- Ahrefs analyzed 66.6 million free, logged-out ChatGPT responses from August and September. Only 14.4% triggered a web search. [1]
- Responses that searched ran about 1.6 background queries on average, and 6.8% of those queries used the site: operator. [1]
- Nectiv ran about 4,000 prompts through a different model over the API and counted 7.61 queries a prompt, 64% of them site: searches. [2]
- The two samples differ in model, prompts, and collection method, so the numbers are not directly comparable. [3]
- Reddit appeared in 19.1% of US ChatGPT answers in March and 2.3% in September, in Ahrefs’ citation data. [3]
- OpenAI says a site must allow OAI-SearchBot to be eligible for ChatGPT search results. [4]
- Every figure here comes from a company that sells AI visibility tools or services. Treat each as one sample.
Two studies of how ChatGPT searches the web, published two months apart, appear to contradict each other. Ahrefs says ChatGPT searched for one answer in seven. [1] Nectiv says it now runs 7.61 searches per prompt, more than three times last year’s 2.17. [2]
Both can be true, because they did not study the same product. The difference between them is the useful part for anyone deciding where to spend on AI visibility.
What Ahrefs measured
Ahrefs collected 66.6 million responses from GPT-5.6 Luna in free ChatGPT sessions that were not logged in, across all countries. It also collected the 15 million fan-out queries those responses produced. [1] A fan-out query is a search the model writes for itself. OpenAI’s help page describes the step: when ChatGPT search works with partner search providers, it “typically rewrites your query into one or more targeted queries,” and may send more specific ones after reviewing the first results. [4]
The headline numbers are small. Of all responses, 14.4% searched, and those ran about 1.6 queries each. Of the 15 million queries, 6.8% used site: to look inside one website, spread across about 116,000 domains. [1]
Put those together and the picture gets smaller still. By our arithmetic, 15 million queries over 66.6 million responses is about 0.23 searches per answer, and 6.8% of 15 million is roughly one million site: searches, enough to touch at most 1.5% of the responses.
Ahrefs’ reason for studying free sessions is that most people use ChatGPT that way. It cites a figure from an interview in Search Engine Land, which we could not read at the source, that more than 90% of weekly users are on the free plan. [1]
Why Nectiv counted far more
Nectiv, an SEO agency for B2B companies, took about 4,000 prompts from a study it ran last year, balanced across verticals such as software, e-commerce, and travel, and ran them through GPT-5.6 Sol using OpenAI’s API. [2] Its results:
| Ahrefs [1] | Nectiv [2] | |
|---|---|---|
| Model | GPT-5.6 Luna | GPT-5.6 Sol |
| How it was reached | Free ChatGPT, logged out | OpenAI’s API |
| Sample | 66.6 million responses, all topics and countries | About 4,000 prompts in chosen verticals |
| Searches per prompt | About 1.6, counting only responses that searched | 7.61 on average, 29 at most |
| Share of searches using site: | 6.8% | 64% |
Nectiv’s highest average was software, at 10.7 searches a prompt. [2] Ahrefs offers the explanation itself: many other studies used commercial “best X” prompts that naturally trigger brand lookups, and free, logged-out sessions “search less than paid or reasoning models.” [1] Search Engine Journal draws the same line: the samples differ in models, prompts, and collection methods, so “their numbers aren’t directly comparable.” [3]
Two smaller studies point the same way, and both come with warnings from their own authors. WordLift, which sells structured data software, sorted 56 ChatGPT traces by the model that served them and found a web search in 10.8% of the free-like group and 47.4% of the paid-like group. It calls the sample small and “a directional signal, not a final verdict.” [5] Suganthan Mohanadasan, a co-founder of Snippet Digital, read the traffic of a single logged-in Plus account and reports that fan-outs in those captures dropped from 12 searches per answer to 4 in early August. [6]
Where the searches go
When free ChatGPT does look inside one site, it mostly picks the company that makes the thing. Among the roughly 160 domains it searched most with site:, 75.9% of the searches pointed at a brand site. Retailers such as Walmart, Costco, and Amazon took 11.9%, and Reddit and Facebook 1.9% between them. [1]
Ahrefs counts 9,442 site: searches for apple.com, 7,361 for ikea.com, 5,210 for nike.com, 4,619 for adidas.com, and 4,432 for samsung.com, with reddit.com ninth at 2,586. [1] Against roughly one million site: queries, even the leader is under 1%. The list is a long tail.
The wording of the queries moved the same way in a month. The word “official” appeared in 13.1% of English fan-out queries in September, up from 10.1% in August, and “price” went from 6.04% to 8.88%. [1] Nectiv saw “official” as the second most common added term in its own sample. [2]
What ChatGPT cites has shifted too
Ahrefs’ citation figures come from a separate dataset: US answers from March to September, 2.4 to 3.8 million a month. In March about 1 in 5 answers cited Reddit. By late September it was about 1 in 43. Wikipedia went from 1 in 7 to 1 in 35. [1] Search Engine Journal reports the same data as 19.1% falling to 2.3% for Reddit, and 14.1% to 2.9% for Wikipedia. [3]
The citations went to sources that produce information. PubMed climbed from 64th to 4th among the most cited domains and the FDA entered at 8th, while WebMD fell from 18th to 79th. [1][3] Publishers with OpenAI licensing deals more than doubled their combined share among the 200 most cited domains, from 7.5% in March to 15.8% in July, and held 12.5% in late September. [1]
Ahrefs checked that new prompts in its panel were not causing the shift, by comparing about 25,000 prompts present in both months. Every site in its chart moved the same direction. [1] One limit, in Search Engine Journal’s words: “Ahrefs’ citation ranks don’t measure traffic.” [3]
Two jobs, and which one is yours
Your presence depends on what the model already associates with your category. Mohanadasan’s captures show ChatGPT writing brand names into its first search before any result came back, and a brand named there reached the answer 68.9% of the time, against 2.1% for brands that were only fetched. One account, and the author calls every percentage directional. [6]
It goes looking for prices, specifications, and official pages, often on your own domain. Ahrefs’ advice is that those facts need to be on your site, in plain text, on a page it can find. [1]
Which job matters more depends on your buyers and your prompts. Ahrefs shows this with one category: for “running shoes,” ChatGPT’s searches used “best” 206 times and “review” 117 times, against 65 for “official,” which is a different pattern from its overall data. [1] An average across 66.6 million answers does not describe your category. It is the same caution we apply to every single-vendor number in generative engine optimization.
