Ahrefs looked at 66.6 million answers from the free version of ChatGPT and found that 14.4% of them triggered a web search. When the model did search, it went more often to the site that owns the fact. Reddit’s share of citations fell 88% in six months.
Table of contents
Key takeaways
- In 66.6 million free, logged-out GPT-5.6 Luna answers from August and September 2026, 14.4% ran a web search, at about 1.6 queries each.
- Only 6.8% of those 15 million queries were limited to a single site. That works out to about one such query per 64 answers.
- In a separate US citation set, Reddit fell from roughly 1 answer in 5 in March to 1 in 43 in late September.
- The study gives no repeat runs and does not publish its prompts. Treat the shares as one vendor’s measurement, not a constant.
What this is about
On October 12, 2026, Louise Linehan of Ahrefs published a large study of how ChatGPT searches. Most free answers never touch the web. When ChatGPT does search, it leans on brand sites, regulators and research bodies. For a site owner, a search you are not part of is now the rare case. The usual case is an answer with no search at all.
What the source says
The study looks at query fan-out. That is the step where the model breaks a prompt into its own search queries before it answers. Ahrefs logged those queries for free, logged-out sessions on GPT-5.6 Luna, the default ChatGPT model. Logging out removes chat history and memory, so the answers show default behaviour.
Linehan’s summary is that ChatGPT “seems to be moving from sites that summarize information to the sources that produce it”. She gives the model two ways into an answer. A brand can be known well enough that ChatGPT never looks it up. Or it can be the original source the model finds when it does look.
The study also draws on a second dataset, Ahrefs’ own record of which domains US ChatGPT answers cite. That data covers March to September 2026 and 2.4 to 3.8 million answers a month. The fan-out queries and the citations come from different samples. They should not be read as one funnel.
Ahrefs sells the tracking product behind both datasets. Part of the article explains how to use it. We leave that part out.
What the data shows
The fan-out figures, in Ahrefs’ own words where the wording matters:
- 14.4% of 66.6 million answers triggered a search. Searching answers ran about 1.6 fan-out queries each, for roughly 15 million in total.
- 6.8% of those queries used the site: operator, which limits a search to one domain. They spread across about 116 000 domains.
- Among the roughly 160 domains searched most often this way, 75.9% of searches went to a brand’s own site. Retailers took 11.9%, and Reddit and Facebook 1.9% together.
- Apple’s domain led with 9 442 site: searches, then IKEA with 7 361 and Nike with 5 210.
- “Price” rose from 6.04% of queries in August to 8.88% in September. “Official” went from 10.1% to 13.1% of English queries.
A separate test of 20 000 US prompts split questions by type. For factual questions about price, features or company details, 84.4% of answers cited at least one official source. For comparison questions, 68.3% did. Official sites took 63.4% of all citations on factual questions and 46.4% on comparisons.
The citation data tells the second half of the story:
| Domain or group | March 2026 | Late September 2026 | Change |
|---|---|---|---|
| Reddit, answers citing it | about 1 in 5 | about 1 in 43 | −88% |
| Wikipedia, answers citing it | about 1 in 7 | about 1 in 35 | about −80% |
| YouTube, answers citing it | about 1 in 370 | about 1 in 500 | about −26% |
| OpenAI partner publishers, share of citations | 7.5% | 12.5% (15.8% in July) | up |
Source: Ahrefs, US ChatGPT citations, March to September 2026, 2.4 to 3.8 million answers a month. The Wikipedia and YouTube percentages are our arithmetic from Ahrefs’ ratios.
PubMed rose from 64th to 4th among the 200 most-cited domains, and the FDA entered at 8th. Health explainers such as WebMD and Verywell Health fell. Prompt sets change from month to month, so Ahrefs reran the comparison on about 25 000 prompts present in both March and September. Every domain in its chart moved the same way on those identical prompts. That check is the strongest part of the study.
