How to build a site an AI assistant can actually cite

Answer engines do not read the web the way a person does. A retrieval system fetches candidate documents, splits them into passages, scores those passages against the question, and hands the winners to a model with an instruction to answer and attribute. Every decision below is about surviving one of those steps.

1. Be fetchable

The most common reason a site is never cited is the least interesting one: the crawler was refused. AI crawlers are separate user agents from Googlebot and a robots.txt written years ago may be blocking them by accident — or a WAF rule, or a bot-protection product, or a rate limit that answers 429 to anything unfamiliar.

• GPTBot and OAI-SearchBot — OpenAI's training and search crawlers, documented separately.
• ClaudeBot — Anthropic's crawler.
• PerplexityBot — Perplexity's index.
• Google-Extended — not a crawler, a permission token that controls Gemini's use of content Googlebot already fetched.

2. Put the answer in the HTML

Many retrieval pipelines read the HTML response and nothing else. If your product description, your price, your opening paragraph or your FAQ only exists after hydration, it does not exist. Server-render the content that carries meaning — this is one of the strongest practical arguments for a framework that renders on the server by default.

3. Write passages that stand alone

This is the part almost nobody does, and it is the one with the largest effect. A retrieval system quotes a chunk, not a page. A paragraph that begins "As we saw above, this approach…" is unquotable: lifted out of the page it refers to nothing. A paragraph that begins "A WooCommerce checkout slows down for three reasons…" can be quoted verbatim and still make sense.

4. One page, one question

A page that covers nine topics competes with itself: no single passage in it is the best answer to anything. A page that answers one question completely — stated in the title, restated as the first heading, answered in the opening paragraph — is a much stronger retrieval candidate. This is also why a homepage that contains the entire site is a liability rather than a convenience.

5. Make attribution easy

An answer engine has to decide who to credit. A page with a named author, a publish date, a modified date, a canonical URL and consistent structured data gives it something to credit. A page with none of those is a text file of unknown origin, and a system deciding between two equally useful passages will attribute the one it can describe.

6. Cite your own sources

Pages that link to primary documentation read as derived from something checkable. It costs nothing, it is honest, and it puts your page inside a citation graph rather than outside it. Every article on this site ends with real links to the specifications and vendor documentation behind its claims.

7. Be fast, for a boring reason

Retrieval happens under a deadline. An assistant fetching six candidate pages while a user waits will drop the ones that did not answer in time. Time to first byte matters here in a way it does not in a ranking algorithm — not as a quality score, but as a timeout.

8. Describe yourself the same way everywhere

Mentions come from the description of you that the web agrees on. If your site, your LinkedIn, your directory listings and your clients' pages describe you in four different vocabularies, there is no agreement to learn. Pick the words a buyer would use for what you do and use those words everywhere, including in the places you do not control but can correct.

9. Measure, then decide

None of this should be done on faith. Ask the buyer questions your category actually generates, on more than one assistant, and record what came back before you change anything. Otherwise you are optimising against a story rather than a measurement — which is the same mistake the SEO industry spent fifteen years making.