How to appear in AI-generated search answers

A category of advice has appeared over the last two years promising to optimise your site for AI answers — the summaries that appear above search results, and the responses assistants give when someone asks a question instead of searching for one.

Some of it is reasonable. A great deal of it is invented, because the honest position is that nobody outside these companies knows the ranking mechanics, and an industry cannot sell "nobody knows" as a service.

This is what is actually established, what follows reasonably from it, and what is being sold that has no basis. It is a spoke of AI and SEO.

What is actually known

Four things, stated at the level of confidence they deserve.

AI answers cite sources, and citation drives clicks. When an answer names and links its sources, some readers follow them. Being cited is therefore worth something, in a way that being absent is not.

For Google's AI features, the underlying index is the same index. Google's own documentation on AI features says the same fundamentals apply: pages need to be crawlable and indexable, and the guidance for appearing in AI experiences is the guidance for appearing in search. There is no separate submission process, no distinct markup, no second index to get into.

Assistants that browse are constrained by what they can fetch. A model retrieving pages at query time can only cite what it can reach and read. A page blocked from the relevant crawler, or one whose content only appears after heavy client-side rendering, is a page that cannot be quoted.

Models cite what is quotable. This is an observation about their behaviour rather than a documented rule, but it is consistent: a clear, self-contained, factual passage that answers the question directly is easier to lift and attribute than the same information dissolved across a thousand words of preamble.

That is the reliable set. Everything else is inference.

What follows reasonably

Given the above, a short list of things that are defensible.

Answer the question early and plainly. If a page's answer is stated in two clear sentences near the top, it can be extracted. If it is implied over six paragraphs, it cannot. This is also better for readers, which is the reason to do it that survives regardless of what the models do next.

Be specific and attributable. Vague claims are not worth citing. Concrete ones — a number, a mechanism, a named condition — are. "SEO takes time" cites nothing; "on a new domain with no external links, expect impressions before clicks and clicks before conversions" says something a model can attribute to you.

Structure the page honestly. Headings that state what the section contains, questions asked as questions, tables where the content is genuinely tabular. Not as a trick — because it makes the content's structure legible to anything reading it, human or otherwise.

Cover the sub-questions. AI answers are often assembled from several sources covering different parts of a question. A page that answers the whole question, including the awkward parts everyone else omits, has more surfaces on which to be cited.

Say things nobody else says. A page that restates the consensus adds nothing to a summary of the consensus. A page with first-hand knowledge, original data from your own operations, or a genuinely different position is the only kind that has to be cited rather than optionally cited.

Keep the site crawlable. The technical baseline is unchanged and is in technical SEO for people who are not developers.

Notice what that list is. It is, almost exactly, the advice for writing a page that deserves to rank. That is not a coincidence and it is the most useful conclusion here: the work does not fork.

What is being sold that has no basis

Said plainly, because the cost of following it is real.

"AEO/GEO optimisation" as a distinct discipline with its own techniques. The named techniques are, in every version I have seen, existing content and technical practice relabelled. If a service cannot say what it does differently from good content work, it does not do anything differently.

Specific formatting rules presented as known. Claims that a certain sentence length, a particular paragraph structure, or a specific proportion of lists increases citation rate. These are not published by anyone who would know, and the ones circulating are pattern-matching on small samples.

Guaranteed citation. Nobody can guarantee this. There is no submission, no placement, no relationship that produces it.

"Optimise for LLM training data." Training corpora are assembled from large crawls at intervals; you cannot meaningfully target them, and a model's training data is not what most cited answers come from anyway — retrieval at query time is.

Schema markup as an AI ranking lever. Structured data helps machines parse a page and can affect eligibility for enhanced search displays. Presenting it as the route into AI answers overstates what it is known to do.

Volume as a strategy. Publishing more pages to increase citation surface produces the same result it produces in search: a lot of pages nothing cites.

