We measured 58 pages that AI actually cites.

Across 58 cited pages the median carried 2,813 words, 11 H2 sections, and 110 list items. 72% had a visible FAQ section and 95% had exactly one H1. None had zero H1. The single most distinctive trait was list density: a median of 110 list items per page.

The median cited page was 2,813 words. Only 3 of 58 exceeded 6,000 words, while 25 sat between 2,000 and 3,500. Longer is not better. The common advice to publish 5,000-plus-word pages for AI visibility is not supported by what is actually being cited.

No. Only 41% of the cited pages had FAQPage schema, while 72% had a visible FAQ section. The visible section is far more common than the markup. Google states directly that no special schema is needed for AI features. Ship FAQ schema because it is cheap and honest, not because it buys citations.

Not in our data. Measuring referring domains against LLM mention counts across nine buyer prompts, netalico.com with 557 referring domains was cited 17 times, while outerboxdesign.com with 4,718 referring domains barely appeared on the same prompts. Authority buys organic position. It did not buy citation here.

Not for commercial queries. Only 18% of the cited pages had a numbered listicle headline. The listicle format is widely recommended for AI visibility and it was a clear minority pattern in this sample. Use it where the topic genuinely enumerates, not as a default.

32% of the cited pages contained a table. That makes tables common but not required. They are worth adding where the topic genuinely compares things, since a table is an unusually clean extractable unit, but the data does not support treating them as mandatory.

We ran 12 commercial e-commerce queries through Google with AI Overview loading enabled, captured 232 AI Overview citation URLs, excluded publisher and marketplace domains so the sample reflected pages a services company could realistically compete with, then fetched the 58 most-cited remaining pages and measured their rendered HTML.

Rendered word count after stripping script, style, noscript and SVG elements; heading counts by level; list items; tables; ordered and unordered lists; JSON-LD schema types; presence of a visible FAQ; listicle headline patterns; and whether dateModified was present. Everything was measured on live rendered HTML, never on source files.

Because source files include markup, styles, and structured data that inflate word counts by roughly double. We made that exact mistake mid-study and corrected it: a page measuring 6,883 "words" in source was 3,904 rendered. Any study comparing content length has to state which it measured.

Four worth stating. The sample is 58 pages in one vertical, e-commerce and agency commercial queries. It reflects Google AI Overviews only, not ChatGPT or Claude. It is capped at the 60 most-cited URLs, so long-tail cited pages are excluded. And it is correlation: these pages are cited and share these traits, which does not prove the traits caused the citation.

Our capture of citation URLs from ChatGPT and Perplexity answers returned no URLs on the run used for the structural analysis, so those pages could not be fetched and measured. The structural findings are Google AI Overview only. We have separate LLM mention-rate data, but not the page structures behind it.

Yes, deliberately and disclosed. We excluded publisher, marketplace, and social domains such as Forbes, Reddit, Clutch and the platform vendors themselves, because the question we were answering was what a services company can compete with. A study including those would produce different medians.

Three things with the clearest support. Raise list density, since the median cited page carries 110 list items and most service pages carry far fewer. Aim for 2,500 to 3,500 words rather than padding higher. Add a visible FAQ section, which 72% of cited pages had. Everything else in the data is weaker.

No. Authority still governs organic ranking position, and our own site sits at 53 referring domains against a competitor median of 1,799, which is exactly why we rank poorly on competitive terms. The finding is narrower: authority did not predict AI citation in this sample. The two are different games with different inputs.

Google states there are no additional requirements to appear in AI Overviews or AI Mode and no special schema.org structured data to add. Its guidance is ordinary SEO fundamentals: allow crawling, keep important content in text, use internal links, and ensure structured data matches visible content. Our data is consistent with that.

Not as a visibility lever. Google has stated no AI system currently uses llms.txt, and an Ahrefs analysis of 137,000 domains found 97% of llms.txt files received zero requests in May 2026. It is cheap and harmless to publish, but it is not a ranking or citation factor.