Short answer
There is no separate ranking system to game. ChatGPT Search retrieves from the web and synthesises an answer with citations, so the work is making your pages crawlable to its user agent, genuinely useful, unambiguous about what they are, and structured so a passage can be quoted without losing meaning.
How answers with citations get built
The pattern is retrieval followed by synthesis. A question is interpreted, often expanded into several sub-queries; candidate pages are retrieved; passages are selected; and a language model composes an answer from those passages, attributing parts of it to the sources it used.
Each stage gives you something to influence. Retrieval rewards ordinary discoverability. Passage selection rewards content that answers a question in a self-contained way. Synthesis rewards clarity. Attribution rewards specificity — a claim only a particular source can support is far more likely to be named.
Crawler access
OpenAI documents separate user agents for different purposes. OAI-SearchBot is the one associated with surfacing sites in search results within ChatGPT. ChatGPT-User represents a user-initiated fetch during a conversation. GPTBot is the crawler associated with model training. These are distinct controls, and search discoverability is a separate decision from training.
- Decide deliberately which agents you allow, rather than inheriting a copied robots.txt.
- If you want to be discoverable in ChatGPT Search, do not block OAI-SearchBot.
- Verify your robots.txt actually says what you think it says — test it, do not assume.
- Remember robots.txt is a public request, not access control.
Allowing a crawler makes you eligible, not selected. No robots directive causes a citation.
What appears to help
- 01Answer in the openingTwo or three sentences that resolve the question completely. This is the passage most likely to be extracted, and it helps human readers too.
- 02Name your entitiesAvoid pronouns where the subject should appear. A passage read in isolation must still be obviously about the right thing.
- 03Be specific enough to attributeOriginal data, a named method, a documented process, a stated limitation. Generic advice is paraphrased without credit because a hundred pages say the same thing.
- 04Structure for extractionOrdered headings, real lists, real tables, an FAQ block. A comparison in a table survives quoting; the same comparison in prose usually does not.
- 05Date your claims"As of 2026" is more quotable than an undated assertion, and it signals maintenance.
- 06Stay consistentContradictions between your own pages reduce confidence. Internal agreement is one of the cheapest wins available.
Technical prerequisites
- Server-rendered HTML — check view-source, not the inspector. Retrieval pipelines may not execute JavaScript.
- Fast, stable responses so fetches succeed reliably.
- Correct canonicals so signals are not split across URL variants.
- Accurate structured data describing what is genuinely on the page.
- A current sitemap so new pages are discovered quickly.
Measuring it
Fix a set of representative questions, run them on a schedule, and record whether you appear and how you are described. Add referral traffic from assistants as a floor, and watch branded search as a proxy for exposure that produced no click. See AI visibility tracking.
The same fundamentals apply across surfaces — see AI search optimization for the comparison, or what is GEO for the underlying discipline.
Last updated