Research

Do AI-Cited Pages Open With the Answer? 78% Open Short Against 73% on the Same Websites

Across 707 same-site pairs, 77.7% of AI-cited pages opened with a short paragraph against 73.0% of neighbours, and a number in the opening did not separate them.

Published Machine Relations Research, founded by Jaxon Parrott
Index Analysis
TopicsCitation BehaviorContent StructureAnswer FirstMachine Relations Index

Answer-first openings separate cited pages from their same-site neighbours by about five points, not by the margin that length does. In the Machine Relations Index answer ledger, 707 pages that AI engines cited were each paired with another page from the same website that the ledger never recorded as cited. 77.7 percent of the cited pages opened with a first paragraph of 60 words or fewer, against 73.0 percent of the neighbours (+4.7 points, 0.7 to 8.3). A number in the first 100 words did not separate them: 77.5 percent against 79.1. The median cited page ran 1,745 words against 1,011, and was longer in 481 pairs and shorter in 219.

The small opening gaps shrink when both pages sit in the same folder of a site, and their intervals then include zero. This is a count of what cited pages look like, not a test that rewriting an opening earns a citation, and it measures which page gets cited, not where on a cited page an engine copies from. That is a different question, which the text-overlap study answers for ChatGPT.

What was counted #

  • Population. Cited addresses in the Index's buying-question panel, runs from 2026-07-06 to 2026-11-18, five engines (ChatGPT, Claude, Gemini, Google AI Mode, Perplexity) through their APIs, runs that name a brand in the prompt excluded, social platforms removed (Reddit, YouTube, LinkedIn, Facebook, Instagram, X, TikTok, Quora, Medium). That leaves 79,587 distinct cited address strings on 31,116 root domains in the daily export of 2026-11-19 08:11 UTC. This is not the consumer ChatGPT or Claude app, and Google AI Overviews are left out.
  • Sample. 1,000 root domains drawn at random (fixed seed 1119), and one cited address drawn at random from each. Each domain counts once, so the sample describes websites that get cited, not citations weighted by how often a large publisher is cited.
  • Cited pages fetched. On 2026-11-19 between 08:35 and 08:39 UTC each address was fetched by plain HTTP (12 second limit, browser user agent). 799 of 1,000 returned readable HTML with at least 30 words of main content. The other 201 were refused, timed out, missing, not HTML or empty, and are unmeasured, not zero.
  • Same-site neighbours. For each readable cited page, one page was drawn at random from its website's sitemap, excluding any address the ledger records as cited, and from the same top-level folder when the sitemap held at least five addresses there (365 of the 707 pairs). Up to three draws were tried and the first readable page kept. 92 cited pages found no readable neighbour. All comparisons are on the 707 pairs.
  • What was measured. The page's main content as served (the main element, else the article element, else the body, with navigation, header, footer, sidebar, form and script content removed). Detection rules are in the method note. Content added by JavaScript after load is not seen.

The paired result #

Share of pages with each feature, 707 cited pages against their 707 same-site neighbours. Difference intervals resample pairs, 2,000 draws, fixed seed.

Feature Cited pages Same-site neighbours Difference, points
Number in the first 100 words 77.5% (548) 79.1% (559) -1.6 (-5.0 to +1.7)
Percentage, dollar figure or the word percent in the first 100 words 18.8% (133) 16.7% (118) +2.1 (-1.6 to +5.5)
First paragraph of 60 words or fewer 77.7% (549) 73.0% (516) +4.7 (0.7 to 8.3)
Definition-style verb in the first 60 words 56.4% (399) 50.6% (358) +5.8 (0.8 to 10.5)
Short first paragraph and a number or definition-style verb 71.6% (506) 66.5% (470) +5.1 (0.8 to 9.6)
Summary marker in the first 300 words 20.2% (143) 14.6% (103) +5.7 (2.5 to 8.9)
Heading 1 phrased as a question 7.4% (52) 4.1% (29) +3.3 (1.0 to 5.5)
FAQ markup or an FAQ heading 49.9% (353) 38.0% (269) +11.9 (8.1 to 15.7)
Link to a primary-source host 13.6% (96) 7.9% (56) +5.7 (3.1 to 8.1)

The first paragraph ran a median 37 words on cited pages and 33 on neighbours, with a paired median difference of 0 (0 to 0): longer on the cited page in 324 pairs, shorter in 308. Numbers in the first 200 words had medians of 4 and 5. Main content ran a median 1,745 words against 1,011, a paired median difference of +358 words (274 to 506).

