Ahrefs’ study of ChatGPT query fan-out counts 15 million fan-out queries in 66.6 million free, logged-out responses, and the Reddit drop reported alongside that count comes from a second dataset. The post, published October 12, draws on two data sources and also reports a separate 20,000-prompt test and one control subset, all under one title, and each headline number is a share of something different. Below, each figure sits next to its base.
What is query fan-out?
In ChatGPT, query fan-out is the set of web searches the model runs before it answers a prompt: Search Engine Journal describes fan-out queries as “background searches ChatGPT performs before giving an answer”. Ahrefs’ post uses the term without a formal definition. In its sample of free, logged-out GPT-5.6 Luna responses, 14.4% of the 66.6 million responses required a web search, and those that searched ran about 1.6 fan-out queries on average.
One post, several bases
The fan-out sample covers August and September, all countries. The Reddit and Wikipedia figures come from what Ahrefs calls “a separate dataset”: Brand Radar’s ChatGPT citations in the US from March to September, 2.4 to 3.8 million answers a month. The official-source figures come from “a separate test of 20,000 US prompts”, whose chart credits the same fan-out tracker as the fan-out charts. Neither the text nor the charts name a model or access mode for that test or for the citation data. This is a vendor study built on the vendor’s own product data: Brand Radar is Ahrefs’ own tool.
| Figure | Its base |
|---|---|
| 14.4% | 66.6M free, logged-out Luna responses: those that required a web search |
| About 1.6 | Fan-out queries per response that searched (15.0M queries from 9.6M searches) |
| 6.8% | Of 15 million fan-out queries: used the site: operator |
| 75.9% | Site: searches to the ~160 most-searched domains: aimed at a brand site |
| 13.1% “official” | English fan-out queries in September (10.1% in August) |
| Reddit 19.1% → 2.3% | US answers citing Reddit, March vs September |
| Partners 7.5% → 15.8% → 12.5% | Citations among ChatGPT’s 200 most-cited domains (US): March, July, September |
| 84.4% | Factual questions in the 20,000-prompt test: cited at least one official source |
How the site: figures nest
By our arithmetic, 6.8% of 15.0 million fan-outs is about 1.0 million site: queries, which matches Ahrefs’ “around 1 million site: queries”. The 75.9% covers only the ~160 most-searched domains, out of ~116K different domains. The fifteen domains in the post’s top-domain table add up to 55,767 mentions by our arithmetic, under 6% of that million. The post does not say what share of all site: queries the ~160 domains cover, or how many responses contained a site: query.
Ahrefs glosses the 6.8% as “(i.e. 6.8% of 14.4%)”. The 6.8% is a share of fan-out queries and the 14.4% is a share of responses. Multiplied together, by our reading, they give a share of neither.
Ahrefs reports that “official” rose from 10.1% to 13.1% of English fan-out queries between August and September, and “price” from 6.04% to 8.88%. In the vocabulary table that lists both words, “india” moves from 3.0% to 6.2%, “australia” from 1.4% to 4.3% and “philippines” from under 0.8% to 3.2%. The post describes a same-prompt control for its citation ranks and none for the fan-out sample. It does not say whether August and September drew on the same prompts or the same mix of countries.
What the Reddit comparison carries
Reddit went from 19.1% to 2.3% of US answers, and Wikipedia from 14.1% to 2.9%, between March and September in Brand Radar. That is the share of answers citing a site. It is not a count of citations and not traffic. Search Engine Journal makes the same point about the rank data: “Ahrefs’ citation ranks don’t measure traffic.”
Ahrefs ran a control: “~25,000 prompts that appeared among the 150,000 highest-volume prompts in both March and September”, and wrote that “Every site in the chart moved in the same direction on those identical prompts.” That sentence follows the rank chart of sixteen domains. Reddit, Wikipedia and YouTube are not among them. The post does not say whether 19.1% → 2.3% was recomputed on the shared prompts, and it gives no size for any move on them.
The post names no model for the March answers. The fan-out sample is Luna, and OpenAI said on August 6 that Luna would become the default chat model for ChatGPT’s Free and Go tiers, as we reported on August 9. The post does not relate that change to its citation data. Ahrefs states no cause for the drop and offers a reading of the wider pattern: “ChatGPT seems to be prioritizing original sources over sites that summarize them.”
Partner share: “doubled” by July
A section heading in the post says partner citation share “has doubled”. The chart title is more specific: “OpenAI’s publisher partners doubled their ChatGPT citation share by July”. The values are 7.5% in March, 15.8% in July and 12.5% in September. The September figure is about 1.7 times March by our arithmetic, after two monthly declines from July. This is a different kind of share again: citations among the top 200 domains, not answers.
Why fan-out numbers differ
Ahrefs’ comparison table puts site: share at 6.8% of fan-outs in its data, against 64% of queries (Chris Long, Nectiv, GPT-5.6 Sol via API, about 4,000 prompts), 76% of fan-outs (Wil Reynolds, Luna, a small tracked subset) and about 23% (Peec data). We did not read those sources. Ahrefs says many other studies “used commercial ‘best X’ prompts that naturally trigger brand lookups.” Search Engine Journal says of Ahrefs’ and Nectiv’s samples that they differ in “models, prompts, and collection methods”. Two ghost-citation studies we covered in July disagreed for the same kind of reason: different samples and engine mixes.
What a marketer can use
Ahrefs says its fan-out numbers are “a snapshot of one model between August and September 2026”, and that free, logged-out sessions “search less than paid or reasoning models”.
The fan-out vocabulary describes what the model searched for. Ahrefs says the same of site: searches, which “show what ChatGPT looks up during query fan-out, but not what it ends up citing.” A share of answers citing Reddit is likewise not traffic. Ahrefs’ advice, which is advice and not a finding, is to keep prices, specs and support details on your own site, in plain text, on a page the model can find.
Three questions fit any fan-out study. Which model and access mode produced the sample? Is each percentage a share of responses, queries, answers or citations? Were the same prompts used in both periods?