The Anatomy of a ChatGPT-Cited Page: 33,000 URLs Analyzed

On May 1, 2026, Evertune published findings from a study that most content teams have not seen yet. Researcher Will Robinson analyzed 33,000 to 40,000 URLs tracked through the Evertune platform that received heavy ChatGPT citations, across nearly 50 product categories, over 60 days.

The question was not which topics get cited. It was: what does the page itself actually look like?

The answer is nine data points describing a structural profile. Not content recommendations. Not topic guidance. The physical architecture of the pages that ChatGPT chose to cite, measured at the median.

This is the most specific dataset on ChatGPT citation structure published in 2026. Here is what it found.

The nine structural patterns

Evertune documented nine consistent structural patterns across the cited URL set. The median values are below, with what each number actually means:

941 words. The median cited page is not a 3,000-word pillar page. It is focused, specific content in the 800 to 1,200 word range. AirOps' analysis of 16,851 queries and 50,553 responses arrived at the same conclusion from a different dataset: the optimal range for ChatGPT citations is 500 to 2,000 words. Pages over 5,000 words underperform even short pages under 500 words. The assumption that longer, more thorough content is what AI rewards does not hold up.

18 paragraphs. At 941 words across 18 paragraphs, that averages roughly 52 words per paragraph. Short, discrete blocks. The passages-beat-pages principle captures why this matters: AI models extract 40 to 60 word passages, not full articles. Pages structured around dense paragraphs produce more extractable units per section.

17-word average sentence. Not fragmented one-liners. Not winding academic sentences. Seventeen words holds one complete idea. It is the sentence length where meaning is clear without requiring a second read.

4 H2 headers. Four section headings across a 941-word page works out to roughly one heading every 235 words. Enough structure for AI to parse topical organization. Not so much that each section becomes too thin to contain a complete answer.

2 H3 headers. The second level of hierarchy is used sparingly. H3s add context within sections without over-fragmenting the content into subsections.

28 internal links. This is one of the higher numbers in the dataset. A page with 28 internal links is part of a publishing program where the team has built out the topic over time. Topical authority correlates at 0.76 with AI citations in EMGI Group's April 2026 analysis of 150 SaaS companies. Internal links are one of the clearest signals of topical depth.

15 external links. This finding gets less attention than it deserves. Fifteen external links to primary sources on a single page means the median ChatGPT-cited page actively links out to credible sources. It is not self-referential. It points to original research, named studies, and primary data.

10 images. This is a consistent number across cited pages. It reflects content depth and production investment, not decorative design. Pages with charts, data displays, and visual aids signal that the publisher took the material seriously.

2 embedded lists. Structured lists are one of the formats AI models favor for extraction. The median cited page uses two of them, not ten. List format is used selectively.

The external links finding

Most GEO strategy advice focuses on getting others to cite you. The Evertune data points to something additional: pages that get cited also link out aggressively.

Fifteen external links per cited page is not the behavior of a publisher trying to keep traffic on their own site. It is the behavior of a publisher treating their content as a research document. The page cites a study. It references the primary source. It links to original data.

This is consistent with what analysis of AI citation half-life patterns found about reference-grade content: specific data, disclosed methodology, and conclusions supported by cited evidence receive three to five times more citations than standard informational content. The external link count is a proxy for whether a page is self-referential or genuinely grounded in primary sources.

For content teams, the practical standard is: every factual claim in a published piece should trace back to a named source with a direct link. Not "studies show." A specific study, linked to the original finding. The Evertune data suggests that pages operating at this standard are the ones ChatGPT cites.

Half of all ChatGPT citations go to listicles

Format matters as much as structure. Evertune found that 50% of ChatGPT citations come from listicle-format content. Of those listicle citations, 58% are ranked lists specifically ("Top 10," "Best 7," "5 Most Reliable").

Self-promotional listicles have a different story. Google's AI Overviews penalized vendor-produced "best of" lists that only include the publisher with 30 to 50% visibility drops. ChatGPT does not have the same enforcement mechanism, but the cited listicles in Evertune's dataset are structured around third-party evaluation, not self-serving rankings. The format works when the content earns its list status. It does not work when the publisher places themselves at the top.

The 50% listicle finding also lines up with the broader platform-level data from AirOps' 2026 State of AI Search: 74.2% of all AI citations across platforms go to list-format content. Two different datasets, two different methodologies, and the same directional conclusion.

Cross-validation with AirOps

The Evertune findings do not stand alone. AirOps published its "2026 State of AI Search" report in April, based on 16,851 queries, 50,553 responses, and 1.5 million data rows from ChatGPT. The overlap with Evertune's structural profile is direct:

Two datasets arriving at structurally similar conclusions from different methodologies gives this data more weight than either study would carry alone.

