Where does ChatGPT get its information from?

Every GPT model has a training cutoff, a date after which nothing new was in its training set. Ask a question that depends on the model's memory alone, no browsing, no plugins, and the answer reflects whatever the web said about the topic before that date, filtered through everything else the model learned at the same time.

This is why an old address, a discontinued product line or a former business name can persist in answers long after the change happens. The model isn't checking a database. It generated an answer from patterns learned once, and it stays that way until a newer model is trained on newer data.

There's no way to edit this directly. OpenAI provides feedback controls on individual answers and separate channels for personal-data requests, but neither functions as a correction service for a single fact. The lever that actually moves things is what live search finds when it looks for you today.

ChatGPT search works differently. When it's active, the system retrieves current web pages relevant to the question and builds the answer from what it finds, citing sources inline. OAI-SearchBot is the crawler behind that retrieval, and OpenAI's documentation is explicit that blocking it removes a site from those cited results specifically, separate from whatever GPTBot does with training data.

That means a page published this week can show up in a ChatGPT search answer this week, which isn't true of the trained model at all. It also means the answer depends on what the crawler can actually read. A page that renders its content with client-side JavaScript, sits behind a paywall, or sits behind a bot-blocking CDN setting gives the crawler nothing to cite, even with a perfectly configured robots.txt.

Exactly how much of that retrieval runs on OpenAI's own crawled index versus a licensed search partner isn't something OpenAI has broken down publicly in a way stable enough to state as fact here. That detail has shifted before. Check OpenAI's current documentation rather than any fixed description of the architecture, including this one.

OAI-SearchBot, GPTBot and ChatGPT-User are separate robots.txt tokens with separate jobs. OAI-SearchBot pulls pages into live search answers. GPTBot gathers material for training future models and has nothing to do with what ChatGPT cites today. ChatGPT-User fetches one page at a time when a person or a Custom GPT asks for it directly.

A business that wants to stay out of model training but still show up in live answers can disallow GPTBot and allow OAI-SearchBot in the same robots.txt file. A business that blocks all three, often by accident through a security plugin or CDN default, disappears from ChatGPT search entirely while potentially still existing in whatever an older model already learned.

Check which of the three a current robots.txt allows before assuming a site is invisible for the wrong reason, or visible for a reason nobody intended.

Ask ChatGPT about the same business twice and the answers can differ, sometimes within the same day. Live search retrieves whatever pages rank or match best for that specific phrasing at that moment, and a slightly different question can pull in a different set of sources.

The model itself also introduces some variation in how it phrases and weighs what it retrieves. That's normal for a language model, not a sign anything is broken. It's also why judging visibility from one conversation is unreliable; a fixed set of questions tracked over weeks tells you far more than a single exchange.

None of that makes the system arbitrary. An answer reflects the sources available at the moment of the query, and those sources are the thing worth improving.