ChatGPT answers from three separate stores, and which one it uses depends on the phrasing of the question rather than a setting. Product questions usually trigger browsing; category questions often do not.
You cannot control which store it reaches for, and you can tell which one it used from the answer itself. That is enough to know whether this quarter’s work can reach a given question at all.
People ask what ChatGPT “knows” about their company as though there were one answer. There are three, they update on completely different timescales, and the same product gives you a different one depending on how the question is worded.
What are the three stores?
Training data
What the model absorbed before its cutoff.
- Updates on
- Model releases. Not your schedule.
- Tell it by
- No sources cited, confident tone, possibly out of date.
Browsing and search
Pages fetched during the request.
- Updates on
- Crawl cycles. Days, sometimes faster.
- Tell it by
- Inline citations and links you can click.
Memory
What this user told it in earlier conversations.
- Updates on
- That one person’s usage.
- Tell it by
- Suspicious familiarity with your company.
The third store deserves a warning rather than a strategy. Memory is per-user, so it inflates your own results and nobody else’s. If you have discussed your company in that account before, the engine may name you for reasons no buyer will ever benefit from. Test in a temporary session with memory off, always.
What decides whether it browses?
The question, not a toggle. And the mechanism is documented, if not by OpenAI in detail then by a directly comparable system: Anthropic states plainly that “Claude determines when to search based on the prompt”, and searches when a request “depends on information that is current, changing, or outside its training data” while answering directly from stable knowledge otherwise.
That is the shape of the behaviour across this class of product. Something in the question signals currency, specificity or uncertainty, and the system looks things up. Something stable and general, and it does not bother.
Stable, general, unchanged for years. Nothing about your last quarter can reach an answer like this.
Specific, constrained, and asking for something that changes. This is the question shape your work can influence.
This is worth stating as a caution rather than a rule, because the trigger conditions are not published in detail and they change. Treat it as something you verify per prompt rather than a law you can rely on.
Which crawler feeds the browsing store?
A specific one, and it is separate from the one people usually think about.
OpenAI documents that OAI-SearchBot is “used to surface websites in search results in ChatGPT’s search features”, while GPTBot collects data for model training and ChatGPT-User fetches pages when a person asks for one directly. Crucially, the settings are independent: a webmaster “can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot”.
OAI-SearchBot
Feeds store 2. This is the one that decides whether browsing can find you at all.
Allow
GPTBot
Feeds store 1, on a model-release timescale. Blocking it is defensible and does not affect search results.
Your call
ChatGPT-User
Fetches a page because somebody named it. Robots rules may not apply to this class of visit.
Ignores robots.txt
Worth knowing before you edit anything: OpenAI notes that a robots.txt change can take around 24 hours to register on their side, so a fix made this evening is not a result you can test tonight.
Is ChatGPT unusual in working this way?
No, and that is useful, because it means what you learn here transfers.
The three-store structure is close to universal among assistants that can search. What differs is the mix and the documentation. Perplexity is built search-first and documents the same crawler split, between a bot that surfaces and links sites and a fetcher for user-initiated visits. Google’s AI surfaces are grounded in the Search index rather than a purpose-built crawl, and Google states there are “no additional requirements to appear in AI Overviews or AI Mode” beyond ordinary Search eligibility.
Where ChatGPT is genuinely distinctive is the memory store, which is per-user and therefore invisible in anybody else’s results. No other major engine gives an individual tester such an easy way to fool themselves.
The store structure, the crawler split, and the fact that prompt shape decides which store answers.
The exact trigger conditions, which are undocumented and change, and the memory behaviour, which is ChatGPT’s own.
How do you tell which store answered you?
Read the answer rather than the content of the answer. Three tells, none of which require any tooling.
It browsed. You are in retrieval territory, the sources are inspectable, and your work can reach this question.
Probably training data. Check whether the facts are stale, because that is the signature failure of this store.
It knows things about your setup that a stranger would not. That is memory, and your test is contaminated.
What does this mean for what you track?
Two things, and both change how you build a prompt set.
Write prompts that trigger the store you can actually influence. Specific, current, constrained buyer questions reach browsing. General definitional questions reach training data, and measuring those is measuring something you cannot move for another model generation.
And keep the full answer text rather than a tally, because the tells above live in the text. A yes-or-no record of whether you were mentioned cannot tell you afterwards whether the mention came from a page you can fix or from a store you cannot touch.
The three-prompt version of this check is in the five-minute test, the wider retrieval architecture is in the RAG explainer, and how the engines split into families that behave this way is in the engine breakdown.
Can you make it browse more often?
Not directly, and the honest framing is that you influence the question rather than the engine.
Nothing you publish changes whether a given prompt triggers a lookup. What you can do is recognise that your buyers ask specific, constrained, current questions far more often than they ask definitional ones, so the questions that matter commercially are already the ones most likely to browse. The general definitional questions that answer from stored knowledge are also the ones least likely to end in a purchase.
That is a more comfortable position than it first sounds. The store you can reach is the store your buyers are using at the moment they are deciding, and the store you cannot reach mostly serves people who are still learning what the category is.
One caveat about this whole page. OpenAI documents its crawlers thoroughly and its answer-routing behaviour barely at all, so the browsing and memory mechanics above are read from documentation where it exists and from observable behaviour where it does not. Treat the crawler facts as settled and the routing behaviour as current best reading.