Unlike schema markup, length and structure do separate a cited page from its neighbours. In the Machine Relations Index answer ledger, 706 pages that AI engines cited were each paired with another page from the same website that the ledger never recorded as cited. The cited page had a median 1,560 words of main content against 968 for its neighbour. It was longer in 440 pairs and shorter in 227. The cited page was also more likely to carry a table (37.7 percent against 23.1), at least three H2 headings (79.0 against 64.0) and a modification date inside the last twelve months (53.1 against 40.7).
The gaps are not a length requirement. 11.0 percent of cited pages ran under 300 words, and 24.5 percent under 800. And they are not a cause: the ledger counts what cited pages look like, and it cannot say that adding words, tables or headings earns a citation.
What was counted #
- Population. Cited addresses in the Index's buying-question panel, successful runs from 2026-07-06 to 2026-11-07, five engines (ChatGPT, Claude, Gemini, Google AI Mode, Perplexity) through their APIs, runs that name a brand in the prompt excluded. This is not the consumer ChatGPT or Claude app. Google AI Overviews are left out. The ledger holds 78,194 distinct cited addresses on 29,282 root domains, the same population as the schema study, the subdomain study and the URL slug study.
- Sample. 1,000 root domains drawn at random (fixed seed) from the cited domains, social platforms removed (Reddit, YouTube, LinkedIn, Facebook, Instagram, X, TikTok, Quora, Medium), 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 big publisher is cited.
- Cited pages fetched. On 2026-11-09 each address was fetched by plain HTTP (15 second limit, browser user agent). 807 of 1,000 (80.7 percent) returned readable HTML. The other 193 were refused, timed out, missing or not HTML, 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 (425 of the 706 pairs; the rest come from anywhere in the sitemap). Up to three draws were tried and the first readable page kept. 706 of the 807 cited pages ended up with a readable neighbour, and all comparisons below are on those 706 pairs.
- What was measured. Words, headings, paragraphs, tables, lists, images and outside links were counted in 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. Dates come from JSON-LD, meta tags and time elements. Content added by JavaScript after load is not seen.
The paired result #
Share of pages with each feature, 706 cited pages against their 706 same-site neighbours. Difference intervals resample pairs, 2,000 draws, fixed seed.
| Feature | Cited pages | Same-site neighbours | Difference, points |
|---|---|---|---|
| 1,000 words or more | 68.1% (481) | 49.2% (347) | +19.0 (15.0 to 22.7) |
| 2,500 words or more | 27.3% (193) | 15.7% (111) | +11.6 (8.2 to 15.0) |
| Under 300 words | 11.0% (78) | 23.5% (166) | -12.5 (-15.3 to -9.6) |
| Three or more H2 headings | 79.0% | 64.0% | +15.0 (11.6 to 18.3) |
| A heading phrased as a question | 53.7% | 41.5% | +12.2 (8.4 to 16.1) |
| At least one table | 37.7% (266) | 23.1% (163) | +14.6 (10.9 to 18.0) |
| At least one list | 77.6% | 68.8% | +8.8 (5.4 to 12.2) |
| At least one image | 77.6% | 74.1% | +3.5 (0.8 to 6.1) |
| At least one outside link | 62.0% | 56.7% | +5.4 (1.7 to 9.3) |
| Author named in markup or meta tags | 67.0% | 57.9% | +9.1 (6.1 to 12.2) |
| Visible byline or author element | 40.1% | 34.4% | +5.7 (3.0 to 8.4) |
Medians tell the same story. Main content ran 1,560 words on the cited page against 968 on its neighbour (paired median difference +226 words, 138 to 350), with 8 H2 headings against 5 and 37 paragraphs against 23. The cited page's title was 60 characters against 56.
The length gap is a shift, not a floor. Share of pages by main-content length:
| Words | Cited pages | Same-site neighbours |
|---|---|---|
| Under 300 | 11.0% | 23.5% |
| 300 to 799 | 13.5% | 20.4% |
| 800 to 1,499 | 23.7% | 20.0% |
| 1,500 to 2,999 | 33.7% | 24.6% |
| 3,000 or more | 18.1% | 11.5% |
A quarter of cited pages (24.5 percent) ran under 800 words, so a short page can be cited. The pattern is that the cited page more often sat in the long, structured end of its own website.
Does the gap survive a fairer comparison? #
Sitemaps hold more than articles. Tag pages, category pages and utility pages sit in them, and 23.5 percent of neighbours had under 300 words. Two cuts remove most of that.
| Restriction | Pairs | Median words, cited against neighbour | Tables | Three or more H2 | Modified in last 12 months |
|---|---|---|---|---|---|
| All pairs | 706 | 1,560 against 968 | +14.6 | +15.0 | +12.5 |
| Both pages 300 words or more | 520 | 1,812 against 1,462 | +13.7 | +8.3 | +7.9 |
| Same top-level folder only | 425 | 1,704 against 1,331 | +13.2 | +10.1 | +9.9 |
Gaps are in percentage points except the word medians. On the 520 article-like pairs the length gap shrinks (paired median difference +164 words, 115 to 262) and the share with 1,000 words or more is 79.4 percent against 65.2. The table gap barely moves. So some of the headline length gap is cited articles against thin utility pages, and some is not.
Dates: a gap in showing one, not in being new #
Fresh content is a common bet, and here it is more about keeping a visible date than about being new. 53.1 percent of cited pages carried a modification date within the last 365 days, against 40.7 percent of neighbours (+12.5 points, 8.9 to 15.9). 26.9 percent against 20.4 had one inside 90 days.
