What ChatGPT search retrieves, as opposed to what the model just remembers
ChatGPT answers from two places: pages fetched live by its search crawler, and what the model absorbed in training. A prompt sorts your questions by path.
- Works in
- ChatGPT, Claude, Gemini
- You need
- Five questions a buyer would ask an assistant · One sentence on what you sell · Your domain and the pages that should answer those questions
- Written for
- how to rank on chatgpt search
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
ChatGPT answers from two places and only one of them fetches your page. Retrieval runs a search, pulls live pages and cites them. Recall answers from what the model absorbed in training, with no fetch and nothing to link to. Almost everything written about ranking on ChatGPT search collapses the two, which is why the advice so often does not match what people see. The prompt below sorts your buyer questions between them, because only one path is reachable this month.
Why the distinction decides what to work on
The two paths fail for opposite reasons. A retrieval answer that does not name you failed because your page was not found, not fetched, or not the best candidate. A recall answer that does not name you failed years earlier, in data you did not write, and no edit to your site changes it.
That is why a single action list for “ranking on ChatGPT” is usually half useless. Rewriting a pricing page is a sensible response to a retrieval miss and an irrelevant one to a recall miss. Sorting the questions first tells you which half of the advice applies to you.
The routing between them is not published. OpenAI documents its crawlers and says nothing about when the product searches and when it does not, so the prompt reasons from the shape of the question instead and is required to say how confident it is. A question about a current price pulls towards retrieval. A question about what a term means pulls towards recall. Those are tendencies, not rules, and the prompt is not allowed to dress them up as rules.
The retrieval path is mostly plumbing
Discovery and access come before writing. A page that OAI-SearchBot cannot reach, or that returns an empty shell to a plain fetch, is not a content problem yet.
OpenAI’s bots documentation names OAI-SearchBot as the crawler behind the search index that surfaces links in ChatGPT, separately from GPTBot for general crawling and ChatGPT-User for a single fetch a person asked for. Those are three different permissions and most robots.txt files treat them as one, which is the rule that quietly breaks the file.
Check the discovery layer with the llms.txt checker, which reads llms.txt, robots.txt and your sitemap together. Then take the wider view from the overview page, or go straight at the site level and technical side if the plumbing is what came back broken.
You are an AI visibility analyst. ChatGPT can answer a question from two
different places and they behave nothing alike. One is retrieval: the product
runs a search, fetches live pages and cites them. The other is recall: the
model answers from what it absorbed during training, with no fetch and no
citation. I want my buyer questions sorted between the two.
For each question I give you, return a block in exactly this shape. No table.
QUESTION: repeat it as I wrote it.
LIKELY PATH: RETRIEVAL, RECALL, or EITHER.
WHY: two sentences at most. Base this on properties of the question itself,
not on guesses about the product. Questions that turn on current facts, named
products, prices, availability, recent events or "best in 2026" pull towards
retrieval. Questions about stable definitions, established methods and how
something works pull towards recall.
CONFIDENCE: LOW, MEDIUM or HIGH, and say what would change it. You are
reasoning about tendencies, not reading a documented routing rule, because no
operator publishes one.
IF RETRIEVAL: the one page of mine that should be the fetched result, and the
single thing wrong with it today. If no page of mine could plausibly be
fetched for this question, say that instead and say what page would need to
exist.
IF RECALL: say plainly that this question is largely out of my direct reach
today, and name what would have to change on the wider web rather than on my
site. Do not offer me an on page fix for a recall question.
Then two closing sections.
THE RETRIEVAL SHORTLIST. A prioritised list of the actions that affect only
the retrieval path, most important first, each with the question it serves.
Discovery and access come before content: a page that cannot be fetched cannot
be retrieved regardless of how well it is written.
WHAT I CANNOT REACH. Everything above that sits on the recall side, stated
plainly as being slow, indirect, and mostly a matter of what other sites say
about me.
Constraints:
- Do not describe OpenAI''s routing, ranking or selection logic as if it were
documented. It is not. Say "not published" wherever you would otherwise
guess.
- Do not invent statistics of any kind, including how often either path is
used.
- Do not tell me a file, a schema type or a markup pattern guarantees
retrieval. Nothing does.