What being cited is actually worth

Worth putting numbers-shaped thinking against, since the decision is a budget decision.

Some cited answers produce no visit at all. The reader gets what they needed from the summary and stops. This is not new — featured snippets have done it for years — but it is more pronounced when the answer is synthesised from several sources rather than lifted from one.

The visits it does produce are unusually well qualified. Someone who reads a summary, sees you cited, and clicks through has chosen to go deeper on the specific thing you were cited for. That is a better visitor than a position-three click on a broad term.

Citation carries a credibility effect that is hard to measure and probably real. Being named as the source, even without a click, is the same category of value as being quoted in an article somebody reads.

So the honest summary: fewer visits than a ranking, better ones, plus an unquantifiable reputational component. That is worth having and it is not worth reorganising a strategy around.

Where this genuinely changes how you write

Three specific habits that are worth adopting, distinct from general good practice.

Make each section answer one question completely, in place. A section that only makes sense after reading the previous three cannot be extracted. This is a mild constraint, and it happens to make pages easier to skim, which readers were already doing.

Put the number, the range or the condition in the sentence, not in a chart. "Most sites see this within three to six months" can be quoted. A graph cannot, and neither can "it varies considerably".

State your position explicitly rather than implying it. Writing that hedges its way to a conclusion is hard to attribute a view to. If your answer is "usually no, for these two reasons", say that in those words somewhere on the page.

None of these are tricks and none of them are new. They are what clear writing looks like, which is the recurring conclusion of this whole subject.

Whether this is worth your time

An honest cost-benefit, because the answer is not "yes, urgently" for everyone.

The traffic is smaller than the discourse. AI answers frequently satisfy the question without a click. Being cited is real and it is not equivalent to a ranking.

The work is not additional. If the actions are the same actions that make a page rank, then you are not choosing between two programmes. You are doing one and being eligible for both.

The landscape changes faster than anyone can optimise for it. Interfaces, citation behaviour and prominence have all shifted repeatedly. Building a strategy on the specifics of how one product behaved last quarter is building on sand. Building on "be the clearest, most specific, most genuinely knowledgeable page on this question" survives the changes.

So: worth understanding, worth letting influence how you write, not worth a separate budget line or a specialist retainer.

What to actually do

Six things, in order, none of which are new:

1. Make sure the page is crawlable and indexed. Nothing can cite what it cannot fetch.
2. State the answer to the page's question in the first hundred words, plainly, in a passage that stands alone.
3. Put the specifics in. Numbers, conditions, mechanisms, the thing you know from doing the work.
4. Use headings that state what the section answers.
5. Cover the whole question, including the parts other pages skip.
6. Say something only you can say. First-hand experience, your own data, an honest position others avoid.

Then check occasionally whether you appear — ask the question in the assistants your customers plausibly use and see who is cited. That is the only measurement available, it is crude, and it is more informative than any tool currently claiming to measure it.

The part that has not changed

Every route to being cited runs through being worth citing. There is no configuration, no markup, and no service that substitutes for having written the page that actually answers the question better than the alternatives.

Which is the same constraint as search, with the same consequence: if your subject's territory is held by sites with more authority, more history and better coverage, being marginally more quotable does not displace them. Where that is the honest reading and the timeline matters, buying visibility directly remains the alternative — the trade is in when renting attention beats owning it, with current prices on the rankings board.

Common questions

Does blocking AI crawlers hurt my search rankings? Google documents separate controls for some AI uses and its search index. Blocking broadly can remove you from AI answers; read the specific controls before applying them rather than copying someone's robots.txt.

Should I add an FAQ section to every page? Only where readers genuinely ask those questions. An invented FAQ block is padding, and it reads as padding.

Do AI answers replace SEO? They change how some questions get answered without a click. The work of being the page worth citing is unchanged.

Is there a tool that tracks my citations? Several claim to. Treat the numbers as indicative rather than measured — none of them has access to the systems producing the answers.