Does the gap survive a fairer comparison? #

Sitemaps hold more than articles, and a site's own folders differ. Two cuts test that.

Feature, difference in points All 707 pairs Both pages 300 words or more (553) Same top-level folder (365)
Number in the first 100 words -1.6 -0.7 -2.2
First paragraph of 60 words or fewer +4.7 (0.7 to 8.3) +5.4 (1.1 to 9.2) +3.8 (-0.8 to 8.8)
Definition-style verb in the first 60 words +5.8 (0.8 to 10.5) +3.8 (-1.1 to 8.9) -1.6 (-7.7 to 3.8)
Short first paragraph and a number or definition-style verb +5.1 (0.8 to 9.6) +5.8 (1.3 to 10.5) +1.6 (-3.8 to 7.1)
Summary marker in the first 300 words +5.7 (2.5 to 8.9) +5.4 (2.0 to 8.9) +3.0 (-1.4 to 7.4)
FAQ markup or an FAQ heading +11.9 (8.1 to 15.7) +7.4 (3.3 to 11.8) +6.8 (2.2 to 11.2)
Link to a primary-source host +5.7 (3.1 to 8.1) +5.2 (2.4 to 8.3) +3.0 (-0.5 to 6.8)

Inside the same folder of a site, a page's opening style told cited pages from neighbours less than it did across the whole sitemap: the definition-style gap reverses and the combined answer-first gap falls to +1.6 with an interval that includes zero. The FAQ gap shrinks to +6.8 but its interval stays above zero in all three cuts. Median words were 1,934 against 1,373 for the 553 article-length pairs and 1,853 against 1,424 for the 365 same-folder pairs.

How this fits the other studies #

The length study found that depth, headings and tables separate cited pages from their neighbours on the same websites, and the schema study found that JSON-LD does not. This study finds the opening sits between the two: a short answer-style opening is somewhat more common on cited pages, and what is clearly more common is a longer page with an FAQ.

The finding does not contradict the ChatGPT result that, once a page is cited and its wording reused, 68.3 percent of ChatGPT's longest shared passages began in the first 100 words. That result is about where text is copied from a page that was already chosen. This one asks whether the opening helps a page be chosen from among a website's pages, and the opening features here barely separate chosen pages from the rest. The two are consistent with an engine that picks the page first and then reuses wording from near its top, which this ledger does not test. The structural-changes evidence and the citation-factors page place both beside third-party studies.

The ledger cannot say whether a short opening or an FAQ is what the engines reward, and features may travel together: this study did not test whether longer pages are the ones that carry an FAQ or a summary marker.

What this changes #

  • Do not rewrite an intro to get cited. A number in the opening, a common answer-first tip, did not separate cited pages from neighbours at all.
  • Keep the answer first for the reader and for ChatGPT's reuse. The paired gap is small but not negative, and the text-overlap study shows where ChatGPT copies from.
  • Put the effort where the gap is larger. Depth and a real FAQ separated cited pages more clearly than any opening feature, and the FAQ gap held in all three cuts.
  • Do not read this as a recipe. Cited pages are a selected set. A page that ranks and matches more questions may also be longer and carry an FAQ.

Limits #

  • Opening features are detected by simple text rules, so a short first paragraph is not proof of a direct answer, and a number is not proof of evidence.
  • 20.1 percent of the sampled cited pages could not be read and are unmeasured.
  • A neighbour is a page the ledger did not record as cited in this panel, not a page no engine has ever cited, and it may be a tag, category or utility page.
  • Each of the 1,000 domains contributes one pair, so large publishers weigh the same as small sites.
  • The fetches read live pages on 2026-11-19, later than the runs that cited them.

