Ahrefs: ChatGPT fans out 15 million queries to answer
Ahrefs' study of 66.6 million ChatGPT answers shows most run on multiple hidden queries, meaning brand visibility hinges on sub-questions users never type.
Key takeaways
- Only 14.4% of ChatGPT responses studied ran on a single retrieval query; the rest used query fan-out.
- Ahrefs logged 15 million fan-out sub-queries across 66.6 million GPT-5.6 Luna responses in August and September.
- Brands optimised for one canonical topic phrase can be invisible if they never surface in the model's generated sub-queries.
- Narrow, specific pages are more likely to match fan-out queries than broad thought-leadership content.
- Topic clusters built around likely sub-questions now matter more than single definitive pages.
One question in, several out. Ahrefs reports that across 66.6 million free, logged-out GPT-5.6 Luna responses gathered in August and September, only 14.4% triggered a single retrieval query. The rest set off what the industry calls query fan-out: ChatGPT silently rewriting a user's prompt into a cluster of related searches, sometimes a dozen or more, before composing an answer. The study logged 15 million of these sub-queries. The headline number is not the 14.4%. It is what happens to the other 85.6%, because that is where most brand visibility is won or lost.
The mechanics of being asked about sideways
Fan-out means the words a user typed are rarely the words that determine whether a brand gets cited. Ask ChatGPT "best CRM for a 50-person startup" and the model may not search that phrase at all. It may generate variants: "CRM pricing for small teams," "HubSpot vs Pipedrive comparison," "CRM with lowest per-seat cost 2025." Each spawns its own retrieval and its own shortlist of sources. The final answer is a synthesis across that whole cluster, not a response to the literal query. A brand optimised to rank for the original phrase can be invisible in the answer if it never surfaces in the three or four derivative searches the model actually ran.
This is the uncomfortable part for anyone still thinking about AI visibility the way they think about Google rank-tracking. There is no single keyword to own. There is a constellation of adjacent, intent-shifted queries, generated at inference time, that changes response to response even for near-identical prompts. Ahrefs' own framing is that this makes classic one-to-one keyword targeting close to obsolete for this channel. That is a strong claim, and the data supports it: with fewer than one in six answers running on a single query, optimising for the literal prompt is optimising for the minority case.
Why financial services and multilaterals feel this first
Consider how this plays out for a bank's asset-management arm, or a multilateral like CGAP publishing on financial inclusion. These organisations tend to write for the canonical question: "What is financial inclusion?" "How does microfinance reduce poverty?" Fan-out means ChatGPT may never retrieve against that canonical phrasing at all. It may instead split the question into "financial inclusion statistics 2025," "World Bank microfinance evaluation," "informal lending vs microfinance outcomes," and build its answer from whichever sources best answer each slice. An institution with one authoritative explainer page, however well it ranks for the parent topic, can lose the citation to a competitor or a secondary source that happens to answer the specific sub-query better, even if that source is thinner on the main topic.
The same logic applies to industrial groups. A company like Holcim publishing on cement decarbonisation is not just competing to be the best answer to "sustainable cement." It is competing across whatever fan-out produces: "cement CO2 emissions per tonne," "carbon capture cement technology," "EU ETS impact on cement producers." Each sub-query is a separate retrieval contest with its own winners. Breadth of coverage, not depth on one flagship page, is what gets cited across the cluster.