What gets quoted is a paragraph, not a page
Nearly half of Google's AI citations link to a specific sentence. Here is what those sentences have in common.
Pillarbase, July 2026 · 15.7 million AI Mode citations · June 2, 2026 · 4 min read
For twenty years, the unit of optimisation was the page. You wrote a page, it ranked, someone clicked it. Everything about how sites are built assumes that — the title, the meta description, the internal links, the whole apparatus points at a page as the thing being chosen.
That assumption has quietly stopped holding.
In July 2026 Pillarbase analysed 15.7 million AI Mode citations across 148 industries and resolved them to 4.6 million unique passages sitting on 2.7 million pages. The first number they published is the one that reframes everything else: 47.7% of citations used a scroll-to-text fragment rather than a plain link. Google is not sending a reader to your page. It is sending them to one highlighted sentence on it.
Which raises the obvious question — what makes a sentence the one that gets picked?
The dataset answers it with unusual clarity. Among the passages that get reused across many different questions:
Median length 117 words
Answer stated in the first sentence 80.9%
Fully self-contained, no dependency on context ~85%
Question-led passages, recycled 48% vs 22% for others
The single most-recycled passage in the entire dataset appeared 661 times across 483 distinct queries. One paragraph, answering hundreds of unrelated questions, for months.
Read those four lines again, because between them they describe a way of writing that almost nobody uses.
A hundred and seventeen words is a full paragraph. Not a snippet, not a bullet, not a one-line definition. The retrieval unit is a substantial chunk of prose — long enough to contain a complete thought and short enough to drop into an answer without editing.
The answer comes first. Four in five of the reused passages state their conclusion in the opening sentence. This is the reverse of how almost every business writes about itself. The standard service page builds: context, then credentials, then the offer, then finally what it actually does. A model reading that page finds nothing usable until paragraph six, and by then it has moved on.
It stands alone. Around 85% of recycled passages have no dependency on what came before them. No “as we mentioned above”. No “this approach”. No pronoun whose referent is two paragraphs back. If a paragraph only makes sense in sequence, it cannot be lifted, and a paragraph that cannot be lifted cannot be quoted.
It leads with the question. Passages that open by naming the question they answer get recycled roughly twice as often.
That last finding needs a careful distinction, because it looks like it contradicts something else in the literature. A separate controlled study measured pages formatted as FAQs and found a −5.74% relative difference — slightly negative. So question-led passages help and FAQ formatting doesn’t?
Both are true, and the difference is the thing worth understanding. A paragraph that opens by stating the question it answers, in ordinary prose, reads like a person explaining something. An FAQ block is a template: a list of collapsed rows, each with a question in bold and two lines under it, usually written to hit keywords rather than to answer anything. The first is writing. The second is furniture. Models appear to reward the first and be indifferent to the second.
Now the part that keeps this honest. Passage engineering does not replace ranking.
The same dataset shows pages with 21 or more highlighted passages had a median rank of #1, and among passages cited more than a hundred times, 76% came from pages ranking first. The author is explicit that this is correlation and not a proven mechanism — but the direction is unambiguous. You do not write your way into an AI answer from page four of Google. What good passage construction does is convert a page that already surfaces into a page that gets used.
There is one more measurement worth putting alongside this. A separate study of 21,143 citations found that ChatGPT cites a mean of 6.88 sources per answer while Perplexity cites 16.35 — but the mean influence of a fetched page on the final answer was 0.27 for ChatGPT against 0.065 for Perplexity. ChatGPT cites narrowly and absorbs deeply; Perplexity cites broadly and shallowly. Which means a citation count is not a visibility score, and the pages that get genuinely used are a much smaller set than the pages that get listed.
