How does AI decide what to cite?
There are two loud camps on this, and both are selling you a third of the truth. One says nothing has changed — the AI just reads Google, so keep doing SEO and you will be fine. The other says everything has changed — it is a black box now, an unknowable game of "narrative" and "brand", and your rankings are irrelevant. I have spent years being paid to know which marketing levers actually move a number, and the honest answer is that neither camp has looked closely enough at what the machine actually does.
A source gets into an AI answer through one of exactly three doors. Only one of them is rank, which is why the "just do SEO" camp is a third right. The other two are real, measurable and largely ignored — which is why the "it is a black box" camp gets to sound clever. We measured how much each door is worth on a live consumer market. Here they are, widest first.
Door one: rank — the front door
The strongest single predictor of whether an AI answer cites you is still where you rank on the underlying search. It is not close.
In a UK health market we measured across thousands of questions, a page sitting in the top three organic results was pulled into the AI Overview about 67% of the time. A page ranking fourth to tenth, about 41%. A page ranking eleventh or worse, about 16%. The curve is steep and it points one way: climb the ranking and your odds of being cited climb with it.
The logic is not mysterious. The model assembles its answer from sources it retrieves, and the cheapest, most trusted pool to retrieve from is the one the search engine has already ranked. Your rank is the model's shortlist. This is the whole basis of the coupling between classic SEO and getting cited by AI — and it is why "SEO is dead" is a headline, not a finding.
But a door is not a house. Two-thirds cited at the top is also a third not cited at the top, and — the part that matters — a large share of what the AI cites did not come through this door at all.
Door two: own the follow-up question
The second way in is to own the answer to the question after the question.
Search results have long carried a block of follow-up questions — the "People Also Ask" list — each with a short, structured answer lifted from a real page. Owning that answer is a distinct asset from ranking for the main term, and it pays a distinct dividend: in our capture sets, holding the People Also Ask answer lifts a source's odds of being cited in the AI answer by roughly 1.3 times for a generic question, and up to 2.4 times for a branded one. As far as we can find, no one else has published that number — it is ours.
The logic: the follow-up answer is already a machine-readable, search-vetted answer to a precise sub-question. When the model builds its response, that is exactly the shape of thing it wants to quote. Ranking gets you onto the shortlist; owning the follow-up answer hands the model a sentence it can use verbatim.
One honest wrinkle, because measuring it honestly is the whole point: the follow-up box is itself turning into a little AI answer. Through 2026 the live version increasingly renders as a generated summary with no single page to "own", so that lift is measured from historical captures and stamped with its era. The asset is shifting under everyone's feet — which is exactly the argument for measuring it rather than assuming it.
Most brands never contest this door. They optimise the headline term and ignore the twenty follow-up questions underneath it — which is where a surprising amount of the answer is actually built.
Door three: the topic authority the model reaches for
The third door is the one the "black box" camp is really pointing at, and it is real. The model does not only retrieve from the ranked results — it reaches past them for sources it treats as authorities on the topic, not the query.
We can size it. Just over a third of the citations under an AI answer — about 37% in our health capture sets — go to sources that rank nowhere for that exact question. The model went and got them anyway. It retrieved at the level of the subject, not the search box.
And here is what comes through that back door in consumer health, which should stop any marketer cold: it is overwhelmingly the community. In the market we measured, Reddit was cited on 63% of the questions where the AI answered at all — most of them without ranking for the question in any normal sense. Facebook, Quora, Mumsnet, YouTube and the patient forums follow it in. The model, asked a health question, reaches past the clinics and the medical databases and pulls in what actual people said to each other.
The logic is that a discussion thread is dense with the lived, specific, first-person detail a good answer needs — and the model has learned to trust it for exactly the questions where an official page is thin. You cannot "rank" your way through this door. You earn it by being genuinely present, and credible, where your buyers actually talk.
So how do you get cited by AI?
Put the three doors together and the brief writes itself — and it is not "rank higher", though that is part of it.
- Earn the shortlist. Rank is door one and the widest, so classic search visibility is the foundation of the whole thing. If you are not ranked and not retrievable, you cannot be selected. This is the half the "SEO is dead" crowd gets wrong.
- Own the follow-up. Contest the People Also Ask answers in your category, not just the head term. It is a separate, under-fought asset that pays a measurable citation dividend.
- Be where the model reaches. Build real topic authority off your own domain — and in consumer categories, take the community seriously, because that is the door the model is quietly using for a third of what it says.
None of that is a black box, and none of it is "just SEO". It is three levers, each measurable, each with a different owner inside most marketing teams — which is precisely why most teams pull only one of them.
Some take-aways
Stop asking "how do I get cited by AI" as if there were one answer. Ask which of the three doors you are actually working. Almost everyone is working the first — rank — and ignoring the two that account for a large and growing share of what the answer is built from.
There is one more thing the three doors do not tell you, and it is the one that decides where to spend: the doors are not the same width on every surface. The community door that is thrown wide open on Google's AI Overview is far narrower inside the chatbots, where authority and reviews walk in instead. Which is a measurement in its own right — we ran the numbers on whether SEO still matters for AI, and the gap between your search share and your answer share turns out to flip sign depending on the engine.
For the framework this sits inside, see the AI-visibility ladder; for the whole-market view, what market intelligence actually is, or explore the platform.
Method note: figures are Theia measurements across live consumer markets — a UK private-healthcare market and a global camera-and-imaging brand's category — captured with Google's asynchronous answer layer requested explicitly and reported as rates over repeated captures. Surface-level results are never pooled. We publish the result and the reasoning, never the method that produces them.
Frequently asked
- How do AI Overviews decide what to cite?
- Through one of three routes. A source is cited because it ranks well on the underlying search (the strongest single predictor — from positions 1-3 it is cited about two-thirds of the time), because it owns the People Also Ask answer for a follow-up question, or because the model reaches past the search results for a topic authority it trusts. Just over a third of AI citations, in the markets we measure, go to sources that rank nowhere for the exact question.
- Does ranking on Google get you cited by AI?
- It is the biggest lever, but not a guarantee. In a UK health market we measured, a page ranking in the top three was cited by the AI Overview about 67% of the time; a page ranking 11th or worse, about 16%. Rank strongly predicts citation — but two of the three ways into an answer are not rank, so 'rank higher' is roughly a third of the job.
- How do I get cited by AI if I can't rank first?
- Use the other two doors. Own the People Also Ask answer for the follow-up questions in your category — in the UK health market we measured, holding that answer lifted citation odds up to 2.4 times for a brand term. And build genuine topic authority where the model reaches for it, which in consumer-health answers is very often the community: forums and discussion threads are cited far more than their search rank would predict.