Think of ChatGPT as two things: a model with a memory that stops at its training cutoff, and a search feature that reads pages when you ask.
The search clock
OpenAI's crawler page describes two fetchers. OAI-SearchBot builds the index that ChatGPT search answers from, and OpenAI says a robots.txt change takes about 24 hours to show in search. ChatGPT-User fetches a page at the moment a user asks something that needs it, and OpenAI says it "is not used for crawling the web in an automatic fashion". So a page can be read today, and its contents can appear in an answer today, if the question triggers a search and your page is retrieved.
In our check on 2026-10-21, ChatGPT searched for all 25 buyer questions and cited 47 sources, so for questions about businesses and products it does lean on live pages.
The training clock
What ChatGPT says without searching comes from training and changes when OpenAI releases a new model. Between releases, publishing a correction does nothing to that memory. It does something to the next release, and it does something today for any question where search runs.
What this means in practice
- Re-check the questions you care about monthly, in a fresh session, and note whether the answer cites a page. A cited answer is fixable this week; an uncited one is not.
- Expect variation. The same prompt gives different answers from one run to the next, which is why the methodology page reads trends over several checks rather than single results.
- Keep a visible "last updated" date on the pages you want cited. It costs nothing and gives the assistant a reason to prefer the current version.
