Covering more topics does not earn more citations
We reviewed the largest public dataset on AI citation and coverage. AirOps scraped 16,851 queries and 353,799 pages from ChatGPT. We read their numbers against the retrieval literature.
No. Pages that cover all the sub-questions get cited 34.0% of the time. Pages that cover 26 to 50% get cited 38.2%. More coverage does not earn more citations.
Pages with full coverage earn fewer citations than pages with moderate coverage. AirOps reads this as a generalist signal. Their gloss says exhaustive coverage may signal generalist content that addresses many topics without depth.
Source: AirOps, “The Fan-Out Effect” (primary sim ≥ 0.8). Each rate is the share of pages in that band that ChatGPT cited. The 100% band does not beat the 26–50% band.
What “covering the sub-questions” means
When you ask ChatGPT a question, it runs several hidden sub-searches. Google calls this query fan-out. Each sub-search is one sub-question.
Coverage is the share of sub-questions a page covers. AirOps scored coverage against the page headings. A close heading match counts as covered.
Most queries fan out small. AirOps found 88.6% of queries generate exactly two sub-questions. Only 2.5% generate four or more. The “cover everything” target is usually two questions, not dozens.
The spread is narrow
The full coverage range moves citation by 4.2 percentage points. That is the gap between the best band and the worst. AirOps calls fan-out coverage a weak signal.
Their report says coverage adds 4.6 percentage points uncontrolled. The signal disappears when query match is held constant.
The whole coverage range moves citation by 4.2 percentage points. AirOps calls fan-out coverage a weak signal.
Coverage is the weakest signal
Two other signals move citation far more. Retrieval position moves it 44 percentage points. A page at position 1 gets cited 58% of the time. A page at position 10 gets cited 14%.
A strong query-heading match lifts citation to 41%. A weak match sits at 29%. Coverage is the weakest of the three.
The spread between the best and worst case shows how much each signal moves citation. Coverage moves it least.
Source: AirOps, “The Fan-Out Effect.” Coverage spread is the 34.0%–38.2% gap. Position spread is position 1 vs position 10. Heading-match spread is weak vs strong.
Why coverage does not drive citation
AI retrieval reads one passage at a time, not the whole page. Karpukhin and colleagues split text into 100-word passages in 2020. The retriever matches one passage against one sub-question.
The rest of the page stays invisible to that match. A page that covers many topics does not help one passage answer. This is why coverage breadth does not drive citation.
The retrieval literature warned about breadth 28 years ago. Mitra, Singhal, and Buckley named query drift in 1998. Blind expansion pulls the query away from the real intent. More expansion can make retrieval worse, not better.
What this means for content creators
The advice to answer every hidden sub-question does not hold. Covering more topics does not earn more citations. The data points the other way.
Focus and depth beat exhaustive breadth. A page that develops one idea fully earns more citations than a generalist page. The strong-heading-match signal is a focus signal. ContentGrapher measures a related property: concept integration.
How this reads against Clearscope-style advice
Clearscope tells you to cover what ranks. The hidden-sub-question advice shares the same premise. Both say more coverage helps. The data says it does not.
ContentGrapher is internally anchored, not competitor-anchored. It asks what concepts this audience needs, at what depth. It does not benchmark your page against what ranks. The fan-out data is the external evidence for that stance.
What this does not mean
This review does not say coverage is useless. Pages with zero coverage still get cited 35.5% of the time. Coverage is a weak signal, not a null signal.
The data is ChatGPT-only. AirOps scraped the ChatGPT interface. ChatGPT uses Bing search under the hood. Google AI Overviews may select sources differently. Our AIO citation study tested Google directly and found a separate null.
The finding is one vendor's disclosed methodology. No one has replicated it. The findability study shows structural quality still matters for retrieval. This study narrows what kind of quality matters: depth, not breadth.
What we cannot claim
- 01Causal proof that writing to sub-questions fails. The data is observational. AirOps measured pages as they existed. No one wrote new pages to test the advice.
- 02Generalization beyond ChatGPT. The study scraped the ChatGPT interface. No other model was tested. Gemini, Claude, and Perplexity may behave differently.
- 03Generalization to Google organic rank. AirOps measured ChatGPT's internal retrieval position. That is not Google's SERP rank. The two are related but separate.
- 04A replicated result. One vendor ran one study. No head-to-head test of the writing advice exists. The convergent read across the retrieval literature supports the direction, not a single experiment.
- 05Even coverage bands. The 100% band holds 120,572 pages. The 26 to 50% band holds 28,785. The sparse band is a weaker base for comparison.
The answer
Covering more of the AI's sub-questions does not earn more citations. Pages with moderate coverage cite at 38.2%. Pages with full coverage cite at 34.0%. The real levers are retrieval position and heading match.
Both reward a focused page that answers one question well. Build for depth, not for breadth. That is what ContentGrapher measures, and what the data rewards.
References
AirOps. (2026). The fan-out effect: What happens between a query and a citation. AirOps report. https://www.airops.com/report/the-fan-out-effect-what-happens-between-a-query-and-a-citation
Google. (2025). AI features and your website. Google Search Central. https://developers.google.com/search/docs/appearance/ai-features
Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W. (2020). Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (pp. 6769–6781). Association for Computational Linguistics. https://arxiv.org/abs/2004.04906
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459–9474. https://arxiv.org/abs/2005.11401
Mitra, M., Singhal, A., & Buckley, C. (1998). Improving automatic query expansion. In Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 206–214). ACM. https://doi.org/10.1145/290941.291006