64% of the Passages Google's AI Overviews Quote Start at a Heading or Right Below One, Against 31% of Other Passages on the Same Page (Study of 3,224 Quotes)

Finding 0164% of the passages AI Overviews quoted start at a heading or right below one

The passage Google's AI Overviews quoted started at a heading, or in the first block under a heading, 64% of the time. Other passages of the same length on the same page did 31% of the time.

Counting only starts at the heading itself: 40% of the quoted passages in AI Overviews, against 16% of other passages. Of all the quotes found, 40% started in a heading, 40% in a paragraph and 13% in a list item. A quote in AI Overviews ran across a median of two blocks: a heading and the text below it, or a sentence and the next.

Finding 02The passage AI Overviews quoted held 43% of the searcher's words, against 25% for the rest of the page

I counted how many of the question's words (leaving out words such as "how", "to" and "the") appear in the quoted passage, and in each of the other passages. The passage Google's AI Overviews quoted held 43% of them on average; other passages on the same page held 25%. The quoted passage held more of the question's words than the rest of its page for 2,295 of 3,243 quotes, and fewer for 880.

The heading above an AI Overview quote carries the question's words too: 33% of them, against 22% for the headings above other passages.

Finding 0364% of the quotes in AI Overviews start in the first 30% of the page

Measured as the share of the page's main text above the quote, the median quote in Google's AI Overviews started 19% of the way down, 284 words in. 33% of the quotes in AI Overviews started in the first tenth of the page, 64% in the first 30% (62% to 66%) and 18% in the second half.

Answers sit higher for the SEO agency searches: 69% of their quotes in AI Overviews started in the first 30%, against 61% for the how-to questions, where pages often open with context before the steps.

Finding 049.8% of the passages AI Overviews quoted sat under a question-style heading, against 10.3% of other passages

Headings written as questions ("How do I ...?") did not set the quoted passage apart: 9.8% of the passages Google's AI Overviews quoted sat under one, against 10.3% of other passages on the same page. Lists were less common where the quote started: 13% of the quoted passages in AI Overviews started in a list, against 23% of other passages.

That is not the same as saying question headings or lists never help: a page may get cited for other reasons, and this study compares passages within pages that were already cited.

Finding 0538% of the source cards in AI Overviews were lists Google made from the page

Most source cards in Google's AI Overviews were excerpts, the page's own words in order: 59%. 38% were lists Google made from the page, an opening line followed by items, and 3% were tables that show some of the page's columns.

In list cards, the opening line was Google's own words in 23% of the list cards in AI Overviews whose opening line could be checked; the items themselves were the page's words.

Finding 0662% of the quotes that start inside a paragraph of three or more sentences start at its first sentence, against 28% by chance

Kevin Indig found the opposite for ChatGPT: 53% of citations came from the middle of a paragraph, 24.5% from its first sentence and 22.5% from its last. Google's AI Overviews start at the top of the paragraph. Of the 479 quotes in AI Overviews that started inside a paragraph of three or more sentences, 62% started at its first sentence, 18% at a middle sentence and 20% at its last.

Chance matters here. A start picked at random among the sentences of those same paragraphs would land on the first sentence 28% of the time, on a middle sentence 45% and on the last 28%. In a long paragraph the middle sentences outnumber the first and the last together, so a middle-sentence share needs its chance share next to it, and the article reporting Kevin's figure does not give one.

A blind second reading of 40 of these quotes, by an AI model that never saw my answers, put the start in the same sentence 39 times. Where the readings differed, the sentence splitter had counted a label (such as "A." or a step number) as the paragraph's first sentence, so the first-sentence share above is, if anything, low: the reader found 32 first-sentence starts where the splitter found 28.

Finding 07The passages AI Overviews quoted held fewer names than the rest of their page: 16% of their words, against 26%

Kevin Indig found cited ChatGPT text dense with names: heavily cited text had an entity density of 20.6%, against about 5 to 8% for normal English text. Against ordinary English, the passages Google's AI Overviews quoted look dense with names too. Against the rest of the same page they go the other way: by a word tagger, 16% of the words in the quoted passages were names, against 26% in other passages on the same page, and the quoted passage held fewer in 2,352 of 3,139 quotes.

The tagger counts too many names. A blind hand count of 100 of these passages, by an AI model that never saw which one was quoted, found 8% of their words were names where the tagger found 19%: it reads capitalised words in lists and labels as names. The direction held by hand: 6.4% of the words in the quoted passages were names, against 9.3% in another passage of the same page, fewer in 18 of 50 pairs and more in 10. So a share of names depends on the tool that counts it, and a comparison means most when both sides are counted the same way, on the same page.

Part of the gap comes from where passages start: other passages start in list items and loose page text more often than quotes do, and those hold more names. Keeping only passages that start in a paragraph on both sides, a check I added after seeing the first results, the gap shrinks but holds: 12% of the words in the quoted passages in AI Overviews were names, against 15%.

Kevin's other signals, measured the same way on both sides of each page:

Position is the clearest: the quotes in Google's AI Overviews start in the first 30% of the page more often than Kevin found for ChatGPT, and in the last 30% far less often. Subjectivity was about the same for the quoted passage and the rest of its page, and below Kevin's figure. The quoted passages read at a slightly lower grade, but among passages that start in a paragraph on both sides the gap reversed: 8.45 against 8.37. Question marks were a little more common in the quoted passages than in the rest of the page, a smaller gap than Kevin's. The definition phrases he names ("is defined as", "refers to", and a few like them) were rare on both sides. His article counts a wider set of definitive wording that it does not list, so that row does not test his figure.