Pages get cited when a retrieval system can fetch them, extract a self-contained answer, and treat the source as credible. The first is an engineering property, the second is a writing property, and the third is mostly your existing search reputation. Everything beyond those three is currently speculation.
Key takeaways
- Retrieval crawlers are simple. Most do not execute JavaScript, so client-rendered content is often invisible to them.
- Extractability beats persuasion. A self-contained 40–60 word answer under a question heading is the single highest-leverage change.
- Your search reputation carries over. Answer engines lean heavily on existing search infrastructure; there is no separate ladder to climb.
- Specific, dated figures get quoted because they are exactly what a generated answer needs and cannot invent safely.
- Most published "AEO tactics" have no evidence behind them. Treat confident claims about ranking mechanics with suspicion, including some in this article's own field.
Why most writing on this is unreliable
Classic SEO accumulated two decades of large-scale correlation studies, public statements from search engineers, patent filings and leaked documentation. Answer engines have almost none of that. The retrieval pipelines are closed, they change without announcement, and results are personalised and non-deterministic — ask the same question twice and you can get different sources.
That combination makes rigorous measurement genuinely hard, which has not stopped an industry of confident advice appearing. When you read that some tactic "increases AI citation by 40%", ask what the sample was, how citation was measured across a non-deterministic system, and what the control was. The answer is usually that there was none of the three.
What actually happens when someone asks a question
The engine turns the question into one or more search queries, retrieves candidate documents, extracts passages from them, and generates an answer that cites the passages it used. Your page has to survive four separate stages, and most optimisation advice addresses only the last one.
- Query fan-out. The question becomes several conventional searches, often more specific than what the user typed.
- Retrieval. Candidate pages are gathered, in most cases through existing search infrastructure rather than a separate index.
- Passage extraction. Chunks of each page are pulled out and ranked for how well they answer the query on their own.
- Generation and citation. The answer is composed, and the passages that contributed are attributed.
Stage three is where most pages lose. A page can rank respectably and still contribute no usable passage, because its answer is spread across four paragraphs, depends on earlier context, or never states a conclusion plainly.
What is documented
| Claim | Status | Basis |
|---|---|---|
| Crawlers identify themselves by user-agent and respect robots.txt | Documented | Published by operators; verifiable in your own server logs |
| Most retrieval crawlers do not execute JavaScript | Documented | Operator documentation; testable by serving different content |
| Answer engines draw on conventional search indexes | Documented | Stated by several operators; visible in cited-source overlap |
| Self-contained passages are easier to extract and cite | Inferred | Follows directly from how passage retrieval works |
| Recency matters more for volatile topics | Inferred | Consistent with observed behaviour; no published mechanism |
| A specific word count improves citation odds | Unsupported | No evidence; recycled from classic SEO folklore |
| Adding "AI-friendly" markup increases citation | Unsupported | No published mechanism connects markup to passage selection |
What is worth doing anyway
The useful part: almost everything defensible here is also just good technical writing. None of it is wasted if the mechanism turns out to differ from what we assume.
Answer the question immediately, in one place
Put a self-contained answer of roughly 40 to 60 words directly beneath a heading phrased as the question. No pronouns pointing at earlier paragraphs, no "as discussed above". The passage has to make sense lifted out of the page entirely, because that is exactly what happens to it.
Serve real HTML
If the content only exists after hydration, a crawler that does not run JavaScript sees an empty document. This is an architecture decision, made early, usually by someone not thinking about visibility — see the rendering section of the pillar guide.
Publish specific, dated figures
Generated answers need concrete detail and cannot safely invent it. A page that states "$102.24 per month at 250,000 views, list pricing July 2026" gives a retrieval system something quotable in a way that "hosting can get expensive at scale" never will.
Make the source attributable
Consistent authorship, a real bio, and one entity referenced site-wide. This is the same E-E-A-T work classic SEO already asks for, which is convenient — you are not maintaining two strategies.
How to measure it
Imperfectly, and that is not a failure of method. There is no Search Console for citations. The workable approach combines referral traffic from answer engine domains, server-log analysis of AI crawler user-agents, and manually asking the engines your own target questions on a schedule.
- Referrals in analytics. Segment traffic arriving from answer engine domains. Small numbers, but they are real people who chose to click through.
- Server logs. Filter for AI crawler user-agents to see which of your pages are being fetched, and how often.
- Manual prompting. Keep a list of the questions you want to own, ask them monthly, record whether you appear. Crude, non-deterministic, and still the most direct signal available.
The short version
Serve real HTML, answer questions plainly and self-containedly, publish specifics with dates, and be a source a system can attribute. That is defensible whether or not the mechanisms are what we currently think they are — which is the only sensible way to optimise for a system nobody outside it can actually see.