The one setting OpenAI documents
Whatever the true search rate is, eligibility is not in dispute. OpenAI’s help page says: “To make a website eligible for inclusion, allow OAI-Searchbot to crawl the site and confirm that the website host or content delivery network allows traffic from OpenAI’s published searchbot IP addresses.” It adds that placement is not guaranteed. [4]
The crawler documentation gives the consequence of blocking it. Sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers, though can still appear as navigational links.” [7] Search and training are separate switches, so a site can allow one and refuse the other, and a robots.txt change can take about 24 hours to reach search. [7]
User-agent: OAI-SearchBot Allow: / User-agent: GPTBot Disallow: /
What to check this week
- Read your robots.txt and your CDN or firewall rules for OAI-SearchBot. A block there removes you from search answers whatever else you do. [7]
- Put prices, specifications, and support facts on your own site as text, not only in images or scripts. [1]
- Ask ChatGPT your category question several times and read the searches it writes. Mohanadasan’s method uses the browser’s developer tools, and the list changed between runs. [6]
- Note whether your brand appears in those searches. If it never does, page fixes are not the gap.
- Report free and paid behavior separately, and name the model, when you quote a search rate.
- Repeat after each model change. Ahrefs calls its own figures “a snapshot of one model between August and September 2026.” [1]
Sources
- Ahrefs · How ChatGPT Does Query Fan-Out (15 Million Queries Studied)Vendor study, October 12, 2026. Ahrefs sells Brand Radar, the tool the data comes from.
- Nectiv · New Research: ChatGPT Tripled It’s Fan-Out Queries + Looks For Authoritative SourcesAgency study, August 13, 2026. About 4,000 prompts run through GPT-5.6 Sol over the API.
- Search Engine Journal · ChatGPT’s Most-Cited Sites Tilt Toward Primary Sources, Data ShowsTrade press, October 12, 2026. Reports percentages from Ahrefs’ charts.
- OpenAI Help Center · Searching the web with ChatGPTPrimary. Read October 13, 2026.
- WordLift · The New Outskirts of the Web: AI Free TiersVendor analysis of 56 traces, May 4, 2026. The author calls the sample small.
- Suganthan Mohanadasan · ChatGPT Already Knows Who’s In The Running Before It SearchesOne logged-in Plus account, 57 conversations, July 2026, updated August 17. The author sells a tool for reading fan-out queries.
- OpenAI · Overview of OpenAI CrawlersPrimary. Crawler documentation.
Frequently asked questions
What is a query fan-out in ChatGPT?
It is a search ChatGPT writes and runs in the background before it answers. OpenAI’s help page says that when ChatGPT search works with partner search providers, it typically rewrites your query into one or more targeted queries, and may send additional, more specific queries after reviewing the first results.
How often does ChatGPT search the web?
In Ahrefs’ sample of 66.6 million free, logged-out GPT-5.6 Luna responses from August and September 2026, 14.4% triggered a web search. That figure describes one model and one kind of session. Ahrefs says free, logged-out sessions search less than paid or reasoning models.
Why did Nectiv find 7.61 fan-out queries per prompt when Ahrefs found 1.6?
The setups differ. Nectiv ran about 4,000 prompts from verticals such as software, e-commerce, and travel through GPT-5.6 Sol over the API. Ahrefs watched the free consumer product on broad prompts from every topic and country, and its 1.6 is an average over only the responses that searched. Search Engine Journal notes the two numbers are not directly comparable.
How often does ChatGPT use the site: operator?
In Ahrefs’ data, 6.8% of 15 million fan-out queries used site: to search a single website, spread across about 116,000 domains. Nectiv reports a site: search in 64% of fan-out queries in its own sample. Among the roughly 160 domains ChatGPT searched most that way in Ahrefs’ data, 75.9% of the searches pointed at a brand site.
Did ChatGPT stop citing Reddit?
No, but it cites it much less in Ahrefs’ US data. Search Engine Journal reports the figures as 19.1% of answers in March 2026 and 2.3% in September. Reddit was still the ninth most searched domain in ChatGPT’s site: queries.
How do I make my site eligible to appear in ChatGPT search?
OpenAI says to allow OAI-SearchBot to crawl the site and to confirm that your host or content delivery network allows traffic from OpenAI’s published searchbot IP addresses. Sites that opt out of OAI-SearchBot are not shown in ChatGPT search answers, though they can still appear as navigational links. OpenAI adds that placement is not guaranteed.
Can I allow ChatGPT search and still block AI training?
Yes, according to OpenAI’s crawler documentation. OAI-SearchBot and GPTBot are controlled separately in robots.txt, so a site can allow the first to appear in search results and disallow the second to indicate that its content should not be used for training. A robots.txt change can take about 24 hours to take effect for search.
Mikołaj Salecki
Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.
