How AI Overviews Choose What to Cite

A page can sit comfortably in the top ten for a query and never once be quoted in the generated answer above it. This is not a bug, and it is not a ranking penalty. Ranking and citation are two different selections, made by two different systems, optimising for two different things.

Ranking orders documents by relevance to a query. Citation picks a passage that can support a specific sentence in an answer that has already been drafted. The second job has requirements the first does not, and most pages fail them for structural reasons that have nothing to do with quality.

The unit of retrieval is a passage

The most consequential thing to understand is that the page is not the unit.

Before anything is generated, documents are split into passages and indexed individually. When the model needs support for a claim, it retrieves passages, not pages. A carefully built argument, where section four depends on the setup in sections one through three, arrives at the model as section four alone, stripped of everything that made it make sense.

This single fact explains most of the observed behaviour. A paragraph that reads perfectly in place can be incoherent in isolation, and an incoherent passage cannot be used to support a sentence.

	Ranking	Citation
Operates on	The whole document	An individual passage
Question asked	Is this page relevant to the query?	Does this text support this sentence?
Rewards	Topical coverage, authority, links	Self-containment, specificity, directness
Penalises	Thin or duplicated pages	Passages that need surrounding context

What survives extraction

The passages that get used share a shape, and it is not a subtle one.

They answer before they elaborate. A paragraph that opens with the answer and then qualifies it can be truncated anywhere and still be useful. One that builds to its conclusion in the final sentence is useless if cut, and it will be cut.

They carry their own referents. "This approach fails under load" is unusable in isolation, because this is gone. "Synchronous tool calls fail under load" survives.

They are specific enough to be checkable. Concrete claims with numbers, names and mechanisms give the model something to attach a sentence to. Hedged generalities give it nothing, which is why survey content that covers a topic broadly is frequently beaten for citation by a narrower page that says one definite thing.

This is the mechanism behind an otherwise confusing pattern: comprehensive guides rank, and specific posts get quoted. A broad survey of generative AI in SEO accumulates impressions across dozens of related queries precisely because it covers everything shallowly, and that same breadth is what leaves it without a passage worth lifting. Breadth is a ranking asset and a citation liability.

Why the top ten and the citation set diverge

Once you accept that these are different selections, the divergence stops being mysterious.

An answer is assembled from several sources, each supporting a different sentence. The engine is not looking for the ten best pages; it is looking for a handful of passages that between them cover the claims it wants to make. A page can be the best document on the topic and contribute nothing, because every claim it could support was already covered by a passage that was easier to lift.

There is also a consensus effect worth naming honestly. Passages that agree with other retrieved passages are safer to cite, so genuinely contrarian material is under-selected relative to its quality. If you are writing something that disagrees with the received view, expect it to rank before it gets quoted, if it gets quoted at all.

What does not work

Keyword density. It was already a weak ranking signal. For citation it is close to irrelevant, because the selection is made on whether a passage supports a sentence, not on term overlap.

Adding FAQ schema to everything. The structural benefit of a question followed by a short complete answer is real. The markup itself is not what earns the citation, and emitting FAQPage schema for content that is not actually on the page is a manual-action risk rather than a shortcut.

Writing for the model. Attempts to produce deliberately model-friendly prose, stripped of transitions and stuffed with declaratives, reliably produce text that reads badly and does not perform better. The properties that make a passage citable are the properties of clear writing. There is no separate register to learn.

What cannot currently be measured

Search Console does not tell you which impressions came from an AI Overview citation. There is no citation report, no field to filter on, and no way to separate a citation impression from an ordinary one.

So anyone quoting a precise citation rate is inferring it from third-party sampling of generated answers, not measuring traffic. Those samples are useful directionally and should not be treated as instrumentation. We do not have a reliable way to close this loop, and we would rather say so than model around a number we cannot verify.

The practical position

Structure is the lever, not style. Put the answer first in every section, write paragraphs that survive being lifted out of the page, and prefer one definite claim over five hedged ones. That is the whole intervention.

It is worth calibrating what it buys. A citation sits inside an answer that has often already satisfied the reader, so it converts worse than a comparable organic position — much the way a model's benchmark score predicts less about production behaviour than its headline suggests. Visibility and attribution are real. Treating a citation as equivalent to ranking third is not.