Getting cited by AI answer engines: what actually seems to matter

People now ask assistants questions they would previously have typed into a search box. The output is an answer with a handful of citations rather than ten links. Being one of those citations is a different problem from ranking, and the advice around it is full of confident claims that nobody can currently support.

Here is what is reasonably well established, and where the honest uncertainty is.

The mechanics, briefly

Most assistants answering a factual question do some version of: reformulate the question into searches, retrieve results from an underlying index, fetch some pages, and synthesise an answer with citations.

Two things follow directly from that, and they are the least speculative advice available:

1. You must be in the underlying index. Several assistants use Bing's. Being absent or stale there means you are not a candidate at all.
2. Your page must be fetchable and readable. If your content only exists after client-side rendering, many AI crawlers will not see it.

Almost everything else is downstream of these two.

What appears to help

Answering the question directly, early

A model extracting an answer needs a passage that is the answer. A page that opens with three paragraphs of throat-clearing before reaching the point is harder to quote than one that states the answer in the first sentence and then explains it.

This is the inverted pyramid, and it is genuinely more useful for AI extraction than for human readers, who will scroll.

Structure that maps to questions

Headings phrased as the questions people actually ask, each followed by a self-contained answer, produce passages that survive being lifted out of context. Rambling narrative sections do not.

Specific, checkable facts

Numbers, dates, named limits, explicit conditions. "Submissions are capped at 10,000 URLs per request" is citable. "We handle large volumes efficiently" is not — there is nothing in it to quote.

Being unambiguous about what you are

Clear statements of what you do, who you serve and where. A model synthesising an answer needs to be confident your page is about the thing being asked.

Being current

Assistants favour recent sources for anything time-sensitive, and many display a date. A page last updated in 2021 is less likely to be chosen for a question about how something works now — and this is where indexing freshness feeds back into citation.

What is unproven

Stated plainly, because the field is full of confident nonsense:

• llms.txt improving citation rates. Plausible, no evidence. Adoption is early. Worth publishing because it is nearly free, not because anyone can show it works.
• Schema markup improving AI citation. Structured data helps traditional search features. Whether it influences LLM retrieval is unestablished.
• "GEO" as a distinct discipline. Most of what is sold under that label is ordinary content quality advice with new vocabulary.
• Specific word counts, keyword densities or formatting tricks. No.

If someone quotes you a percentage improvement from an "AI optimisation" technique, ask what the control group was.

A realistic priority order

1. Be indexed and current in Bing. Foundational, measurable, and directly actionable — this is the concrete step most people skip.
2. Serve content in the initial HTML. Do not make retrieval depend on JavaScript execution.
3. Allow retrieval crawlers in robots.txt, whatever you decide about training crawlers.
4. Restructure key pages around real questions, answering each directly and early.
5. Add specifics. Replace vague claims with facts a model can quote.
6. Keep it updated, and make sure the update is noticed.
7. Publish an llms.txt. Cheap, harmless, possibly useful.

The uncomfortable part

Being cited does not reliably produce a visit. The assistant answers the question; the user may never click. For informational queries this is a genuine and unsolved threat to traffic.

The content that survives it tends to be the kind that cannot be fully summarised away: tools people need to use, data they need to explore, services they need to buy, communities they need to join. Content whose entire value is a paragraph of explanation is, increasingly, content the assistant simply delivers itself.