Usually one of three things: OpenAI cannot crawl you, your pages have no cleanly liftable answer, or nothing independent corroborates that you exist. ChatGPT also answers from two different sources, training memory and live browsing, and the fix differs depending on which one is failing you.

The short version

Why this happens

The most common mistake is treating ChatGPT as one system. It is at least two, and they fail for different reasons.

Training memory. When ChatGPT answers without searching, it is drawing on what it absorbed during training. That snapshot is fixed, undated and cannot be edited. If your business was invisible, tiny or differently described when the snapshot was taken, that is what it knows. This is why firms sometimes see ChatGPT confidently state something about them that has been wrong for two years.

Live browsing. When it searches, it retrieves current pages and can cite them. This is the half you can actually influence in a reasonable timeframe, and it is where nearly all practical work happens.

Then there are the crawlers, and this catches people out constantly. OpenAI operates three distinct user agents: GPTBot for training data, OAI-SearchBot for the search index, and ChatGPT-User for fetches triggered by a user in conversation. They are governed separately in robots.txt. A firm that blocked GPTBot in 2023 because of understandable training concerns may have unintentionally kept itself out of nothing at all, or out of everything, depending on which agents the rule actually covered.

The second common cause is extraction. Your page is reachable, it is retrieved, and nothing in it can be quoted. If your answer sits below three paragraphs of context, there is nothing clean for the model to lift, so it lifts from somebody else.

The third is corroboration. ChatGPT is noticeably reluctant to name businesses it cannot verify from more than one place. If you exist only on your own website, that is a single unverified source, and it will frequently describe your category without naming anyone in it.

How to check it yourself

Four checks, about ten minutes:

  1. Ask it directly: “Tell me about [your company].” Then ask, “Did you search the web for that or answer from memory?” It will usually tell you, and that identifies which half is failing.
  2. Force a browse: “Search the web and tell me which UK firms do [your specialism].” If you appear here but not in the first test, your live retrieval is fine and your training footprint is thin. That is normal and it improves over time.
  3. Check your robots.txt for all three agents by name: GPTBot, OAI-SearchBot, ChatGPT-User. Our AI crawler guide lists the current set.
  4. Check Bing, not just Google. Search your key terms in Bing. Historic weakness there is a known blind spot for firms who have only ever optimised for Google.

What to do about it

  1. Unblock all three agents if any are restricted, and confirm no firewall or CDN rule is returning 403 to them. A permissive robots.txt proves nothing on its own.
  2. Put a liftable answer in the first 60 words of every page that matters.
  3. Fix Bing indexing. Claim Bing Webmaster Tools and submit your sitemap. It takes an afternoon and is routinely skipped.
  4. Build corroboration on sources beyond your own domain, so there is something to verify you against.
  5. Accept the training lag. Live retrieval responds in weeks. Training memory updates on OpenAI’s schedule, not yours.

Once you have diagnosed it, the full method is in how to appear in ChatGPT answers and how ChatGPT retrieves and cites sources.

Platform behaviour changes. This page reflects how ChatGPT worked in August 2026 and is reviewed quarterly.

Back to all GEO questions →