AI search quoted a passage, not a page.
Only 3 of 14 came from the first paragraph.

Where in a page does the passage an AI answer uses actually sit?

Shallow, but not at the very top. Across 14 usable observations measured on 9 September 2026, the median quoted passage sat 9.3% of the way through the source page’s body text, and 10 of the 14 fell inside the first quarter of the document. Seven of those landed in the first tenth. The tail is real though: one passage sat 66.7% down a pricing page and another 83.2% down a comparison post, so depth is a tendency and not a rule.

That gap matters more than the median does. If the rule were simply “put the answer at the top,” the first bar would be made almost entirely of opening paragraphs. It is not. Most of those shallow passages were the second, third, or fourth block on the page, sitting under a subhead, after an intro the engine ignored.

What did the quoted passages have in common?

A question above them. Nine of the 14 quoted passages sat immediately under a heading that was phrased as a question or opened with a question word: “What is a canonical tag?”, “How much does Marketing Hub Professional cost?”, “How ChatGPT decides what to cite”, “Is Airtable better than Notion?”. That held across depth. The passage 83.2% down a comparison post still sat under a question heading. The heading, not the position, is what the two deepest hits had in common with the shallow ones.

The five that did not have a question heading were mostly product roundups and pricing tables, where the heading was a brand name or a plan name. Those pages still got quoted, so a question heading is clearly not a requirement. It is just the single most common thing in the set, and it is the cheapest thing on this list to go and change. Google’s own documentation describes featured snippets as programmatically extracted from a page, which is the same extraction behavior at a different surface.

Worth naming the thing this does not license. Rewriting your copy “for AI” is not the move, and neither is spinning up a page per question. Google’s guidance on AI features in Search is explicit that its AI surfaces run on the same core ranking systems, so a page has to be indexed and snippet-eligible before any of this matters. We made the same point about the file everyone wanted to be a shortcut when we wrote that nothing is reading your llms.txt yet.

What exactly was measured, and what are the limits?

On 9 September 2026 we ran 20 buyer-intent and explainer queries through a synthesized web-answer surface, the WebSearch tool, which returns a written answer plus the source set it drew on. For each query we fetched every cited source, split the page into block-level paragraphs in document order, and scored each paragraph against the answer text using inverse-document-frequency-weighted token overlap. The highest-scoring paragraph across all of that query’s sources was recorded as the observation, and we logged three things about it: the midpoint of the paragraph as a percentage of the page’s total body text, whether it was the first paragraph, and whether the nearest heading above it was question-shaped.

Six of the 20 queries produced nothing usable and were logged as misses, so N is 14, not 20. In five cases no paragraph in any cited source cleared the 0.25 overlap threshold, which usually meant the answer had synthesized across several sources rather than leaning on one. In the sixth the pages were client-rendered and returned no parseable body text. Because the misses skew toward answers with no single dominant source, treat every number here as a floor on how concentrated passage retrieval is, not a rate.

Other caveats, plainly. This is one surface on one day, and it is a synthesis-plus-citation surface rather than ChatGPT, Perplexity, or Google AI Overviews measured directly. Overlap matching identifies the paragraph most likely to be the source of a claim; it does not prove the engine read that paragraph. Depth is measured against extracted body text, not rendered pixels, so a page with a long boilerplate footer would read as shallower than it looks in a browser. At N equals 14 a single reclassified observation moves a bin, which is why the shape of the distribution is the finding and the exact percentages are not.