How do you get cited by AI answers?

The only major provider that has published guidance on this says there is no special technique. Google's own words: optimising for generative AI search is optimising for the search experience, and thus still SEO. No special files, no special markup, no special writing style.

What does differ is narrower and more practical: which crawlers you let in, and understanding that the one which trains on you is not the one that cites you.

Everything below is from published provider documentation, checked 13 October 2026, with the sources named.

What does Google actually say?

Four things, each of which contradicts a popular tactic.

And the framing underneath them: the generative features are built on the same ranking and quality systems as ordinary search, so the work that makes a page rank is the work that makes it eligible to be summarised.

That is a less exciting answer than the industry around this would like, and it is the one from the company operating the surface.

So does anything technical matter?

Yes — one thing, and it is permission rather than optimisation.

Each provider runs several crawlers for different purposes, and they are controlled separately in robots.txt. Most people who set rules here set the wrong ones.

OpenAI, per its documentation:

Anthropic, per its documentation:

Which is the mistake people make?

Blocking the training crawler and assuming they have opted out of everything.

GPTBot is training. OAI-SearchBot is the agent that puts you in an answer with a link next to it. They are different tokens and they obey different rules. Block GPTBot and you have declined to be training data while remaining fully citable — which is, for most people, exactly the position they actually wanted.

The reverse mistake is worse: a blanket block that catches the search and user agents removes you from the surface where someone was about to click through to you. If a model is going to answer the question either way, you would rather be the footnote than be absent.

Which you choose is a real decision and not a technical one. This site allows all of them, because being quoted is the point.

What does a page need to be worth citing?

Something specific that is yours.

An answer engine summarising a question has a dozen pages saying roughly the same thing and one that has a number, a date and a method. The second is quotable; the first is interchangeable.

In practice that means:

• A figure with a source and a date next to it. "Stripe charges 1.5% + €0.25 for standard European cards, checked 2 October 2026" survives being lifted out of context. "Payment fees are low" does not.
• A method someone can repeat. The arithmetic shown, not just the conclusion.
• Saying when not to. The paragraph explaining who should skip this is the part nobody else writes, and it is disproportionately what gets quoted.
• A heading that is the question. Not because a model needs it, but because it makes the passage self-contained — and a self-contained passage is a quotable one.

None of that is a trick. It is the same thing that makes a page useful to a person who found it from a search result, which is Google's point.

What should you actually do this week?

Three things, in descending order of value.

1. Check your robots.txt reflects a decision you made. Not a copied block list. Decide separately about training, search and user-initiated fetching.
2. Put a date and a source next to every number you publish. This is the single highest-leverage habit here, and it is a writing habit.
3. Stop reading GEO advice that promises a mechanism. The provider operating the largest surface has published that there is no mechanism. Anyone selling one is selling the absence of evidence.

And the things not worth doing?

The ones in the table at the top of this page, on Google's own account: AI-specific files, special markup, chunking, and writing differently for machines.

Two of those have a separate justification worth knowing about, which is why they are not simply wrong — just not an SEO tactic. Does llms.txt do anything? covers the file, and which structured data actually gets used? covers the markup, including a type on this very site that stopped producing results in 2023.

Where this came from

Google's guide to optimising for generative AI features in Search, OpenAI's crawler documentation, and Anthropic's crawler support article — all checked 13 October 2026. Quotations are from those pages.

This area moves quickly and most of what is written about it is speculation presented as technique. The test to apply to any claim, including the ones on this page: who published it, and when did someone last check.

The AI Mastery Playbook is about building systems rather than chasing tactics, which is more or less the same argument as this article.