Where both pages of a pair carried a date (408 pairs), the cited page was newer in 214 pairs and older in 126. But the median time since the newest date was 120 days on cited pages and 156 on neighbours, a paired median difference of -3 days (-16 to 0). Among pages that show a date, the cited ones were only slightly fresher. What differs is how often a page shows a date at all and keeps it recent: 66.9 percent of cited pages had a published date against 59.5 percent of neighbours, and among article-like pairs the difference in having any date fell to +2.5 points (-0.2 to 5.4).
How this fits the other studies #
The schema study found that a signal engines are widely assumed to want, JSON-LD, was on 87 percent of cited pages and 85 percent of their neighbours: no separation. This page finds signals that do separate: depth, headings that break a topic into parts, tables, and a visible recent update. They are about the content a page can offer an engine's passage retrieval, not about labels on the page. An October 2026 preprint, analysing 158,099 answers and 468,037 citations, found the same direction inside a website: pages whose title and paragraphs match the questions people ask are cited more. A page with eight sections and a comparison table has more places to match a question.
The ledger cannot rank the signals against each other, and several travel together: long pages tend to have more headings, more tables and more recent edits. The citation-factors page places this beside the other ledger studies. Third-party findings on length and format, which vary by industry and by engine, are collected on the format-divergence page.
What this changes #
- Do not pad. A quarter of cited pages ran under 800 words. The gap is in depth and structure, not a word count to hit.
- Give a long answer its sections and a table. The table gap (+14.6 points) held in every cut. Where a page compares options, a table is the structure the cited pages more often had.
- Show a real date and keep it real. Cited pages more often carried a modification date inside a year. The data does not show that changing a date changes a citation, and a date moved without an edit would be a false claim.
- Do not read this as a recipe. Cited pages are a selected set, and engines retrieve by matching a question. Longer, structured pages may simply be the ones that rank and match more questions.
What this does not show #
- Cause. Nothing here shows that adding words, headings or tables changes whether a page is cited.
- Rank. Cited pages may be longer because they rank higher in the search step that feeds the engines, and the ledger does not record rank.
- Content added by JavaScript. Pages that fill in text after load count as short. If that were common among cited pages, the cited length would be understated, and the cited side is where that hurts: 11.0 percent of cited pages fall under 300 words.
- The 19.3 percent not read. Sites that refuse plain requests may differ from the ones that answer. Both pages in a pair were fetched the same way.
- Neighbours are not guaranteed uncited. They are addresses the ledger did not record, drawn from a panel of 1,008 questions. A neighbour might be cited on a question the panel never asked.
- Page type. Page type was not matched beyond the folder cut and the 300-word cut. Product pages, listings and archives still sit among the neighbours.
- All websites. Every site in the sample was cited at least once. This page says nothing about websites the engines never cite.
Method note #
Source: the Index answer ledger's citation-events table, daily export of 2026-11-08 08:10 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 2026-11-07, runs marked successful, which reproduces 78,194 distinct cited addresses on 29,282 root domains. 1,000 root domains drawn with a fixed seed (20261109) from domains outside the social list above, and one cited address drawn at random per domain. Fetch: plain HTTP GET on 2026-11-09 between 03:38 and 03:41 UTC, 15 second limit, a page counted as readable when status 200, a content type containing html and a body that parsed. Same-site neighbour: robots.txt sitemap entries (or /sitemap.xml), up to six sitemap files per site read with sitemap indexes sampled three children at a time, image, PDF and XML files removed, any address in the ledger's cited set removed, shuffled, first readable of three tried. Words: main, else article, else role-main, else body, with navigation, header, footer, aside, form, script, style and template elements removed, counted as runs of word characters. Question heading: an H2 or H3 whose text ends in a question mark. Dates: earliest published and latest modified value read from JSON-LD datePublished and dateModified, article and og meta tags, and time elements, with years before 1995 or after 2026 discarded; age counted to 2026-11-09. Intervals on differences resample pairs with replacement, 2,000 draws, fixed seed. The same fixed seed is used in the schema study; the two samples were drawn by the same method and each reports its own pairs. The fetches read live pages on 2026-11-09, so a later re-run reads pages that may have changed.
Frequently asked questions #
Do AI engines cite longer pages? #
More often than not, compared with other pages on the same website. In 706 pairs, the cited page had a median 1,560 words against 968, and was longer in 440 pairs and shorter in 227. On the 520 pairs where both pages ran 300 words or more, the medians were 1,812 and 1,462.
How long should a page be to get cited by AI? #
This ledger sets no minimum. 24.5 percent of the cited pages ran under 800 words and 11.0 percent under 300. 51.8 percent ran 1,500 words or more.
Do tables help a page get cited by AI? #
Cited pages carried them more often: 37.7 percent against 23.1 percent of same-site neighbours (+14.6 points, 10.9 to 18.0), and the gap held when both pages were article-length and when both sat in the same folder. The ledger counts co-occurrence and does not show that adding a table earns a citation.
Does updating a page's date help AI citation? #
The data does not show that. Cited pages more often carried a modification date within a year (53.1 percent against 40.7), but where both pages showed a date the median age differed by only 3 days. Moving a date without changing the page would also be untrue.
Is content length a ranking factor for AI search? #
This page cannot say. It counts what cited pages look like next to uncited pages of the same website, not how any engine ranks or selects. The schema study found no such separation for markup, so the signals that do separate here are worth testing directly.