Boost your skills with Growth Memo’s weekly expert insights. Subscribe for free!
For years, SEOs have operated on a simple assumption: The more ground your content covers, the more likely it is to surface in AI-generated answers. In fact, every “best practice” in classic SEO content pushes you toward more: more subtopics, more sections, more words. Build the “ultimate guide.”
An analysis of 815,000 query-page pairs across 16,851 queries and 353,799 pages says otherwise:
- Fan-out coverage is nearly irrelevant to citation rates.
- Two signals actually predict whether ChatGPT cites your page.
- Six concrete changes to your existing content library help.
1. The Study
AirOps ran 16,851 queries through ChatGPT three times each through the UI, capturing every fan-out sub-query, every URL searched, every citation made, and every page scraped. Oshen Davidson built the pipeline. I analyzed the data.
Each query generates an average of two fan-out queries. ChatGPT retrieves roughly 10 URLs per sub-search, reads through them, then selects which ones to cite. We scored how well each page’s H2-H4 subheadings matched those fan-out queries using cosine similarity on bge-base-en-v1.5 embeddings. That score is what we call fan-out coverage: the share of subtopics a page addresses at a 0.80 similarity threshold. (The 0.80 similarity threshold cutoff was used to decide whether a subheading counts as a match to a fan-out query. Think of it as a relevance bar.)
The question: Do pages with higher fan-out coverage get cited more?
You’ll find even more information in the co-written AirOps report.
2. Density Barely Moves The Needle
Across 815,484 rows, the relationship between fan-out coverage and citation is weak.
Covering 100% of subtopics adds 4.6 percentage points over covering none. That gap shrinks further when you control for query match (how well the page’s best heading matches the original query). Among pages with strong query match (>= 0.80 cosine similarity):
Moderate coverage (26-50%) outperforms exhaustive coverage. Pages that cover everything score lower than pages that cover a quarter of the subtopics. The “ultimate guide” strategy produces worse results than a focused article that covers two to three related angles well.
3. What Actually Predicts Citation
These two signals dominate: retrieval rank and query match.
1. Retrieval rank is the strongest predictor by a wide margin. A page at position 0 in ChatGPT’s web search results (the first URL returned by its search tool) has a 58% citation rate. By position 10, that drops to 14%. We ran each prompt three times consecutively for this analysis, and pages cited in all three runs have a median retrieval rank of 2.5. Pages never cited: median rank 13.
2. Query match (cosine similarity between the query and the page’s best heading) is the strongest content signal. Pages with a 0.90+ heading match have a 41% citation rate compared to the 30% rate for pages below 0.50. Even among top-ranked pages (position 0-2), higher query match adds 19 percentage points.
Fan-out coverage, word count, heading count, domain authority: all secondary. Some are flat. Some are inversely correlated.
4. The Wikipedia Exception
One site type breaks the pattern. Wikipedia has the worst retrieval rank in the dataset (median 24) and the lowest query match score (0.576). It still achieves the highest citation rate: 59%.
Wikipedia pages average 4,383 words, 31 lists, and 6.6 tables. They are encyclopedic in the literal sense. ChatGPT cites Wikipedia from deep in the search results where every other site type gets ignored.
This is density working as a signal, but at a scale no publisher can replicate. Wikipedia’s content is exhaustive, richly structured, and cross-linked across millions of topics. A 3,000-word corporate blog post with 15 subheadings is not the same thing.
5. The Bimodal Reality
58% of pages retrieved by ChatGPT in this dataset are never cited. 25% are always cited when they appear. Only 17% fall in between.
The always-cited and never-cited groups look nearly identical on most content metrics: similar word counts (~2,200), similar heading counts (~20), similar readability scores (~12 FK grade), similar domain authority (~54). The on-page signals we can measure do not separate winners from losers.
What separates them is retrieval rank. Always-cited pages rank near the top when they surface. Never-cited pages rank in the bottom half. The retrieval system, whatever signals it uses internally, is the gatekeeper. Everything else is a tiebreaker.
6. What This Means For Your Content
Conventional SEO content writing wisdom says cover more subtopics, add more sections, build density. The data says the conventional approach produces “mixed” pages, the 17% in the middle that get cited sometimes and ignored other times.
Mixed pages have the highest word counts, the most headings, and the highest domain authority in the dataset. They are the “ultimate guides.” They are also the least reliable performers in ChatGPT.
The pages that win consistently are focused. They:
- Match the query directly in their headings,
- Tend to be shorter (the citation sweet spot is 500-2,000 words), and
- Have enough structure (7-20 subheadings) to organize the content without diluting it.
Build the page that is the best answer to one question. Not the page that adequately answers 20.
Featured Image: Tero Vesalainen/Shutterstock; Paulo Bobita/Search Engine Journal