Most of what looks like a citation signal on a page is a length effect. In the Machine Relations Index answer ledger, 659 pages that AI engines cited were each paired with another page from the same website that the ledger never recorded as cited, and both were scored on 14 features. Twelve of the 14 were more common on the cited page. Where the two pages were within one and a half times each other's length (294 pairs), two still were: a first paragraph of 60 words or fewer (+9.5 points, 4.1 to 15.0) and a link to a primary-source host (+3.7, 0.3 to 7.5). The short opening is the only one that held in every equal-length cut, including pairs from the same folder of a site.
This is a count of what cited pages look like, not a test that adding any feature earns a citation. It says where the evidence stops: for named authors, dates, tables, FAQ markup, lists and Article-type markup, the ledger shows pages that are bigger and more article-like, not pages that carry a marker the neighbour lacks.
What was counted #
- Pairs. The pairs of the opening study and the named-author study: one cited address drawn at random from each of 1,000 random root domains in the Index's buying-question panel (runs 2026-07-06 to 2026-11-18, five engines through their APIs, brand-prompted runs and social platforms excluded), and one same-site page from the website's sitemap that the ledger never recorded as cited. 659 of the 707 pairs were readable in both fetches and are scored here; the other 48 are unmeasured, not zero.
- Features. Fourteen, none new: seven from the named-author fetch (named author, any date, outbound link, outbound links to three or more sites, Article-type JSON-LD, a table, three or more images), five or more lists, and the opening-study set (first paragraph of 60 words or fewer, a number in the first 100 words, FAQ markup or heading, a summary marker, a question headline, a link to a primary-source host). Detection rules are in the method note.
- Length. Cited-page words divided by neighbour words, from the second fetch. The median cited page ran 1,745 words against 1,085 for its neighbour.
Twelve of 14 look like citation signals until length is matched #
The difference in points between cited page and neighbour, with its interval (2,000 resamples of pairs, fixed seed). The last two columns keep only pairs whose two pages were within one and a half, and one and a quarter, times each other's length.
| Feature | All 659 pairs | Within 1.5x (294) | Within 1.25x (190) |
|---|---|---|---|
| First paragraph of 60 words or fewer | +5.3 (1.1 to 9.1) | +9.5 (4.1 to 15.0) | +6.8 (0.5 to 13.7) |
| Link to a primary-source host | +5.0 (2.4 to 7.7) | +3.7 (0.3 to 7.5) | +5.8 (1.6 to 10.0) |
| Outbound link to another website | +10.3 (6.2 to 14.3) | +3.1 (-1.4 to 7.1) | +5.3 (0.5 to 10.5) |
| Headline phrased as a question | +3.5 (1.2 to 5.9) | +3.4 (0.0 to 7.1) | +3.7 (-0.5 to 8.4) |
| A table | +14.0 (10.2 to 17.9) | +4.8 (-0.3 to 9.5) | +2.6 (-3.2 to 8.9) |
| Any publish or update date | +8.8 (5.6 to 12.3) | +1.7 (-1.4 to 4.8) | +3.7 (0.0 to 7.9) |
| Article-type JSON-LD | +12.1 (8.8 to 15.3) | +0.7 (-2.4 to 3.7) | +2.6 (-1.1 to 6.3) |
| Five or more lists | +12.6 (8.3 to 16.8) | +2.0 (-3.4 to 7.5) | 0.0 (-7.4 to 7.4) |
| A named author | +8.3 (5.5 to 11.4) | +1.4 (-1.7 to 4.4) | +1.1 (-3.2 to 5.3) |
| Summary marker in the first 300 words | +5.5 (2.1 to 8.6) | +1.0 (-3.7 to 5.4) | +1.6 (-4.2 to 6.8) |
| Outbound links to three or more websites | +8.8 (5.2 to 12.4) | -0.7 (-5.1 to 3.7) | +0.5 (-4.2 to 5.3) |
| FAQ markup or an FAQ heading | +10.8 (7.0 to 14.6) | -0.7 (-5.4 to 3.7) | -1.1 (-6.8 to 4.7) |
| A number in the first 100 words | -0.9 (-4.6 to 2.7) | -0.3 (-4.8 to 4.4) | -3.2 (-8.4 to 2.1) |
| Three or more images | +1.7 (-1.7 to 4.9) | -1.7 (-5.1 to 1.4) | -3.2 (-7.4 to 0.5) |
Twelve features have an interval above zero across all pairs. Within 1.5x the median pair was 1,611 words against 1,645, and two features keep an interval above zero. Within 1.25x the same two do, plus an outbound link to any other website.
The gaps follow how much longer the cited page is #
Where the cited page was more than one and a half times longer (303 pairs, median 2,153 words against 548), 11 of the 14 features were more common on it: an Article-type block +25.1 points, FAQ markup +25.7, a table +27.1, five or more lists +29.4, a named author +16.2. In those pairs the neighbour was often a tag, category or utility page: 42.6 percent carried Article-type JSON-LD against 67.7 percent of the cited pages. The one feature that does not rise there is the short opening (-1.3, -7.9 to +5.0). It separates two comparable articles, not an article from a stub.
What survives, and how firmly #
- A short opening held in every equal-length cut. +9.5 within 1.5x, +6.8 within 1.25x, and +8.6 (2.5 to 15.2) among the 197 pairs from the same top-level folder and within 1.5x. The opening study's own length cut, run on its first fetch, gave +8.7 (3.5 to 13.8).
