remove my company from ChatGPT search? No. OpenAI documents GPTBot as a training crawler and OAI-SearchBot as the crawler for ChatGPT search, and states that each setting is independent. You can disallow GPTBot to keep content out of training while allowing OAI-SearchBot so your pages stay eligible for search answers. Blocking OAI-SearchBot is what removes you from ChatGPT search answers, apart from navigational links.

Will allowing GPTBot make ChatGPT describe my company correctly? There is no evidence that it will. OpenAI's bots page describes GPTBot only in terms of training and claims no visibility benefit. Any effect would arrive only when a future model is trained, and nobody has measured whether training inclusion changes how a model describes a brand. Treat it as a policy choice about training use, not a fix.

Can I make ChatGPT search my site instead of answering from memory? Not directly. As of October 2026, ChatGPT decides per question whether to run a search, and it rewrites the question into its own search queries. You can't trigger search for other users. What you control is whether, when it does search, it can reach a page that states the correct fact plainly in the HTML.

Does ChatGPT-User follow my robots.txt? Not reliably. ChatGPT-User handles actions a user initiates, such as asking ChatGPT to open a specific link, and OpenAI says robots.txt rules may not apply to it. OpenAI also says ChatGPT-User is not used to decide whether content may appear in search. For search eligibility, the controls that matter are OAI-SearchBot in robots.txt and your firewall or CDN settings.

How do I know if my fix worked? Re-run the same questions logged out, with memory off, several times over a few weeks. Note whether citations appear and which pages they point to. If searched answers now cite your corrected page or show the right fact, the search path has picked it up. Unsearched answers may stay wrong until a model with a later cutoff ships.