Generative Engine Optimization: Getting Cited by AI Search

How do AI answer engines decide what to cite?

They extract passages, not pages. An answer engine retrieves candidate documents, splits them into passages, and assembles an answer from the ones that state something specific, verifiable, and self-contained. A page ranks into the candidate set the ordinary way, crawlable, fast, linked, credible. It gets cited because one passage on it answered the question outright.

That difference is the whole discipline. Classical SEO optimises a page to be chosen. GEO optimises a paragraph to be quoted.

The three-part model

Every citation we have been able to trace back came from a page that got all three of these right.

1. Reachable. The engine's crawler has to be allowed in, and it is not the same crawler as Googlebot. GPTBot, PerplexityBot, ClaudeBot and Google-Extended are separate user agents with separate rules. Blocking them by accident is the most common cause of zero AI visibility we find on audit.

# robots.txt, allow answer engines explicitly
User-agent: GPTBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: Google-Extended
Allow: /

2. Parseable. Client-rendered content is a coin flip. Most answer-engine crawlers do not execute JavaScript, or execute it on a budget that a heavy single-page app exhausts. If the substance of the page is not in the initial HTML response, assume it does not exist. Server-render the content, and check by fetching the URL with curl and reading what comes back rather than trusting what the browser shows you.

3. Extractable. This is the part that is actually new. A passage gets lifted when it can survive being separated from its page:

• A heading that states the question a reader would actually type
• An answer in the first 40–60 words underneath it, with no pronoun pointing back at the previous section
• Concrete nouns and numbers instead of adjectives. "settles in under 900µs" beats "blazing fast", every time
• One claim per paragraph

Structured data still does most of the work

Schema.org markup is the cheapest signal available and remains disproportionately effective, because it removes the parsing step entirely. The types that earn their place on a technical site:

The FAQPage block is the highest-leverage of these on an engineering blog. It hands the engine question-and-answer pairs that need no interpretation. Keep the visible copy and the markup identical, divergence between them is both a policy violation and, in our testing, a reliable way to lose the rich result.

llms.txt: worth adding, not worth arguing about

/llms.txt is a proposed convention: a plain-text map of your most useful pages, written for a model rather than a crawler. Adoption is partial and it is not a ranking factor. It costs about twenty minutes, it cannot hurt, and on a documentation-heavy site it gives a model a clean route to the pages you actually want quoted. We ship one on every build. We would not restructure a site around it.

What we measure, and what we ignore

There is no rank tracker for answer engines, and anyone selling you one is selling you a sampling methodology with a dashboard on it. What works:

1. Fix a panel of 30–50 prompts a real buyer would type.
2. Run them monthly against each engine and record cited / not cited.
3. Cross-check referral traffic by host in analytics, chat.openai.com, perplexity.ai, and friends show up as referrers.

Two things we deliberately ignore: raw impression counts from AI surfaces, which are not exposed reliably anywhere, and "AI visibility scores" from third-party tools, which are that same prompt panel with an opaque weighting applied.

The uncomfortable part

The pages that get cited are usually the ones that say something specific enough to be wrong. A page hedging every claim survives review and gets quoted by nobody. If your content team's incentive is to avoid being contradicted, GEO will not work, and no amount of markup fixes that.

References

• Schema.org. FAQPage, the vocabulary reference.
• Google Search Central, structured data general guidelines.
• OpenAI. GPTBot documentation, including the published IP ranges.
• llms.txt proposal, for the file format itself.