How do you get cited by AI search?

The first thing to accept is that this is not the same contest as ranking, and treating it as one wastes most of the effort. Only around a tenth of sources cited in AI answers also appear in the organic top ten for the same query. That divergence is the whole strategic fact: the work that moves a ranking and the work that earns a citation overlap far less than anyone assumed, and a business optimising only for position is competing in a contest whose audience is shrinking.

What gets lifted is a passage rather than a page. Retrieval systems extract, and what they can extract cleanly is a short, complete, self-contained statement that answers the question without depending on the paragraph before it. A page that builds to its answer over four hundred words of introduction has hidden the only part that could be quoted. Putting the complete answer in the first sentences is not a stylistic preference; it is the difference between being quoted and being read past.

Attribution is the second lever and it is mechanical. A figure with a named publisher, a method and a date attached is safer for a generated answer to use than the same figure floating free, because the answer can carry the attribution with it. Direct quotations are the highest-measured lever of this kind, and the reason is the same: a quotation arrives pre-attributed. Publishing a sentence written to be lifted verbatim, with a name behind it, is a cheap and specific thing to do that most pages never do.

The third is that the material has to be genuinely unavailable elsewhere, and this is where most attempts fail. If your page restates what forty other pages say, a generated answer has no reason to reach for yours in particular, and redundancy suppression means the version it does reach for is one of the others. Original measurement, a reconciliation of sources that disagree, a provenance trail showing where a circulating claim actually came from, a method somebody else can run — these are things a model cannot synthesise from its training data, which is precisely what makes them worth citing.

There is a defensive half to this that matters as much as the offensive one. Publishing large volumes of near-identical pages is now actively harmful — scaled interchangeable page sets lost a very large share of their rankings in the March 2026 core update — and the same interchangeability that triggers that also makes a page useless to a retrieval system, for the same underlying reason. There is no version of this strategy where volume substitutes for distinctiveness.

The practical shape, then, is short and unglamorous: answer in the opening, attach a source and a date to every figure, include at least one sentence written to be quoted, publish something on the page that exists nowhere else, and link related material so a retrieval system can follow the entity rather than guess at it. None of that requires new technology, and almost none of it is what a page optimised for ranking alone would look like.