Does chasing query fan-outs get your brand named?

The advice, and the two questions the academic literature can answer

A widespread piece of content advice says to write pages that answer the extra questions an AI silently asks itself before it replies — what Google calls query fan-out, issuing several related searches across a topic before answering. The advice is to map every one of those hidden sub-questions onto your page.

This piece does not run a new measurement. It reviews 40 published sources spanning 1971 to 2026, each fetched against its primary text on ACL Anthology, arXiv, Google Patents, and the publishers' own pages, with the highest-risk claims — the patent identification and the core industry statistics among them — re-checked by two independent models from separate providers.

It asks two objective questions the academic literature can actually answer: is the underlying mechanism new, and does the evidence show that optimizing for it improves whether an AI names or cites a brand.

Fifty years of query expansion, not a new AI-search mechanism

The mechanism being marketed as a new AI-search idea is older than AI search. In 1971, Rocchio's relevance-feedback method defined the basic move: shift a query toward what a user has marked relevant, then retrieve again.

The query-expansion literature that followed built on that move — add related terms, issue more queries, aggregate what comes back. By 1998, Mitra, Singhal, and Buckley had named its failure mode “query drift”: expanding a query with the wrong terms pulls it away from what was actually asked, and retrieval gets worse.

Google's own 2018 ActiveQA-family work thought about the same aggregation problem — multiple queries synthesized and combined — in a neural setting. Google's own documentation confirms the shape: AI Overviews and AI Mode “may use a ‘query fan-out’ technique — issuing multiple related searches across subtopics and data sources — to develop a response.”

The engineering root of Google's version is a patent, US11663201B2, filed in 2018 and granted in 2023. No other AI provider has published comparable engineering detail, so the patent documents Google's own implementation specifically — not a mechanism confirmed to be identical across every AI system that expands a query before answering.

Industry blogs sometimes name a different Google patent, “Thematic Search,” as the source of fan-out; its own text never uses the term, and the link is a blogger's comparison rather than Google's description.

A 2026 preprint by Venktesh and colleagues runs the 1998 query-drift test on the reformulations that today's large language models generate, and lands in the same place: more reformulations do not reliably improve retrieval, and can degrade it as the irrelevant ones pile up. The throughline is a fifty-year-old idea — expand the query, retrieve more — that the retrieval literature already found has a ceiling, and past that ceiling, hurts.

What fan-out actually is

Stated plainly, Google's patent is a control loop. A trained model generates up to eight types of query variant from a single input query, and a control model decides at each step whether to keep generating or stop.

The fan-out is the set of variant queries; the system retrieves against them and aggregates the results before it answers. That is Google's documented mechanism specifically.

Other AI systems are not required to work the same way, and none has published comparable engineering detail. But the pattern shows up beyond Google: AirOps independently measured ChatGPT breaking a single prompt into several sub-queries before answering — the same general shape, observed rather than disclosed.

Whether covering more of those sub-questions on a page helps is a separate question, and the next section takes it.

Fan-out coverage barely moves the citation rate

The sharpest evidence on the coverage question is the industry's own best-designed study, AirOps' Fan-Out Effect report — serious, disclosed-methodology research drawn from 16,851 queries, 353,799 pages, and 50,553 ChatGPT responses.

Its coverage table has three rows: pages answering none of the hidden sub-questions, 79,024 of them, are cited 35.5% of the time; pages answering 26–50% of them, 28,785 of them, 38.2%; pages answering all of them, 120,572 of them, 34.0%.

Citation rate barely moves across the full range. The moderate-coverage middle sits highest, but the spread from worst to best across the whole table is narrow, and exhaustive coverage does not outperform no coverage.

The sample sizes are uneven: the all-coverage bucket holds more than four times the 26–50% bucket's pages, so the none and moderate-coverage comparison points rest on a smaller base than the all-coverage one — a limitation on the comparison, not a reason to drop the rows.

The same company's foundations course states that fan-out happens in 90% of ChatGPT queries and that about 15% of pages retrieved through fan-out earn a citation, in a register that implies chasing broad coverage helps. The flagship report's own coverage finding does not support that implication.

Why covering more fan-outs doesn't win more citations

The query-drift literature explains why. Blindly expanding a page to cover more of the AI's hidden sub-questions pulls it toward a broader, less focused answer, and away from precisely answering the one question that matters most.

Mitra, Singhal, and Buckley documented the retrieval version of this in 1998: expansion with the wrong terms moves a query off what was actually asked. Venktesh and colleagues found the same pattern in 2026 in the reformulations that large language models generate.

More coverage does not reliably convert to more citation, because the coverage that is added is coverage of questions the page was not primarily trying to answer.

Rank predicts citation — by attention, or just by being the better answer

The AirOps data show rank predicts citation too: pages retrieved first in fan-out are cited 58% of the time, against 14% for pages retrieved tenth. The literature is not settled on why position matters.

One reading is that the model pays more attention near the top of what it is given — the “Lost in the Middle” effect Liu and colleagues documented, and that Hsieh and colleagues showed can be calibrated away.

The rival reading, from a preprint by Cuconasu and colleagues, is more mundane: in real retrieval pipelines, distracting passages routinely surface in the top results, and both relevant and distracting passages take the same positional penalty, so the attention-bias effect is marginal.

On that reading, a page that ranks well and gets retrieved is more likely to actually be the best answer to the question — classic retrieval logic, not a claim about how the model weighs what it reads. The AirOps position numbers are consistent with either reading; they do not decide between them.

No head-to-head test, and a lopsided sample

No single study ran, as its own head-to-head experiment, content mapped to fan-out coverage against content not mapped to it. The finding here is a convergent read across studies that each measured one piece — retrieval rank, heading match, coverage ratio, the stability of keyword-density signals against signals of real expertise — not one paper that says exactly this.

The AirOps coverage comparison also rests on an uneven sample: the all-coverage bucket is more than four times the 26–50% bucket, so the two sparser rows are a weaker base than the dense one.

And AirOps' data is specific to ChatGPT and to ChatGPT's own internal retrieval position, not to Google's organic rank; the coverage finding is one vendor's disclosed methodology, not a replicated result.

Fan-outs are old, and covering them doesn't help

The evidence supports two findings, stated as findings rather than as instructions. The fan-out mechanism is not new: it is a fifty-year-old query-expansion idea whose ceiling the retrieval literature identified in 1998 and re-confirmed for LLM reformulations in 2026.

And the evidence available does not show that building content to cover more of an AI's hidden sub-questions increases how often that content gets used, retrieved, or attributed: in the industry's own best-designed study, citation rate barely moves across the full coverage range, from none to all.

What the evidence does support is older and less proprietary — being genuinely findable, ranking well, and being genuinely good — the same levers classic retrieval always rewarded. The AI's hidden question set is a description of how the system searches, not a new surface to optimize against.