How Does ChatGPT Choose Its Sources? What OpenAI Documents and What Is Guesswork

ChatGPT picks sources in four steps: it decides whether to search, rewrites your question into queries, retrieves candidate pages, then ranks and cites some of them. OpenAI documents the first three in some detail and the last one only in outline.

The practical consequence is that no single ranking factor exists for you to optimize. What you can do is pass the eligibility check, then be the page that is easiest to quote and easiest to corroborate. The sections below take each part in turn.

ChatGPT chooses on its own whether a question needs the web, then sends targeted queries to search providers and reads what returns. OpenAI's help page says it typically rewrites your question into one or more targeted queries, and may send further, more specific ones after reviewing the first results.

OpenAI's own example turns a long question about CCR8 cancer drugs into a short query with a year in it, then follows up with a narrower one about a specific drug and conference. Our inference: the pages in play are those that match the rewritten query, not your original prompt. If memory is enabled, saved memories can also shape the rewrite.

This retrieve-then-write pattern is called retrieval-augmented generation (RAG); our RAG explainer covers the mechanics.

The same help page says ChatGPT search sometimes partners with other search providers and links Microsoft's and Shopify's privacy statements. OpenAI's launch post adds content supplied directly by publisher partners, among them the Associated Press, Financial Times and Reuters. Our page on whether ChatGPT uses Bing covers that layer.

One caution from OpenAI itself: search results and citations can be incomplete, outdated or incorrect. A citation shows where ChatGPT looked, not that the page proves the claim.

Eligibility is documented, ranking is not. A page must be crawlable by OAI-SearchBot and reachable through your host or CDN; after that, OpenAI says only that it ranks results on multiple factors and that placement is not guaranteed.

Documented: the eligibility check.

• Allow OAI-SearchBot in robots.txt and allow OpenAI's published IP ranges at your CDN or firewall.
• Sites that opt out of OAI-SearchBot are not shown in ChatGPT search answers but can still appear as navigational links. Robots.txt changes can take about 24 hours to take effect.
• A disallowed page can still surface as a bare link and title if OpenAI learns its URL elsewhere. OpenAI's publisher FAQ says to use a noindex tag, which the crawler can only read if it is allowed in.

Observed: what research found. A September 2025 University of Toronto preprint compared AI search with Google across many queries. For consumer-electronics queries in the US, a GPT search model drew 92.1% of its citations from earned media (independent reviews and comparison sites); Google's results leaned more on brand and social pages. Those were 2025 API models, so today's ChatGPT may differ.

The GEO benchmark (KDD 2024) found that adding statistics, quotations and cited sources raised a page's visibility by roughly 30 to 40% on one metric, while keyword stuffing did not help. It was tested on a research engine and Perplexity, not ChatGPT.

Our inference. Pages that open with the answer, state checkable facts and are backed by independent pages saying the same thing should be easiest for a model to quote. Nothing published proves ChatGPT weights them that way.

When ChatGPT does not search, it answers from patterns learned in training, so no individual page is chosen. OpenAI says its models learn from public web content, partner datasets and human-generated data, and that they do not store or copy the pages they learned from.

That is why a brand can be "known" without any page being cited. Our inference is that repetition across many independent sources is what survives training, which is why brand mentions matter for this route. It is also slow to change: a repositioned product stays stale until a retrieval step corrects it. See how often ChatGPT updates its knowledge.

Two OpenAI settings are easy to confuse. Disallowing GPTBot signals that content should not be used for training. Per OpenAI's crawler documentation, that setting is independent of OAI-SearchBot, so blocking GPTBot does not take you out of search answers. We explain the trade-off in why blocking GPTBot is usually a mistake.

You cannot see which route ChatGPT used for a given answer. Responses that search may show citations, and absent citations do not prove memory.

Pass the access check first, then make your page easy to quote and your claims easy to corroborate. Access is binary and checkable in an afternoon; the rest is slower and partly inference.

1. Check access. Allow OAI-SearchBot in robots.txt and confirm your CDN lets OpenAI's published IPs through. In AEOeye's crawler access study (384 well-known sites, fetched once on 3 October 2026; per-site CSV), 16 sites (4.2%) blocked OAI-SearchBot at the homepage: 1 of 193 SaaS sites against 11 of 38 news publishers. Brands rarely close the door in robots.txt, but 24 of 88 e-commerce and consumer homepages blocked or challenged our own crawler. A firewall can do what robots.txt does not. It is a convenience sample, so read it as a prompt to check your own logs. The directory hub has the method.
2. Answer first. Open every section with the direct answer, then the evidence.
3. State checkable facts. Add statistics, named sources and dates, the tactics the GEO benchmark tested.
4. Get independent pages to say the same thing. Reviews, comparisons and forum answers are the earned media the Toronto study saw AI search lean on.
5. Measure. ChatGPT adds utm_source=chatgpt.com to referral links, so GA4 can show the visits. To see whether you are named, ask your buyers' questions with search on. AEOeye's free audit asks three unbranded buyer questions live in ChatGPT and shows the full answers and the competitors named instead.

Our position: skip checklists that promise a ChatGPT ranking factor. OpenAI gives you an access checklist and no weights, so test, measure and treat the rest as hypotheses. For the wider playbook, see ChatGPT SEO.