Method note #

Source: the Index answer ledger's citation-events table, daily export of 2026-11-19 08:11 UTC, read with DuckDB. Filters: the buying-question panel, brand-prompted runs excluded, engines ChatGPT, Claude, Gemini, Google AI Mode and Perplexity, run dates 2026-07-06 to 2026-11-18, address starting with http, social hosts removed; no separate success filter was applied beyond the table holding only recorded citations. 1,000 root domains drawn with a fixed seed (1119), one cited address drawn per domain. Fetch: plain HTTP GET on 2026-11-19, 12 second limit, a page counted as readable when status 200, a content type containing html and at least 30 words of main content. Neighbour: robots.txt sitemap entries (or /sitemap.xml), up to five sitemap files per site with sitemap indexes sampled two children at a time, any address in the ledger's cited set for that domain removed, home pages removed, first readable of three draws. Words: runs of non-space characters in the main, else article, else body element with navigation, header, footer, aside, form, script, style, noscript and svg elements removed. First paragraph: the first paragraph element of eight words or more. Number: any digit in the first 100 words. Percent or money: a digit followed by percent sign, a dollar sign followed by a digit, or the word percent, in the first 100 words. Definition-style: the word is, are, means, refers to or defined as in the first 60 words. Short opening: first paragraph of 60 words or fewer. Combined opening: a short opening with a number or a definition-style verb. Summary marker: TL;DR, key takeaway(s), in short, the short answer, quick answer, bottom line, at a glance or summary in the first 300 words. FAQ: FAQPage in JSON-LD, or the phrase frequently asked or the word FAQ in the main text. Primary-source host: a link to a .gov or .edu host, arXiv, DOI, Nature, NCBI, Pew, Census, BLS, OECD, World Bank, SSRN, ACM or IEEE. Intervals on shares resample pairs with replacement, 2,000 draws; intervals on medians, 500 draws; fixed seed. The script and the per-pair file are retained in the Machine Relations Research repository with this analysis. A rerun reads pages that may have changed.

Frequently asked questions #

Do AI engines cite pages that open with the answer? #

Somewhat more often than other pages on the same website, but the gap is small. 77.7 percent of cited pages opened with a first paragraph of 60 words or fewer against 73.0 percent of same-site neighbours (+4.7 points, 0.7 to 8.3), across 707 pairs.

Does putting a statistic in the first paragraph help a page get cited? #

The ledger shows no gap. A number in the first 100 words appeared on 77.5 percent of cited pages and 79.1 percent of neighbours (-1.6 points, -5.0 to +1.7).

This study does not say how any engine ranks. It counts what cited pages look like next to other pages on the same websites. The opening features separate them by about five points; within a site's own folder the gap shrinks to include zero.

What separates cited pages from other pages on the same website more than an opening does? #

Length and an FAQ. The cited page had a median 1,745 words against 1,011 (+358 words, 274 to 506), and FAQ markup or an FAQ heading appeared on 49.9 percent of cited pages against 38.0 percent, a gap that stayed above zero in every cut.

Should I still write answer-first? #

Yes, for readers, and because ChatGPT's reuse of cited wording concentrates near the top of a page. Do not expect the opening alone to make a page get cited.

Published by Machine Relations Research, founded by Jaxon Parrott, and classified as Index Analysis under the Machine Relations editorial standards.

Cite this research: Machine Relations Research, founded by Jaxon Parrott. "Do AI-Cited Pages Open With the Answer? 78% Open Short Against 73% on the Same Websites." November 19, 2026. https://machinerelations.ai/research/do-ai-engines-cite-pages-that-open-with-the-answer-2026

Data: Machine Relations Index, created by Jaxon Parrott, CC BY 4.0.

Machine-readable version at /research/do-ai-engines-cite-pages-that-open-with-the-answer-2026.md.

Machine Relations Index — data & methodology →