Where Does ChatGPT Get Its Information?

By Muntazir · Published 28 October 2026

A specialty import shop asks ChatGPT, on a Tuesday, what it should know about running a small retail business through the winter. The answer is generic but reasonable — the kind of thing a model has absorbed from thousands of business articles. A customer asks ChatGPT about the shop itself the same afternoon. The answer this time cites a review published three weeks earlier. Same model, same hour, two completely different sources behind the words.

Where does an answer about a business actually come from?

Three places, and they are not interchangeable. The first is frozen training data — a fixed snapshot of text gathered up to a cutoff date, baked into the model and never updated between releases. The second is a live web search, run only when the system decides a question calls for one, drawing on an index built by OpenAI’s own crawler rather than on memory at all. The third, narrower than either, is content a publisher has specifically licensed to OpenAI, which can end up quoted or displayed inside an answer by arrangement rather than by either of the first two routes.

A question and its answer give no label saying which of the three produced it. That’s the actual source of the inconsistency above: not that the model is unreliable in some vague sense, but that “ChatGPT” is shorthand for at least three different ways of getting an answer, each with its own blind spots.

How does ChatGPT decide whether to search instead of answering from memory?

Automatically, and not every time the same way. OpenAI’s help documentation on search describes the system deciding on its own whether a live search would help, with a manual option sitting alongside it for when a person wants to force one. That’s a real capability, not a guarantee. A question that reads like a settled fact — what a company does, where it’s based — doesn’t carry the same obvious cue to go looking that a question about this week’s headlines does. Nothing in OpenAI’s own documentation spells out the exact threshold, and that’s worth saying plainly: the criteria aren’t published, so a business can’t engineer a phrasing that reliably crosses it.

Is OpenAI’s own crawler reading the same pages a search engine would?

Not the same crawler for the same reason. The one that builds the index ChatGPT’s search feature cites from is called OAI-SearchBot, and it’s a distinct piece of infrastructure from GPTBot, which only gathers pages to train future models. OpenAI’s developer documentation treats the two as independent: a site can allow one while blocking the other, and neither setting touches the other’s behavior. That distinction, and what it means for a site’s robots.txt file, is covered in more detail in a dedicated look at OAI-SearchBot versus GPTBot — the short version here is only that “ChatGPT” is not one crawler with one policy, so a rule aimed at blocking it can easily miss the half that actually feeds citations.

Do licensing deals with publishers put content straight into an answer?

For a narrow set of partners, yes — and this is the channel most businesses will never touch. OpenAI’s first major news deal, announced jointly with the Associated Press in July 2023, licensed part of the AP’s text archive going back to 1985 for use in training, with the financial terms undisclosed. A later arrangement with Condé Nast went further: OpenAI’s own partnership let ChatGPT display content from Wired, Vogue and The New Yorker directly inside its responses, not just as training material folded into a future model.

Both deals matter for understanding how an answer can be assembled, and neither extends past the specific partner it names. A local accountant or a regional retailer isn’t covered by an arrangement written for a national wire service, and there’s no sign-up path that puts an ordinary business into one. It’s a real channel — just one that exists for a handful of named publishers, not a lever most readers can pull.

Can a test actually show which channel answered a question about a specific business?

Only one of the three, by design. A deeper AI-visibility review runs a plain, unprompted question about a business — no supplied context, no browsing — across several current models at once, precisely to see what training memory alone holds, separate from anything a search or a licensing deal might add. That isolates the first channel cleanly. It does not run the identical question a second time with search switched on to compare the two, so it cannot say what a live search would have found instead — a different, harder thing to measure, and one this particular design doesn’t attempt. How that memory-only test actually sorts responses, including a failure mode easy to miss, is covered separately in a closer look at ChatGPT’s knowledge cutoff.

What does any of this change for a business trying to be found?

Not much for the licensing channel — that door simply isn’t open to most. The other two are where the actual leverage sits, and they call for opposite responses. Training memory can’t be edited after the fact; a business that opened after a model’s cutoff has to wait for a newer model and accept that the gap reappears for whoever opens after that one. Live search is the channel that rewards work done today: being easy for a crawler to reach, clearly described, and structured so a retrieval system can pull a clean answer out of a page rather than guess at one — a free discoverability check scores a page on exactly that basis, and the broader practice sits on the site’s AI-SEO page.

Three sources, one undifferentiated-looking answer, and only one of the three asks anything of the business on the other end of the question.