- A primary-source link held in two of three. +3.7 within 1.5x and +5.8 within 1.25x, but +2.5 (-2.0 to 7.1) in the same-folder cut. A primary-source host here means a .gov or .edu address or a research and statistics site from a short list in the method note, present on 13 percent of cited pages. With 14 features and 95 percent intervals, about one clearing zero by chance is expected, so this one is a lead, not a finding.
- Four are borderline. An outbound link to any site clears zero within 1.25x only. A date touches zero within 1.25x (0.0 to 7.9) and clears it in the same-folder cut (+2.5, 0.5 to 5.1), not within 1.5x. A question headline (+3.4, 0.0 to 7.1) and a table (+4.8, -0.3 to 9.5) sit on the edge within 1.5x and fall back within 1.25x.
- Eight are level. A named author, Article-type JSON-LD, five or more lists, FAQ markup, a summary marker, three or more outbound sites, a number in the opening and three or more images separate no equal-length pair.
How this fits the other studies #
The length study found cited pages longer and more often carrying tables and several H2 headings; this sweep says those are largely the same fact seen twice. The schema study found JSON-LD about as common on both sides; Article-type markup is the page-type marker that is not, until length is matched. The citation-factors page now reads each structural row with this control in view.
What this changes #
- Do not add markers to get a page cited. At similar length a byline, a date, a table, a list, an FAQ block or Article markup did not separate cited from uncited pages.
- Write the substantial article. What most of the gap reflects is being the full page on a site's topic rather than a thin one. Whether length itself drives a citation is not something this design can say: it removes length as a way to see features, not as a cause.
- If you test one change, test the opening. A first paragraph of 60 words or fewer is the one feature that separated comparable pages in every cut. It is an association, and a number in the opening did not help.
- Cite primary sources where they exist. The link to a primary-source host is a smaller, less firm gap; it costs nothing and is worth a dated test.
Limits #
- The length bands are retrospective cuts of one sample, not a matched experiment. The within-1.25x band has 190 pairs and wide intervals; 14 features and three bands make 42 intervals, so a lone result near zero should not be leaned on.
- A neighbour is a page the ledger did not record as cited in this panel, not a page no engine has cited. The ledger counts the pages engines cited in this panel; it cannot say why an engine picked one.
- Detection uses simple rules on pages fetched without running JavaScript; markup added after load is not seen. The opening features come from the first fetch and the rest from the second, on the same day.
- 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.
- Intervals are resampled with a fixed seed and can differ by a few tenths from the same cell in an earlier study.
Method note #
Pairs: the 707 same-site pairs of the opening study (Index answer ledger citation-events table, daily export of 2026-11-19 08:11 UTC, the buying-question panel, runs 2026-07-06 to 2026-11-18, engines ChatGPT, Claude, Gemini, Google AI Mode and Perplexity, brand-prompted runs and social hosts removed, 1,000 root domains drawn with fixed seed 1119). A pair counted when both pages were readable (status 200, an html content type, at least 30 words of main content) in the opening study's first fetch and again in the named-author study's second fetch on 2026-11-19 between 11:40 and 11:46 UTC: 659 pairs. Length bands: cited-page words divided by neighbour words from the second fetch; within 1.5x is a ratio from 0.67 to 1.5 (294 pairs), within 1.25x from 0.8 to 1.25 (190), cited 1.5x+ longer above 1.5 (303); the same-folder cut keeps the 197 within-1.5x pairs whose two addresses share a top-level folder. Features from the second fetch: named author, any date, outbound link, outbound links to three or more registrable hosts, Article-type JSON-LD, a table, three or more images, as defined in the named-author study, and five or more ul or ol elements in main content. Features from the first fetch: first paragraph of 60 words or fewer; a digit in the first 100 words; FAQPage markup or the words frequently asked or FAQ in the first 200,000 characters of text; a summary marker, meaning any of tl;dr, key takeaway, in short, the short answer, quick answer, bottom line, at a glance or summary in the first 300 words; a headline that ends with a question mark; a link in main content to a host matching .gov, .edu, arxiv.org, doi.org, nature.com, ncbi, pewresearch, census, bls.gov, oecd, worldbank, ssrn, acm.org or ieee. Intervals resample pairs with replacement, 2,000 draws, seed 7. The script and both per-pair files are retained in the Machine Relations Research repository with this analysis. A rerun reads pages that may have changed.
Frequently asked questions #
What page features do AI-cited pages have? #
More of almost everything, because they are longer. Across 659 same-site pairs, 12 of 14 features were more common on the cited page, including a table (+14.0 points), five or more lists (+12.6), Article-type markup (+12.1), FAQ markup (+10.8) and a named author (+8.3). Where the two pages were within one and a half times each other's length, only a short first paragraph (+9.5) and a link to a primary-source host (+3.7) still were.
Does page length explain why AI engines cite some pages? #
It explains most of the structural gaps in this sample, though the data cannot say it causes a citation. The median cited page ran 1,745 words against 1,085 for its neighbour, and where the cited page was more than 1.5 times longer, 11 of 14 features were more common on it. Among pages of similar length, only 2 of the 14 cleared zero.
Which on-page change is most worth testing to get cited? #
The opening. A first paragraph of 60 words or fewer was more common on the cited page in every equal-length cut: +9.5 points within 1.5x, +6.8 within 1.25x and +8.6 among pairs from the same folder. It is an association from counting pages, so the test is to change one opening and read the same questions again.
Do FAQ markup, tables or Article schema help a page get cited? #
This sweep does not show it. Their gaps were +10.8, +14.0 and +12.1 points across all pairs and -0.7, +4.8 (-0.3 to 9.5) and +0.7 within 1.5x length. Cited pages carry them more often because they are more often full articles.