One AirOps finding Evertune's study does not cover: pages updated 30 to 89 days before a query are cited at optimal rates. Content older than two years faces a citation penalty. The content refresh cycle is part of maintaining the citation profile, not just achieving it at publication.

What this means for your content brief

The Evertune data translates into a specific structural target for any page aimed at ChatGPT citations:

• •Word count: 800 to 1,100 words, centered on the 941-word median
• •Paragraphs: 15 to 20, averaging 50 words each
• •H2 headers: 4, each addressing a distinct question the target reader would ask
• •H3 headers: 1 to 3, for sub-questions within longer sections
• •Internal links: 20 to 30, pointing to related content in the same topic cluster
• •External links: 12 to 18, pointing to primary research, named studies, and authoritative sources
• •Images: 8 to 12, at least one data visualization or chart
• •Lists: 2 embedded structured lists

The external links target is the most demanding item. Twelve to eighteen links to primary sources means every factual claim traces back to a named, linked source. Reaching that standard requires genuine research before writing, not just restatement of general knowledge.

This brief also implies a different content production approach. A 941-word page with 15 external links is a more focused piece than most content teams currently produce. It does not cover everything. It covers one question thoroughly, with citations.

The off-page dimension

Structure alone does not explain ChatGPT citations. Otterly.ai published a controlled experiment in March 2026 that tested what happens when a brand has none of the structural advantages.

They created a fictional GEO agency with just seven pages, no schema markup, zero domain authority, and no web presence. Then they built 16 external placements over 14 days: directory listings on Sortlist, Clutch, Goodfirms, and DesignRush; Reddit thread participation; LinkedIn Pulse articles; and listicle placements on authority sites.

Result: rank seven out of more than ten named competitors in ChatGPT, 90 total AI mentions across platforms, 10% share of voice, and 12% brand coverage.

One detail from that experiment is important for content strategy: 74% of the brand mentions came from location or year-specific prompts ("best GEO agency in Berlin 2026"), not generic category queries. The fictional brand ranked well on specific prompts but remained invisible on broad ones.

That finding suggests the on-page structure and off-page placement address different citation problems. On-page structure determines whether a page gets cited when ChatGPT retrieves content for a category query. Off-page placement determines whether a brand gets mentioned when buyers ask who the players are. Both matter, and how AI platforms choose which sources to cite covers the selection logic that combines both signals.

FAQ

What word count does ChatGPT prefer for citations?

Evertune's analysis of 33,000 to 40,000 ChatGPT-cited URLs found the median cited page is approximately 941 words. AirOps' separate analysis of 16,851 queries found the optimal range is 500 to 2,000 words. Pages over 5,000 words underperform the shortest pages in citation frequency. The research points in the same direction: focused, dense content outperforms long-form pillar pages for ChatGPT citation.

How many external links should a ChatGPT-optimized page have?

Evertune found approximately 15 external links per cited page, pointing to primary research, named studies, and authoritative external sources. The external link count reflects whether a page is self-referential or genuinely grounded in third-party evidence. Pages that cite primary sources aggressively tend to get cited in return by AI models looking for well-grounded content.

Why do listicles get so many ChatGPT citations?

Evertune found 50% of ChatGPT citations go to listicle-format content, and 58% of those are ranked lists. Lists create discrete, extractable units that match how AI models parse and reconstruct answers from multiple source passages. Each list item is a self-contained passage. AirOps found the same pattern independently, with 74.2% of all AI citations going to list-format content across platforms.

Does content age affect ChatGPT citation rates?

AirOps found that pages 30 to 89 days old are cited at optimal rates. Pages updated in the last 30 days are too new for full indexing. Pages older than two years face a citation penalty. The practical target is a 60 to 90 day refresh cycle for pages intended to maintain ChatGPT citation presence, not a one-time publishing event.

Does schema markup increase ChatGPT citation rates?

AirOps found that pages with JSON-LD schema have a 38.5% ChatGPT citation rate versus 32% for pages without. FAQ schema specifically produced a 350% increase in AI citations in Otterly.ai's controlled experiment (2,379 citations versus 529 without schema). Schema is not the primary driver, but it consistently adds 5 to 10 percentage points to citation rates when the underlying content is already well-structured.

The structural brief is specific now

Before May 2026, GEO content advice was mostly directional: write for the reader, structure with headings, keep content focused. The Evertune dataset makes it specific.

A page that hits the 941-word target, uses four H2s, links to 15 external sources, and includes two embedded lists is structurally competitive with what ChatGPT already cites. A page that ignores all of these and publishes 4,000 words with few external references is not competing for citation, regardless of content quality.

The cross-validation from AirOps, the off-page experiment from Otterly, and the format data from Evertune together describe a content brief that is more actionable than anything available before this quarter. The structural target exists. The question is whether your publishing process is built to hit it.