- If a question is too vague to classify, say so and give me the sharper
version rather than classifying it anyway.
What I sell: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
Five buyer questions: [FIVE QUESTIONS, ONE PER LINE]
My site and candidate pages: [DOMAIN, THEN ONE URL PER LINE WITH A FEW WORDS ON WHAT IT COVERS]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- Sets what counts as a good fetched result. "We sell EV charger installation to Manchester landlords" makes location and current pricing questions obviously retrieval shaped. "We do electrical work" gives the model nothing to reason about and every question comes back EITHER.
[FIVE QUESTIONS, ONE PER LINE]- Five is deliberate. The whole output is the contrast between them, and with two or three you cannot see the pattern. Include at least one question about a current price or a current best option, and at least one about how something works, so the split has something to separate.
[DOMAIN, THEN ONE URL PER LINE WITH A FEW WORDS ON WHAT IT COVERS]- The candidate pages, described in your own words rather than fetched. The model cannot reliably read your site, and a page you describe honestly produces a better answer than a page it hallucinated the contents of. Ten lines is plenty.
How to run it
- 01Separate the two paths in your own head first
Retrieval means the product ran a search and fetched pages, which is why links appear. Recall means the model answered from training with no fetch. The same chat window does both and rarely announces which. Everything in this prompt depends on you holding that distinction while you read the output.
- 02Write five questions with a deliberate spread
Include one that turns on a current price or a current best option, one about how something works, and one naming a competitor. The spread is what makes the classification legible. Five near identical questions produce five near identical blocks and tell you nothing.
- 03Run it and check the CONFIDENCE lines
Most blocks should come back MEDIUM or LOW. A run where everything is HIGH means the model is describing a routing rule it does not have access to, since OpenAI publishes none. That output is confident fiction and worth rerunning with the constraint restated.
- 04Act only on the retrieval shortlist
The recall side is real and it is not this week's work. Take the retrieval shortlist and start at the top, which will almost always be discovery and access rather than writing. A page that no crawler can find is not a content problem yet.
- 05Check the discovery plumbing the shortlist points at
Run the llms.txt checker, which reports your llms.txt, robots.txt and sitemap together. Those three answer the discovery question for anything that fetches pages. A missing sitemap on a site publishing new pages weekly is a bigger retrieval problem than any paragraph on them.
- 06Rerun the same five questions in a month
Keep the questions and the page list identical so the only variable is what you changed. There is no index status you can check and no ranking report you can pull, so a repeated measurement against a fixed input is the closest thing to feedback this problem offers.
Questions people ask
How do I rank on ChatGPT search?
There is no ranking to enter and no published selection logic, so the reachable version of the question is narrower: can ChatGPT search find and fetch your page at all. That depends on OAI-SearchBot being allowed in robots.txt, on your pages being discoverable through a sitemap and internal links, and on real text existing in the response before any JavaScript runs.
What is the difference between ChatGPT search and ChatGPT?
ChatGPT search fetches live pages and shows links to them. Ordinary ChatGPT answers from what the model absorbed during training, with no fetch and nothing to cite. The same chat window does both, and it does not reliably tell you which one produced a given answer, which is why the distinction has to be reasoned about rather than observed.
Which crawler feeds ChatGPT search?
OAI-SearchBot, according to OpenAI's own bots documentation, which describes it as the crawler behind the search index that surfaces links in ChatGPT. GPTBot is separate and is described as general crawling for OpenAI. ChatGPT-User is a single fetch made when a person asks ChatGPT to open a specific page.
If I block GPTBot, can I still appear in ChatGPT search?
That is the configuration many publishers choose, and it is expressible in robots.txt: disallow GPTBot, allow OAI-SearchBot and ChatGPT-User. Whether it produces the outcome you want in every case is not something anyone can verify from outside, because there is no index status to check. It does state your position clearly, which is more than a silent file does.
Can I influence what the model remembers about me from training?
Barely, and not on your own timetable. Training data is fixed at some point before a model ships, and what it absorbed about you came largely from other people's pages. The work that moves it is slow and looks more like public relations than SEO. Treating it as a channel you control is the most common mistake in this area.
Unsubscribe in one click. We never pass your address on.