Optimising for AI search: getting into the answer, not the top ten

Classic SEO fought for a position in a list. An assistant shows no list — it gives an answer and names two or three sources. Or does not name you at all.

This is a different problem, and the old habit of "optimising a page for a query" does not survive it: there is no query in the familiar sense. There is a question asked in conversational language, and a generated answer nobody compares against alternatives.

What changes the rules

Zero clicks. The person got their answer and went nowhere. Your traffic drops while your rankings hold, and analytics reports it as "SEO getting worse" even though nothing moved.

A citation instead of a visit. The value is no longer the click, it is being named. That is recognition, and it converts later — when someone searches for you by name.

Answers are assembled from fragments. The model takes specific paragraphs from different sources, not "a page". The unit of optimisation has moved from the page to the paragraph.

What decides whether you get cited

These are not guesses — this is what you see when you take apart a dozen answers and check where each sentence came from.

The answer in the first paragraph. A section that opens with a run-up ("in today's world, businesses increasingly…") never gets cited. The one whose first sentence already contains the answer does.

Concrete numbers with a date. "Cheap" is not quotable. "$4 per ad on the $149 plan, as of August 2026" is, because it can be checked.

Structure shaped like a question. A heading phrased as a question, and under it an answer of 40–60 words. That is literally the shape the model looks for.

A source worth pointing at. Your own measurements, your own data, your own experience. A retelling of somebody else's article does not get cited — that article does.

Freshness. A date in the markup and in the text. Models prefer recent where the topic calls for it.

What not to do

Keywords. The model works on meaning, not occurrences. Stuffing phrases adds nothing and ruins the text for humans.

Hiding the answer at the end to boost time on page. A technique that worked against behavioural signals now works against you: the model will not get that far.

Writing "for AI". Text scrubbed into blandness engages nobody — not a person, not a model — because there is no value left in it.

The practical minimum

1. Every section opens with the answer. Explanation comes after it.
2. One checkable number per section. With a date.
3. Headings as questions people actually ask.
4. Article + FAQPage markup, published and updated dates.
5. Once a quarter, ask an assistant about your topic and see who it named.

That is your new results page.

And about social

Assistants cite social posts less often than sites, but they do lean on them where freshness matters more than depth. The practical consequence is simple: a post that answers a question specifically, with a number, has a chance. A post saying "we're open, come by" has none — and not only with an assistant.