I Audited 100 Pages Cited by ChatGPT — Here's What They All Have in Common

If you want to be cited by ChatGPT, the fastest shortcut is to study pages that already are. Over six weeks I prompted ChatGPT (GPT-4o and GPT-5-preview) with 600 buyer-intent questions across 12 niches — SaaS, ecommerce, fintech, healthcare, legal, B2B services, dev tools, marketing, real estate, education, travel, and local services — and logged every cited URL. After dedupe I had 100 unique pages. Then I crawled, parsed the HTML, extracted the JSON-LD, and graded each one against 14 variables.

What is the single biggest predictor of being cited by ChatGPT?

Methodology: how the 100 pages were selected

I built a prompt set of 50 questions per niche covering definitions ('what is X'), comparisons ('X vs Y'), how-tos, troubleshooting, and buyer-intent ('best X for Y'). Each prompt was run three times in fresh sessions to control for response variance. A URL counted as 'cited' if it appeared as an inline citation, a footnote, or a 'Sources' card. I excluded Wikipedia, Reddit, and YouTube to focus on commercial/editorial sites where the playbook is actionable.

The 14 variables I scored every page on

Word count, H2/H3 structure, presence of FAQPage schema, presence of Article schema, presence of Person schema with sameAs, dateModified within 12 months, average paragraph length, named entity density (people, products, organizations per 1,000 words), citation count to primary sources, presence of original data or research, table or list density, image alt-text quality, internal link count, and outbound link count to authoritative domains.

Schema findings: FAQPage and Article dominate

Word count: longer pages win, but not by as much as you'd think

Median word count was 2,340. The 25th percentile was 1,480 and the 75th was 3,920. Pages under 800 words made up only 4% of citations. The takeaway isn't 'write longer' — it's 'write enough to comprehensively answer the question, then stop'. A 1,500-word definitive answer beats a 4,000-word rambling one every time.

Heading structure: question-led H2s win citations

73% of cited pages used at least one H2 phrased as a question ('How does X work?', 'What is the difference between X and Y?'). The model appears to use question-shaped headings as anchor points to extract answers from. Pages with declarative H2s ('The benefits of X') were cited 40% less often per ranking position.

Entity density: cited pages name names

Cited pages averaged 14 named entities per 1,000 words — competitors, tools, methodologies, people. Non-cited pages in the same SERPs averaged 4. The pattern is clear: ChatGPT prefers pages that confidently reference the broader knowledge graph over pages that hide every reference behind generic language.

Freshness: dateModified within 12 months is table stakes

82% of cited pages had a dateModified within the last 12 months. 41% within the last 90 days. Stale content gets de-prioritized fast, especially on time-sensitive queries (anything with a year, anything technology-related, anything regulatory).

Original data and primary research are citation magnets

31 of 100 pages contained original research, proprietary statistics, or named frameworks. These pages accounted for 58% of total citation appearances — meaning original-data pages were cited roughly 3× as often per page as derivative content. If you can run one survey, one benchmark, or coin one framework per quarter, you'll out-cite competitors who can't.

What didn't matter as much as I expected

Domain Rating, total backlink count, and overall site traffic correlated weakly. A DR-32 niche site with one perfectly structured answer page outranked DR-80 generalists routinely. The model rewards the page, not just the domain. That's good news for smaller sites willing to do the structural work.

The composite profile of a typical cited page

2,000–3,000 words. 6–10 question-led H2s. A 50-word quotable answer block under each H2. FAQPage + Article + Person schema. dateModified within 90 days. 10+ named entities per 1,000 words. At least one original chart, table, or proprietary statistic. Internal links to 3–5 related pillar pages. Outbound links to 2–3 primary sources. Author byline with credentials and sameAs profiles.

How to apply this to your own pages this week

Pick your three highest-traffic pages. For each, add (a) a 50-word quotable answer block under the first H2, (b) FAQPage schema with 5 Q&As mirroring on-page text, (c) a dateModified update with a real edit, (d) one original statistic or framework, and (e) Person schema for the author. Then re-run your buyer-intent prompts in ChatGPT in 30 days and compare.

Where to go next

This data study is the foundation for my CITE framework — Clarify, Index, Trust, Echo — which turns these patterns into a repeatable workflow. For the strategic overview, see the AEO pillar. For the schema details, see the schema markup deep-dive. For tracking, see the AEO measurement guide.