Google’s AI Overviews do not simply summarize the top ranking page. They assemble an answer from several sources retrieved for the specific question, and the pages that get cited are consistently the ones that answer a question directly, early, and in a structure a model can extract cleanly. Ranking first helps. It does not guarantee citation, and pages ranking eighth get cited regularly.
That gap between ranking and being cited is the whole opportunity, and most businesses have not noticed it exists.
What retrieval actually favors
A direct answer in the first hundred words
If a page spends four paragraphs establishing context before answering the question in its own title, a model has nothing clean to lift. Pages that state the answer plainly and then elaborate get quoted. Pages that build to a conclusion do not.
This runs against a lot of conventional content advice, which encourages narrative and delayed payoff to keep people reading. For AI retrieval, the opposite is true, and the good news is that human readers prefer the direct version too.
Self contained sections
A heading followed by content that makes sense without the surrounding page is extractable. A section that depends on three paragraphs above it is not. Writing each section to stand alone is the single most useful structural habit.
Specificity that other sources do not have
Models synthesize across sources. A page that repeats the consensus adds nothing to the synthesis and gets skipped in favor of whichever source said it first or best. A page with an actual number, a firsthand account, or a stated position gives the model something only that page can supply. That is the mechanism behind citation.
Content that exists without JavaScript
This one silently disqualifies a large number of otherwise good pages. Most AI crawlers do not execute JavaScript, so if your content is injected by a script, those systems see an empty page. It does not matter how good the writing is if it is not there when they look.
An entity the model can resolve
Clear structured data, consistent naming, and an unambiguous statement of what your business is and where it operates. Models are cautious about citing sources they cannot identify. A connected schema graph does more work here than most people expect and it is standard technical SEO practice regardless.
What appears to matter less than people assume
Word count, on its own. Long pages get cited and short pages get cited. What correlates is whether the answer is findable, not how much surrounds it.
Keyword density, which was already a dead concept for classic search and is more so here. Models work on meaning, not repetition, and a page that repeats a phrase unnaturally reads worse to both audiences.
Publication recency, except for questions that are actually time sensitive. An evergreen explanation from two years ago gets cited over a fresher page that says less.
How to check where you stand
Search your main questions and see what the overview says and who it cites. Do it for the ten questions your customers actually ask, not the ten keywords in your rank tracker, because those are different lists.
Record the answers. This is the part people skip, and it is the only way to know later whether anything you did made a difference. The same approach applies across the other assistants, and the results differ enough between them to be worth testing separately.
The honest caveat
This is a fast moving area and anyone claiming settled certainty about how AI Overviews select sources is overstating what is known. Google has not published the mechanism, results shift between tests, and the same query can produce different citations on consecutive days.
What is defensible is that the underlying practices, which are direct answers, clean structure, server rendered content, unambiguous entities and genuinely distinctive material, are all things that improve a page for classic search and for human readers regardless of what any model does next. That is why the work is worth doing now: it is not a bet on one system.
Questions
Does appearing in an AI Overview reduce my traffic?
Sometimes, for purely informational queries where the answer resolves the need entirely. For queries with commercial intent, being cited tends to be worth more than the click it might replace, because it positions you as the source before anyone compares options.
Can I opt out?
There are controls for excluding your content from certain AI features. For almost every small business this is the wrong move, because it removes you from the consideration set entirely while your competitors remain in it.
Is this different from normal SEO?
It overlaps substantially, but the outputs are measured separately because they behave separately. The differences are worth treating as their own discipline.
