Ahrefs: ChatGPT searched the web in 14.4% of 66.6M responses

Ahrefs published a study of how ChatGPT retrieves information on 12 October 2026, written by Louise Linehan and reviewed by Ryan Law. It analyzes 66.6 million ChatGPT responses and their 15 million fan-out queries. This report restates the findings, each with the base Ahrefs gives, so that no figure is read against the wrong group.

Ahrefs is a vendor of SEO and AI visibility software, and its post points readers to its own Brand Radar and Ask Ahrefs tools. The study is a vendor’s data, and the whole of it is Ahrefs’ reporting, not ours.

For what fan-out is and how Google describes it, see our report on Google’s R4T research.

What did Ahrefs study?#

Ahrefs analyzed 66.6 million free, logged-out GPT-5.6 Luna responses and their 15 million fan-out queries, from August and September 2026. It chose free sessions because, citing Olivier de Segonzac of RESONEO in Search Engine Land, more than 90% of ChatGPT’s weekly users are on the free plan. Logging out also strips chat history and memory, so the data shows default behavior.

Ahrefs says its prompts are broad and stem from real search demand across every topic and country. The post does not say how the responses were collected. The citation figures further down come from a different dataset, Brand Radar’s ChatGPT citations in the US from March to September 2026, covering 2.4 to 3.8 million answers a month.

Finding	Figure	Base
Responses that required a web search	14.4%	66.6 million responses
Fan-out queries per searching response	about 1.6	Responses that searched
Fan-outs using a site: operator	6.8%	15 million fan-outs
Domains those site: searches covered	about 116,000	Fan-outs using site:
Common site: searches aimed at a brand site	75.9%	Most-searched sites
Reddit citations	about 1 in 5 to about 1 in 43 answers	Brand Radar, US, March to late September

How often does ChatGPT search?#

Ahrefs reports that only 14.4% of the 66.6 million responses triggered a web search, which it summarizes as 6 in 7 prompts answered without searching. When ChatGPT did search, it ran about 1.6 fan-out queries per response.

Ahrefs reasons that every search costs OpenAI money, so skipping one when the model knows enough is sensible. It cites WordLift research suggesting free-tier style models may lean more on training data. Ahrefs concludes that for many prompts, visibility depends on what ChatGPT already knows about a brand, not on what it can find.

Which sites does it search?#

Of the 15 million fan-outs, 6.8% used the site: operator to search a single website. Among the roughly 160 most-searched sites, Ahrefs says 75.9% of those searches targeted a brand site, such as apple.com or ikea.com, rather than a retailer, government site, forum or review site. Retailers such as Walmart, Costco and Amazon made up 11.9%, and Reddit and Facebook together 1.9%.

Consumer brands fill 8 of the 10 most site:-searched domains. The chart shows the top ten by mentions, from about 1 million site: queries behind the ranking. Ahrefs notes the ranking reflects its own dataset and prompts.

Government domains also feature: gov.br and gov.uk sit in the top 15, and Ahrefs says the NHTSA, the FDA and the SEC are in the top 150. Of the 17 publishers and platforms with OpenAI licensing deals, Reddit is the only one in the top 150 most site:-searched domains.

Ahrefs also ran a separate test of 20,000 US prompts. For factual questions about pricing, product features or company details, 84.4% cited at least one official source, against 68.3% for comparison questions. By citations, 63.4% went to official websites for factual questions and 46.4% for comparison questions.

Which words are growing?#

Ahrefs counted the words inside fan-out queries. Its top-20 table lists “2026” first, then “official” and “price”. The table shows each word’s share of fan-outs in the two months.

Word	August	September	Change
official	10.1%	13.1%	+3.0 pts
price	6.0%	8.9%	+2.8 pts
best	5.0%	5.3%	+0.3 pts
reviews	2.7%	3.6%	+0.9 pts
meaning	2.9%	2.3%	-0.6 pts

Ahrefs says a gain of 1 or 2 points is small in absolute terms, but these words started from a low base. “Price” went from 6.04% to 8.88% of fan-outs, a relative rise of nearly half (47%) in one month. “Reviews” grew by a third and “specifications” entered the top 50 for the first time. “Meaning” and “definition” each fell by about a fifth, which Ahrefs reads as likely a dip rather than a trend.

