What actually gets a page quoted in an AI Overview

Ask most people what it takes to get cited in an AI Overview and you'll get a list: rank well, add schema, publish long, structure your headings. Four answers, and the measured evidence supports roughly one of them. The deeper problem is that all four are answers about a page. An AI Overview does not cite a page in any meaningful sense. It lifts a sentence or two, attaches a link to the domain that sentence came from, and moves on to the next subtopic. The link is a receipt. The thing that earned it is a passage. That distinction sounds academic until you try to act on it, at which point it changes everything about what you'd bother doing.

Ranking buys the ticket. It stopped buying the seat.

Rank still matters, and anyone telling you otherwise is selling something. In a study of 362,000 US keywords that produced an AI Overview, seoClarity found that a page ranking first organically had a 43% chance of being cited in the Overview above it. Position two: 37%. Position three: 31%. By position twenty it was down to 7%. That is a clean, monotonic relationship, and it is the strongest single predictor anyone has published. But the share of citations coming from the top of the SERP has been falling, and falling fast. Ahrefs measured 4 million AI Overview citations in early 2026 and found 37.9% came from URLs in the first ten results. Their own study eight months earlier put that figure at 76%.

Take that trend with the caveat Ahrefs itself attached to it. Even so, the direction is corroborated elsewhere. In the same seoClarity dataset, the share of keywords where the Overview's citations overlapped completely with the top ten fell from 19% in May 2025 to 6% five months later. Whatever the exact number, a growing slice of every AI Overview is being filled from somewhere other than page one.

Three things people optimize that don't appear to move it

This is the part that tends to annoy people, so it's worth being precise about what was actually tested.

Schema: Ahrefs matched difference-in-differences, 1,885 treated pages against 4,000 controls. Word count: Ahrefs, Spearman correlation across 174,048 cited pages. llms.txt: Ahrefs, 38,000 domains publishing a valid file.

The schema result is the interesting one, because the correlation is real and enormous and completely misleading. Across six million URLs, pages cited by AI were nearly three times more likely to carry JSON-LD than pages that weren't. Fifty-three percent of cited pages had it. That looks like a finding. Then Ahrefs took 1,885 pages that added schema between August 2025 and March 2026, matched each against three control pages on other domains with similar prior citation levels, and measured the difference. AI Overview citations went down 4.6%. AI Mode and ChatGPT moved by amounts indistinguishable from zero. The raw AI Mode number before controls was +43%, which collapsed to +2.4% once you subtracted what the untouched pages did over the same period. Google's documentation says the same thing in plainer language: "There's also no special schema.org structured data that you need to add." Keep schema for rich results, which is what it was built for. Just don't expect it to buy you a citation.

What the cited passages have in common

The only study I'm aware of that codes the passages themselves rather than the pages is small, hand-done, and unusually honest about it. Advanced Web Ranking ran 40 prompts across two engines in July 2026, logged 265 citations, and hand-coded 112 passages — 95 that were quoted and 17 that were surfaced on the page but never cited.

Quoted95

Names the entity at first mention

Visible recent date

Answer in the first sentences

Pure consensus restatement

Self-contained. Dated. Answers before it explains.

Surfaced, not quoted17

Names the entity at first mention

Visible recent date

Answer in the first sentences

Pure consensus restatement

Undated. Says what everything else already says.

Two gaps are wide enough to be worth something even at this sample size. Quoted passages carried a visible recent date 80% of the time against 53% for the uncited ones. And the uncited passages were more likely to be pure restatement of what everyone else already says — 82% versus 61%. The second one is the counterintuitive result. Saying the consensus thing clearly makes your passage easy for a model to absorb. It does not give the model any reason to attribute it to you specifically, because forty other pages said the same thing and one of them said it first. Being correct is table stakes. Being the only place a particular sentence exists is what puts your name next to it. The same study found a page ranking first for the exact query, whose definition the Overview clearly tracked, that received no citation at all while lower-ranked pages did. That is what absorption without attribution looks like from the outside.

A large share of the citations were never going to be your article

It's worth knowing what you're competing against before you decide how hard to compete. Profound, analysing 27 million citations across the major engines, found that brand-owned domains account for 69% of AI Overview citations — not publishers, not forums. Media sits at 17%, institutions at 7%, social at 7%. In a separate analysis of 3.25 billion citations collected in March 2026, AI Overviews turned out to be the most aggressive user of social content of any engine measured, at 15.3% of citations. And within that social slice, it isn't Reddit doing the work. It's YouTube, at 38%, with Reddit second at 21%. Ahrefs found the same shape from a different angle: YouTube URLs were 5.6% of all AI Overview citations in their dataset, and 18.2% of the cited URLs that didn't rank in Google's top 100 at all. Which is to say a meaningful chunk of every Overview is being filled from a surface your blog post was never eligible for.

So what's actually worth doing

Short version. Rank, because it still raises your odds more than anything else measured. Write passages that survive being lifted — name the subject in the sentence rather than relying on the paragraph above it, put the answer before the explanation, and date the page honestly. Have at least one thing on the page that only you can say, whether that's a number you collected, a decision you made, or a case you actually worked. And stop spending time on markup and file formats that the engines have publicly said they ignore.

Sources

Related reading

• Not every prompt sends an AI to the web — before a page can be quoted, the question has to send the engine looking in the first place.
• Retrievable, extractable, citable, reusable — the four mechanics sitting underneath the structural patterns above.
• The citation is the conversion — why being quoted is the outcome worth optimizing for.
• The Prism Score — how AIO readiness is graded as one of ten dimensions.