Google AI Mode isn't citing your page — it's citing one paragraph of it, and 80.9% of them are used exactly once

The numbers, from the study itself

The primary source is Pillarbase Research's published study, *The Text-Fragment Phenomenon in Google AI Mode*; Search Engine Land's write-up ran on July 29, 2026. The headline counts:

And the shape of the passages that get reused: a median length of 117 words, 80% opening with the answer rather than building to it, and 83% self-contained — readable with no surrounding context. Question-led passages are reused roughly twice as often as others (48% of high-reuse passages are question-led versus 22% of single-use ones). Pages that accumulate 21 or more distinct highlighted variants have a median rank of #2, and Pillarbase puts the share of heavily recycled passages coming from #1-ranking results at 59–76%.

One note in the interest of not laundering numbers: Search Engine Land's summary reports 85% self-contained and a median rank of #1 where the study page we read says 83% and #2. We have used the figures on the study page throughout, and flag the discrepancy rather than picking whichever is more dramatic.

Why the unit change matters

Classic SEO optimises a page: one URL, one primary keyword, one ranking position. A text-fragment citation does not work that way. AI Mode selects a span of text, links to it directly, and — in the reuse tail — carries that same span into hundreds of different answers. The 661-reuse passage answered 483 distinct queries, which is a level of leverage no page-level metric can even represent. Meanwhile the long tail is brutal: 80.9% of passages are used exactly once, so most of what gets highlighted is a one-off, and volume of highlights is not by itself a signal of anything.

The rank correlation is the part that should stop anyone hoping AI Mode is a way around ranking. If the majority of heavily recycled passages come from pages already sitting at #1, then passage-level optimisation is a multiplier on organic position, not a substitute for it. You do not write your way into the reuse tail from page four.

What to actually change in the copy

• Answer in the first sentence of the paragraph, then justify. 80% of highlighted passages open with the answer. A paragraph that spends two sentences setting up context has already lost the extraction — the model needs a span that stands alone, and the span it can lift starts with the answer.
• Write paragraphs that survive being torn out. 83% of highlights are self-contained. That means no "as we saw above", no pronouns pointing at the previous paragraph, no numbers whose units live in the heading. Every paragraph you want quoted should re-state its own subject.
• Use question-led subheads and answer them immediately. Question-led passages appear at roughly twice the rate in the high-reuse set as in the single-use set. This is also the long-tail-question shape buyers actually type.
• Aim near 117 words per answer block. That is the median highlighted length. Shorter than that and there is nothing to lift; much longer and the answer is buried in the middle of the span.

The takeaway

Pillarbase's dataset says the smallest useful unit of AI-search work is now the paragraph, and that the paragraphs that travel are answer-first, self-contained, question-led and about 117 words long. That is a writing instruction, not a technical one. What it will not tell you is whether any of it moved the outcome for a specific business — for that you have to ask the engines the buyer's question and count. The free 60-second check does exactly that across ChatGPT and Perplexity, and how to show up in ChatGPT recommendations covers the same discipline on the engine that sends the most referral traffic.