In an example for “running shoes”, Ahrefs says Brand Radar returned 1,474 US ChatGPT responses with fan-out queries. “Best” appeared 206 times, “2026” 160 times, “review” 117 times, “vs” 105 times and “official” 65 times. Ahrefs reads that as third-party reviews and comparisons mattering more than brand pages for that topic.

Which sites does it cite?#

Ahrefs reports that ChatGPT’s citations moved too, in its separate Brand Radar data:

• Reddit: about 1 in 5 answers cited it in March and about 1 in 43 by late September, which Ahrefs headlines as an 88% drop in six months.
• Wikipedia: from about 1 in 7 answers to about 1 in 35.
• YouTube: from about 1 in 370 answers to about 1 in 500.
• OpenAI partner titles: among ChatGPT’s 200 most-cited domains, their combined share went from 7.5% in March to 15.8% in July, and was still 12.5% in late September.

Ahrefs says the lost citations went to research bodies, regulators, brands and hands-on testers. PubMed climbed from 64th to 4th among the most-cited domains and the FDA entered at 8th, while health sites that summarize research, such as WebMD and Verywell Health, fell. To rule out new prompts as the cause, Ahrefs compared about 25,000 prompts present in both March and September, and says every site it charted moved in the same direction.

How does it compare?#

Ahrefs lists five studies. It says they agree on direction, toward official sources and site: searches, but differ on scale, because others found far more fan-outs and site: searches per prompt. These rows are Ahrefs’ summary of each study, not our check of them.

Study	Setup	Fan-outs per prompt	site: share
Ahrefs	GPT-5.6 Luna, free, logged out, 66.6M responses	about 1.6	6.8% of fan-outs
Chris Long, Nectiv	GPT-5.6 Sol via API, about 4,000 prompts	7.61	64% of queries
Wil Reynolds	GPT-5.6 Luna, small tracked subset	not given	76% of fan-outs
Peec data	GPT-5.6 vs 5.5, sample not disclosed	56.5% of prompts had 2+	about 23% of fan-outs
Suganthan Mohanadasan	Logged-in, 57 conversations, SaaS-heavy	15 to 40 on Thinking	Frequent on commercial prompts

Ahrefs explains the gap by its broad prompts and free, logged-out sessions, which search less than paid or reasoning models.

How reliable is the study?#

These are the limits, partly stated by Ahrefs and partly our reading of the post.

• One model, one snapshot. Ahrefs calls the numbers a snapshot of one model between August and September 2026 and says OpenAI changes how ChatGPT retrieves information constantly.
• Free, logged-out only. Ahrefs notes these sessions search less than paid or reasoning models, so the 14.4% is not a rate for ChatGPT as a whole.
• Collection not described. The post does not say how the 66.6 million responses were gathered or how its prompts were chosen, and the raw data is not published.
• Vendor study. Ahrefs sells Brand Radar and Ask Ahrefs, and the post promotes both.
• Two datasets. Fan-out figures come from the 66.6 million responses, citation figures from Brand Radar’s US data over March to September. They have different bases and dates.
• Small shifts, short window. The word changes span two months, and Ahrefs itself calls the decline of “meaning” and “definition” likely a dip rather than a trend.
• Inconsistent counts. The post refers to both about 160 and the top 150 most-searched sites.

What should you take from it?#

This section is our own reading, not Ahrefs’ conclusion.

• Do not read the 14.4% as a universal rate. It describes one free model over two months, and Ahrefs notes that others measured far more searching per prompt.
• Put facts on your own site in plain text. Ahrefs’ reasoning is that if ChatGPT runs site: searches for prices, specs and support details, those facts need to be findable on a page it can reach.
• Cover the sub-questions. As in Google’s own fan-out research, pages are retrieved for the narrower queries they answer. The AI answer readiness checker scores whether a page answers a question directly, and it measures structure only.
• Compare with other platforms before generalizing. Citation behavior differs by engine, as in our report on AI Overview citation source overlap.
• Read the wider topics. See AI search and content strategy for related coverage.