How we reviewed this
What kind of study this is
This is a literature and industry-claim review. ContentGrapher spent no measurement budget. We did not run a new probe. We read the public dataset against the retrieval literature.
The data comes from AirOps. AirOps published a disclosed-methodology study called “The Fan-Out Effect.” We audited their numbers and placed them in the academic record.
The AirOps dataset
AirOps scraped the ChatGPT interface, not the API. They sent 16,851 unique queries. They ran each query three times. This produced 50,553 responses.
They scraped 353,799 pages. They scored 815,484 coverage rows. They recorded 1,511,251 fan-out detail rows. The queries span 10 industry categories and four query types.
One model was tested. The data is ChatGPT-only. No other LLM was in the panel.
How coverage was scored
A sub-question is a fan-out query. ChatGPT issues these sub-searches internally. AirOps counted every fan-out query ChatGPT ran.
Coverage is the share of sub-questions a page covers. AirOps scored coverage against the page headings. They embedded the H1 to H4 headings with the BAAI/bge-base-en-v1.5 model.
A page covers a sub-question when a heading matches the query at or above 0.80 cosine similarity. They also tested 0.60 and 0.70. The pattern held at both.
A citation is a URL that appears in the answer. A page appears once per response, regardless of coverage. Density does not earn a page multiple citations in one answer.
The coverage bands
The AirOps report groups pages into three bands at the primary 0.80 threshold. There is no finer breakdown. We report the three bands exactly as AirOps published them.
The bands are uneven. The 100% band holds 120,572 pages. The 26 to 50% band holds 28,785. The sparse band is a weaker base for comparison.
How we verified the numbers
Every number in this study traces to the AirOps source HTML. We fetched the report directly. We did not take it from a blog paraphrase.
The load-bearing claims were re-checked by two models from separate providers. GLM 5.2 and Grok 4.5 each confirmed the four core AirOps statistics. This meets the two-validator rule.
The academic anchor
We grounded the mechanism in the retrieval literature, not in AirOps. Mitra, Singhal, and Buckley named query drift in 1998. They showed blind expansion can make retrieval worse.
Karpukhin and colleagues defined a passage as a 100-word segment in 2020. Lewis and colleagues used the same design for the original RAG architecture. Both treat the page boundary as irrelevant.
These sources explain why coverage breadth does not drive citation. A passage-level retriever sees one chunk, not the aggregate page.
What this review cannot claim
- Causal proof that writing to sub-questions fails. The data is observational. No one wrote new pages to test the advice.
- Generalization beyond ChatGPT. One model was scraped. Gemini, Claude, and Perplexity may select sources differently.
- Generalization to Google organic rank. AirOps measured ChatGPT's internal retrieval position, not Google's SERP rank.
- A replicated result. One vendor ran one study. No head-to-head test of the writing advice exists.
- Even coverage bands. The 100% bucket is more than four times the size of the 26 to 50% bucket.