A buyer asks ChatGPT which contract manufacturers make Class II medical devices in Ireland, and 4 names come back. Each name got there for one of 2 reasons. Either the model remembers the company from training, or a search it ran during the answer returned a page that mentions it.
OpenAI’s documentation describes the second path in enough detail to act on. It says nothing about the first. Neither does anyone outside OpenAI.
We wrote about the difference between being named and being cited in what an AI citation is. This post is about the choosing, read out of OpenAI’s own pages, with what we’d change at each of the 4 steps if the site were ours.
First, ChatGPT decides whether to search at all
OpenAI’s help page on web search, updated in August 2026, says ChatGPT “may search the web automatically when your question would benefit from current information”. The user can also force it. The page describes a Search tool in the composer.
A buyer choosing a vendor usually gets a search. A shortlist is the kind of question that benefits from current information. But if the search doesn’t fire, the answer comes from memory. No page of yours is read.
You can’t control this step. You can only notice, when you run the question yourself, whether citations appeared. We note that for every answer in our check.
Second, the question is rewritten into searches
The same page says that when ChatGPT uses a search provider, it “typically rewrites your query into one or more targeted queries that it sends those providers”. OpenAI lists the Microsoft privacy statement among the providers’ policies, which is as close as the documentation comes to naming Bing.
Two consequences for your pages. The query that finds you is a shorter, more specific rewrite you’ll never see, and there may be 2 or 3 of them per question. And the index being searched is not ChatGPT’s own. A page a web search engine can’t find is a page the rewrite can’t return.
There’s a third detail in the same section. Saved memories, where the user has them switched on, can shape the rewrite. Two buyers asking the same question can get different searches. One of them told ChatGPT last month which sector they work in. So a check run from your own account is a check of your account. We run ours logged out.
Third, pages are fetched, and 2 agents do the fetching
OpenAI’s crawler documentation names the agent that matters here: “OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features”. The same page says a site that opts out of that agent is left out of ChatGPT’s search answers, though it can still show up as a plain navigational link. It gives the fix: “we recommend allowing OAI-SearchBot in your site’s robots.txt file and allowing requests from our published IP ranges below”.
A second agent, ChatGPT-User, visits pages when a user’s action needs it. The documentation says it “is not used to determine whether content may appear in Search”, so allowing it is a courtesy to the user rather than a way in.
This is the one step that’s fully in your hands. It takes 10 minutes. Open your robots.txt. If OAI-SearchBot is disallowed, or a firewall rule blocks it, the rest of this post doesn’t apply to you yet.
Our own file names the agent in a group of its own. It repeats the 2 disallowed paths inside that group, on purpose: a group written for one agent stands alone. OpenAI also publishes the IP ranges, so a firewall can allow the agent by address rather than by name. And it says its search side needs about 24 hours to notice a robots.txt change. The check comes the next day.
Fourth, the answer is written from what came back
Here the documentation stops describing the mechanism and starts warning the user. “Search results and citations can be incomplete, outdated, or incorrect”. The user is told to “Open a cited source to check that it supports the answer, review when it was published or updated”.
Read that as a specification for your page. The assistant lifts a passage. The buyer is told to open the page, check the passage supports the claim, and look for a date.
A page with no date fails the second check at a glance. A page whose first paragraph doesn’t contain the claim fails the first. We run both tests on every post in this series.
What you can pay for, and what you can’t
The question comes up whenever this subject does, so here’s the documented answer. OpenAI sells advertising in ChatGPT. Its developer documentation describes “ChatGPT Ads” as a product that helps “customers discover relevant products and services”. It has its own agent, OAI-AdsBot, which the crawler page says “only visits pages submitted as ads”.
That’s a separate product with its own fetch. The 2 pages we read on 20 September 2026 describe ads as a product of their own. Neither describes a way to buy a place among the vendors an answer names. We’d treat any supplier who sells that the way our terms treat guarantees: “we would not trust a supplier who did”.
The 4 steps as a to-do list
The first, whether a search fires, is not yours. The second, the rewrite, you can’t see. You can write for it: the buyer’s sentence, answered in plain words, on a page a search engine has indexed. The third is robots.txt and the firewall. It’s the only step with a right answer.
The fourth is the page itself: the claim in the first paragraph, a date on the page, a source at the claim.
Most of the effort belongs in the third and fourth. The second is where nobody has the query, so advice about it is guesswork. We’ll show you where you stand today: 3 questions, 3 assistants, 9 answers, the vendors named and the pages cited, from the free check. If a page needs rewriting after that, the AI citations article comes with the ChatGPT and Perplexity answers recorded again 30 days after delivery.