What we see getting cited in Google's AI Overviews across client accounts: answer position, specificity, structure and source type. Plus what does not work.
Getting cited in an AI Overview comes down to four things: answering the question in the first hundred words, being specific enough to be worth quoting, structuring the page so the answer can be lifted cleanly, and being the kind of source the system already trusts on that topic. Ranking well helps but does not decide it.
I have been pulling AI Overview citations across our client accounts for most of this year, and the gap between what people assume and what actually shows up is wide enough to be worth writing down. The most common assumption is that the top organic result gets cited. It often does not.
What we see getting cited
Pages that answer in the first hundred words
This is the strongest pattern and it is not close. Pages that commit to a direct answer near the top get quoted. Pages that build to the answer over four paragraphs of context get summarised anonymously, if they get used at all.
The practical version: whatever question the page targets, answer it plainly under the heading, before the narrative starts. It feels blunt to write and it reads better than most people expect, because it is what the reader wanted too.
Specificity that a general model cannot generate
An AI Overview does not need a source to tell it what SEO stands for. It needs a source for things it cannot produce itself: a number, a threshold, a named process, a constraint, a date, a price range, a specific consequence. Pages full of reasonable generalities are exactly the pages that get replaced by the summary, because the summary can generate the same generalities.
If you want to be cited, put something in the page that only you could have written.
Structure that makes the answer liftable
Short question-shaped headings with the answer directly underneath. Definitions as complete sentences rather than fragments spread across a paragraph. Lists where the content is genuinely a list. Tables where the content is genuinely a table. None of this is new advice. What has changed is that the cost of ignoring it went up.
Being an expected source for the topic
Systems reuse sources they have already treated as reliable on a subject. A site with genuine depth on one topic gets cited across that topic repeatedly, while a site with one good page on it rarely does. This is the slow part and it is the same work as building topical authority for conventional search, which is at least convenient. Our piece on authority building and off-page SEO covers the underlying programme.
What does not seem to work
Stuffing a page with question headings when the answers underneath are thin. The structure gets you considered and the substance decides it.
Schema on its own. Structured data helps machines understand what a page is and it does not make a weak page citable. We have seen no pattern suggesting schema alone earns citations.
Chasing position one. Useful, not sufficient. We regularly see citations going to results outside the top three while the top result is not cited at all, usually because the top result buries the answer.
Writing for the model rather than the reader. Pages that read as though they were assembled to be quoted tend to be generic, because generic is what optimising for a machine produces. The pages that get cited are the ones with a point of view and a specific claim.
How to check your own position
This takes an afternoon and is worth more than any general advice, including this article.
Write down the twenty questions your buyers actually ask before they buy. Not keywords. Questions, in their words, including the awkward ones about price and suitability. Run each one and record whether an AI Overview appears, who is cited, and whether the cited passage is something you could have written better.
Then look for the pattern in who does get cited. In most accounts we run this on, the cited sources fall into two groups: the obvious authority, and one specific page that answers unusually plainly. The second group is the one you can join this quarter.
Finally, segment AI referrals in your analytics if you have not already. The volumes are small and the trend line matters more than the absolute number. Without a baseline you will not be able to tell whether any of this worked.
Questions I get asked
Do AI Overviews reduce clicks?
For informational queries where the answer is short, yes, measurably. For commercial and considered queries the picture is more mixed, because the person still needs to choose a supplier and the overview mostly reframes that choice. The queries you should worry about are the ones where a complete answer removes the reason to visit. The queries worth investing in are the ones where the answer creates a shortlist.
Is AI Overview optimisation different from SEO?
Mostly it is SEO done with more discipline about answering plainly. The overlap is high enough that treating it as a separate workstream usually wastes money. Where it genuinely differs is in the emphasis on specificity and on being machine-readable without JavaScript. Our comparison of GEO against SEO sets out where the disciplines actually diverge.
Can I stop an AI Overview using my content?
There are controls, and using them is almost always the wrong trade. Blocking the systems that summarise you also removes you from the answers where buyers form their shortlist. The better response is to be the source that gets named rather than the source that gets absorbed.
How long does it take to get cited?
Faster than ranking, in our experience. We have seen pages picked up within weeks of a rewrite where the change was mainly moving the answer to the top and adding specifics. That is one of the more encouraging things about the current moment.
Does it help to add an FAQ section?
It helps when the questions are real and the answers are specific, and it does nothing when they are padding. A useful test: if the answer could appear on a competitor's site unchanged, it is padding.
What about ChatGPT and Perplexity rather than Google?
The same fundamentals apply and the sourcing behaviour differs. What has been most striking in our own Search Console data is how many queries now arrive as long conversational prompts generated by an assistant on someone's behalf, and how well specific, problem-shaped content performs against them compared with generic definitional content.
The short version
Answer first. Be specific enough to be worth quoting. Structure so the answer can be lifted. Build genuine depth on the topics you want to own. Then measure it, because most of what is written about AI search, including a fair amount of what I have read this year, is assertion rather than observation.
Our AI search optimisation guide covers the full programme, and the AI search optimisation service sets out how we run it. If you want us to run the twenty-question exercise against your own market, get in touch.