Against the seven method points we ask of vendor studies, this one names the model, the interface, the dates and the regions. It does not publish the prompt set, beyond saying it comes from real search demand in every country. It reports no repeat runs and no measure of run-to-run variation. Linehan calls the results a snapshot of one model, which is fair.
What it does not mean
It does not mean ChatGPT runs site: searches most of the time. Our earlier piece on Luna’s direct searching reported Seer Interactive’s figure: 76% of Luna’s searches limited to one site. Ahrefs finds 6.8%. Both can be right. Seer ran 346 matched prompts that lean commercial. Ahrefs ran broad prompts in every country.
There is also a second 76% in this study, and it measures something else. The 75.9% is the brand-site share among the most-searched site: targets. It is not a share of all searches.
| Study | Setup | Queries per prompt | site: share |
|---|---|---|---|
| Ahrefs | Luna, free, logged out, 66.6M answers, all countries | about 1.6 (searching answers) | 6.8% of queries |
| Seer Interactive | Luna vs GPT-5.5, 346 matched prompts | similar to GPT-5.5 | 76% of searches |
| Chris Long, Nectiv | GPT-5.6 Sol via API, about 4 000 prompts | 7.61 | 64% of queries |
| Peec | GPT-5.6 vs 5.5, sample not disclosed | 56.5% of prompts had 2+ | about 23% |
Sources: Ahrefs, October 12, 2026, including its comparison of other studies; Seer Interactive as covered on this site, September 2026.
It does not mean Reddit stopped mattering. Reddit is still the ninth most site: searched domain in the fan-out data. Suganthan Mohanadasan of Snippet Digital, quoted by Ahrefs, found that ChatGPT still fetches Reddit for real users’ opinions while citing it less. The drop shows up in citations. Our report on Reddit in Google’s Discussions and Forums feature covers a different engine, where Reddit still dominates.
It does not mean partner deals buy citations. The partner share doubled from March to July and then fell back to 12.5%. Ahrefs offers explanations from David McSweeney and Mohanadasan about pre-loaded publisher content. Those are observations from outside the system, not a mechanism OpenAI has confirmed.
It does not tell you how often your brand is searched. The 14.4% is an average across every topic and country. A commercial category may search far more often. Ahrefs’ own running-shoes example found queries led by “best” and “review”, not “official”.
This would be falsified by a repeat run on the same prompts showing a very different search rate. That is the test to look for in the next study.
What to do
1. Put your prices, specs and support facts on your own site in plain text. When ChatGPT runs a site: search on a brand, those are the queries it runs. A PDF or a “contact sales” page does not answer them. 2. Publish what only you can publish. Original data, hands-on testing and first-hand product facts gained citations in this data. Summaries of other people’s work lost them. 3. If your AI visibility runs through Reddit or Wikipedia, check how often ChatGPT still cites them on your topics. A source that was in one answer in five in March was in one in 43 by late September. 4. Measure your own category rather than borrowing the average. Run your core prompts several times a week, logged out, and record whether a search happens. Our guide to how answer engines cite explains why a single run proves nothing.
This week, read the three pages ChatGPT is most likely to search on your domain: pricing, specs and support. Ask whether each one answers the question in its first paragraph.
FAQ
What is query fan-out?
It is the step where an AI assistant turns one prompt into several web searches of its own before it writes the answer. In this study, ChatGPT ran about 1.6 such searches for each answer that searched at all.
Why did Ahrefs test logged-out free sessions?
Linehan cites an estimate that more than 90% of ChatGPT’s weekly users are on the free plan. Logging out also removes memory and chat history. The trade-off is that free sessions search less than paid or reasoning models, so the 14.4% is likely a floor for heavier users.
Is the Reddit drop permanent?
Nobody can say. The data ends in late September 2026, and Linehan notes that OpenAI changes retrieval constantly. Snippet Digital saw an 86% drop in August, so two independent sources agree on the direction. Neither says how long it will last.
