For a decade, content strategy has had one default blueprint: the pillar page. Build one comprehensive guide, "The Ultimate Guide to Running Shoes," point every internal link at it, and let that single URL rank for every query in its topic. It was good advice for Google. We wanted to know whether AI engines reward the same architecture.
So across 500 queries we recorded every URL that Google AI Overview, ChatGPT, and Perplexity cited, 6,083 unique pages. Then, for each page, we counted how many different queries earned it a citation.
The pillar model predicts a set of hub pages cited across dozens of related questions. That is not what the data shows. It isn't close.
// FINDING_1
Finding 1: AI Citation Is Radically Specific
Of the 6,083 pages AI engines cited:
| Cited for | Pages | Share |
|---|---|---|
| Exactly 1 query | 5,451 | 89.6% |
| 2 queries | 406 | 6.7% |
| 3 to 4 queries | 185 | 3.0% |
| 5 to 8 queries | 41 | 0.7% |
| 10 or more queries | 0 | 0% |
Nine out of ten pages that win an AI citation win it for exactly one question. The dream of a page that owns its entire topic doesn't survive this table. Across 500 queries, the single widest reach any page achieved was 8.
// FINDING_2
Finding 2: Even the Champions Cover Almost Nothing
It's worth looking at who those top performers actually are, because they're precisely the pages the pillar model says should dominate.
CNET's best-VPN guide is one of the most-cited pages in the whole study. Institutionally backed, constantly updated, category-defining. It was cited for 8 queries out of 500. SAMHSA's national helpline page, a US government mental-health resource: 7. Carfax's auto-repair hub: 7.
These are among the strongest content assets on the internet in their categories, and each one covers between 1% and 2% of the query space. The most any single page can reach is far lower than most content teams assume.
Why so low? Because AI engines build answers per question, not per topic. For "best running shoes for flat feet," the engine wants the page about flat feet, not the page about running shoes in general. Every extra query an answer page reaches for makes it a slightly worse fit for each individual one. Breadth doesn't get rewarded. Fit does.
// WRONG_CONCLUSION
The Wrong Conclusion, and the Data That Blocks It
At this point a tempting idea shows up. If citation is one page per query, then mass-produce pages. Spin up a template, generate five hundred thin variants, collect five hundred citations. The data blocks that road too.
We compared pages by how many queries they were cited for:
| Cited for | Pages measured | Median lists | Median headings | Median external links | Schema rate |
|---|---|---|---|---|---|
| 1 query | 4,399 | 14 | 16 | 32 | 54.6% |
| 2 to 4 queries | 464 | 18 | 17 | 47 | 55.4% |
| 5+ queries | 29 | 25 | 20 | 42 | 58.6% |
Structure rises with citation breadth. The pages that win multiple queries carry more lists, more headings, more outbound references, and more schema than the single-query winners. And even the single-query winners are substantial: the median cited page has 16 headings, 14 lists, and 32 external links. That is not template output. A thin programmatic page, near-zero references and a flat structure, resembles nothing in the winners' table.
(Honest caveat: only 29 of the 5-plus-query pages returned measurable HTML, so treat that row as directional. The consistent pattern is that multi-query winners beat single-query pages on every metric; external links happen to peak in the 2-to-4 bucket.)
So both extremes lose. The 8,000-word everything guide loses because engines match per question. The 500 thin template pages lose because engines cite pages with real substance. What wins sits in the narrow band between them.
// FORMULA
The Formula: A Deep Page for Every Question
Put the two findings together and the content architecture writes itself. One deep page per specific question. Not one giant page for all the questions, and not a thousand shallow pages splitting one.
For a team that currently maintains a single pillar guide on a topic, this is what that looks like in practice.
- Split by question, not by keyword volume. To an AI engine, "best running shoes" is not one topic. Flat feet, wide feet, marathon training, beginners, plantar fasciitis: each is a separate citation contest with a separate winner. Map the distinct questions your buyers actually ask. That map is your page list.
- Fund each page like it has to survive fact-checking. The winning profile across this whole study is consistent: sectioned structure (the median multi-query winner carries 20 headings), generous outbound references (30 to 50 external links is normal for cited pages), schema markup, and real coverage of the one question the page exists to answer.
- Let the pillar page retire into a hub. A short page that routes readers to the deep pages still earns its place for navigation and internal linking. Just stop expecting it to be the thing AI engines quote. In our data, they don't.
- Measure per question, not per page. If your dashboard tracks "is our guide cited," it's asking a question the engines stopped answering. Track which specific queries you hold a citation for, and treat every lost query as a missing page, not a page that needs to get longer.
The pillar era rewarded breadth. This one rewards precision at depth. The teams that adjust first will be building a page for every question their buyers ask while everyone else keeps polishing one guide that the engines have quietly stopped reading.
Which questions do you hold a citation for?
Run a free AEO audit and see which buyer queries cite your pages today, and which are missing a page entirely.
Run the Free Audit// METHODOLOGY
Methodology
Dataset. 500 queries, 100 in each of five intent categories (informational, commercial, transactional, YMYL, local), US market, English, collected April 2026. Citations captured for Google AI Overview, ChatGPT (web search enabled), and Perplexity, deduplicated to 6,083 unique AI-cited URLs.
Breadth measurement. For each cited URL, the count of distinct queries for which any AI engine cited it. Maximum observed: 8.
Depth measurement. Structure signals (headings, lists, external links) parsed from raw HTML saved during a crawl of all URLs; schema extracted from JSON-LD. The depth-by-breadth comparison uses the 4,892 AI-cited URLs with HTTP status 200 and parseable HTML.
// RELATED
More From the AEO Study
AI Search Isn't One Channel. It's Three.
94.2% of AI-cited pages appear in exactly one engine.
Part 1The Pages That Win Everywhere
The rare dual winners cite sources at twice the typical rate.
Part 2Every AI Engine Plays by a Different Rulebook
ChatGPT's cited pages carry schema at 81%. Google's at 48%.