How ChatGPT uses site: searches in its fan-out

Published 12 November 2026. Updated 12 November 2026.

A site: search is a fan-out query that ChatGPT narrows to one website, such as site:abondance.com seo. Published counts of how often ChatGPT writes one run from 0.14% to 64% of its queries, because each study used a different model, prompt set and way of collecting data, and in every study that names the sites, most of them belong to the brand the question is about.

A site: search is one of the ways query fan-out turns a single prompt into several searches. Below we line up every published count we could find with how it was measured, show the site: searches in one of our own recordings, and set out what each kind of target means for your pages.

What is a site: search in ChatGPT?

A search limited to one domain, URL or path. Google's documentation says the operator asks for results "from the particular domain, URL, or URL prefix specified in the operator" (Google Search Central). ChatGPT writes it into its own search. The person asking never types it.

A restriction to one site can sit in three places, and only two of them are ChatGPT's own choice.

Where it sits	Who sets it	Example
site: in the query text	The model	site:abondance.com seo, from a GPT-5.4 Thinking search reported by RESONEO
The domain field at the end of a web search line	The model	fast|Zendesk AI agents pricing 2026|30|zendesk.com, a line format we cover in ChatGPT's four kinds of fan-out query
A domain filter in OpenAI's API	The developer	Up to 100 allowed domains, subdomains included, set in the request (OpenAI web search guide)

The third row matters when you read a study. An API filter is a setting the researcher chose, so a count taken with one measures that setting. Check whether a study set a filter before you compare its figure with what ChatGPT does in the app.

How often does ChatGPT run site: searches?

Anywhere from almost never to most of the time, depending on who measured it and how. These are the six counts published between July and October 2026.

Study	Setup	Share with site:
DEJAN, 25 August 2026	196,692 queries from GPT-5 models through OpenAI's Responses API, fixed prompts run daily, March to August 2026	6.43% of queries in March, 0.14% in August, 0.15% on GPT-5.5
Ahrefs, October 2026	15 million queries from 66.6 million free, logged-out GPT-5.6 Luna responses, August and September 2026	6.8% of queries
Promptwatch, via Nadia Mohamed, August 2026	The live ChatGPT interface, sample size not published	0.37% of responses on 7 August, 16.8% on 8 August
MJ Cachón, 25 August 2026	1,797 queries from 189 brand prompts for three clients, GPT-5.5 through the API	30.2% of queries
Peec (David Konitzny), July 2026	ChatGPT 5.6 Luna fan-outs, sample size not published	About 43%, against 0.004% on ChatGPT 5.5
Nectiv, 2026	About 4,000 prompts spread across sectors, GPT-5.6 Sol through the API	64% of queries

Why do the site: counts disagree so much?

Four things change between the studies, and each one can move the figure by tens of points.

• The prompts. Cachón's prompts all named a brand, and 30.2% of the queries used site: on GPT-5.5. DEJAN's mixed prompts got 0.15% on the same model. Inside DEJAN's own data, long one-off prompts ran at 6.29% and short daily tracking prompts at 0.44%.
• The model. Both Peec and DEJAN put GPT-5.5 close to zero, at 0.004% and 0.15%. DEJAN found 5.88% on GPT-5.2, and the three studies on GPT-5.6 report 6.8%, 43% and 64%.
• The source of the data. Three studies used OpenAI's API, Ahrefs used free logged-out sessions and Promptwatch sampled the live interface. DEJAN notes it has not run its prompts through the ChatGPT interface.
• The unit. Promptwatch counts responses with at least one site: query. The others count queries, and a response with ten queries and one site: search scores 100% on the first measure and 10% on the second.

The 8 August jump shows how far apart the setups are. Promptwatch saw site: responses go from 0.37% to 16.8% that day. DEJAN's API capture recorded 0 site: queries out of 371 on the same date. OpenAI has announced nothing about it. There is no single site: share for ChatGPT, so the useful number is the one for your own prompts.

What did ChatGPT's site: searches look like in our recording?

In a ChatGPT turn we recorded with Only Fan Out on 14 September 2026, all three of its searches were narrowed to a single website with site:. We read them in the live answer stream. The saved copy of the same chat kept the search steps but none of the query text, which is the change our post on ChatGPT's missing fan-out queries covers.

Those three searches sat in a turn that fetched 78 pages in all and cited 15 of them. Most site: queries name one domain. Only 9% of Peec's named two or three, and DEJAN's averaged 1.13 domains, with up to 10 chained together with OR.

Which sites does ChatGPT search with site:?

Mostly the brand the question is about, then Reddit, review sites and public bodies. The three studies that named the sites agree on the order.

Study	Brand's own sites	Other sites
Peec, ChatGPT 5.6	84% of the domains it searched	Reddit was 71% of the rest. Trustpilot was 54% of the review and comparison sites
Ahrefs, about 1 million site: queries	75.9% of searches on the 160 or so most searched sites, led by apple.com, ikea.com and nike.com	Retailers 11.9%, Reddit and Facebook 1.9% together; gov.uk and gov.br in the top 150
MJ Cachón, brand prompts	55.1% of site: searches went to the brand's main domain and 26.2% to a group or parent domain	Third parties 18.4%; competitors 2 queries out of 1,797

Peec also saw a pattern in how deep the search goes. A whole domain came with broad words such as review, official and comparison, while a specific URL path came with product and pricing queries. Cachón found site: four times as common in the last search of a run as in the first, and 64.4% of brand prompts led to at least one site: search on the brand's own domain.

What should you change when ChatGPT searches your site?

Put the fact it is checking on a page of your own that search engines have crawled, written in plain HTML. A site: search only returns pages already in the index behind it, so an uncrawled pricing page or a fact held in a PDF or behind a script cannot come back.

Where the search went	What to check
Your whole domain, with words like official or review	A clear page that answers the question in the query, linked from your home page
A path on your domain, such as pricing or a product	That path exists, is indexed, and states the price or spec in text
Reddit	Real threads where people talk about you. Do not post fake ones
Trustpilot and other review sites	Your profile is claimed and your recent reviews are visible
Government or research sites	Your page cites the same public source for any fact you share with it

A site: search in Google is a weak check on whether your page is indexed. Google says the results are "not always exhaustive" and that a page missing from them may still be indexed, so use the URL Inspection tool in Search Console. None of the studies above measured how often a site: search led to a citation.

How do you find site: searches on your own domain?

Record your buyers' questions live and look for your domain in the queries. Since 14 September 2026 ChatGPT sends query text only while the answer streams, so a chat saved earlier will not show them.

1. Add Only Fan Out to Chrome and click its toolbar icon on a ChatGPT tab to open the side panel.
2. Add your domain to the watchlist. It highlights every mention across queries, fetched sources, cited sources and answers, so a site: query that names you stands out.
3. Ask the questions your customers ask, one per chat, and let each answer finish.
4. Search History for site: and your domain to pull up every recorded conversation whose queries used them.
5. Export the queries to CSV and count how many name your domain, a competitor, Reddit or a review site.
6. Ask the same questions again on another day. The queries change between runs, so group them by what they look for, as our comparison of fan-out queries vs keywords sets out.

Do Claude and Grok run site: searches too?

Grok does, and Claude rarely or never does on the published data. Peec counted a site: search in 18.3% of Grok chats, and nine in ten of its Reddit searches used site:reddit.com (Peec, April 2026). DEJAN found none in 33,872 Claude queries taken through Anthropic's API between May and August 2026, and 198 in 321,131 Google queries, or 0.062%.

Related questions