Which pages does ChatGPT search for on brand websites?

A fan-out query is a search ChatGPT Search writes for itself while it answers a prompt. A site: lookup is a fan-out query that starts with site: and a web address, so the search engine returns pages from that one website only. A typical lookup reads "site:sentry.io node.js error monitoring crashes alerts Node.js".

Of 155,899 fan-out queries, 50,701 were site: lookups. That is 32.5%, or about one in three. 29.6% of answers ran at least one. Answers that ran any ran 4.0 on average.

Across eight days, 12,749 answers included site lookups covering 8,156 websites.

We matched each lookup to the brands named in the same answer by their website. A brand counts as looked up when a site: lookup targets its website. Own site cited means at least one citation in that answer points to the brand's website.

In answers that ran lookups, the 44,095 looked-up brand mentions had their own site cited 85.4% of the time. The 35,868 brands named in the same answers without a lookup had their own site cited 45.7% of the time. That is 1.87 times more often. In answers with no lookups at all, the rate was 59.4%.

This is an association, not proof that the lookup causes the citation. ChatGPT may look up the brands it already plans to feature. But in the same answer, to the same prompt, the looked-up brands are the ones whose own pages back the answer.

In answers that ran lookups, the chance that a named brand got one falls with the order in which the answer names it. ChatGPT looked up 79.0% of first-named brands, 64.6% of third-named brands, and 21.3% of brands named 11th or later.

Looked-up brands were named first in 22.5% of their answers. Brands named in the same answers without a lookup were named first in 7.4%. The lookup and the top of the answer go together.

To see what pages ChatGPT searches on brand websites, we read the words after the site: address and sorted them into page types. One lookup can carry more than one type. Docs counts docs, help, and developer subdomains too.

56.1% of lookups name no page type at all. They pair the website with words from the buyer's need, such as "site:monday.com cross team visibility dashboards portfolio project management". Docs and setup words appear in 15.8% of lookups. ChatGPT pricing page lookups make up 10.1%. Security and compliance follow at 5.1%, integrations at 4.8%, and the API at 4.6%. Reviews (0.6%) and comparisons (0.4%) are rare.

On average, 48.9% of the words after the site: address come from the prompt itself. Only 5.4% of lookups use prompt words alone. The rest add words the buyer never typed.

The most common added words are category nouns such as management. After those come pricing (added to 2,862 lookups), 2026 (2,307), official (1,413), features (1,186), API (1,051), integrations (845), and docs (831).

Across all answers, ChatGPT looked up 18.2% of named brand mentions. In answers that ran any lookup, it looked up 55.1%. Of 35,129 distinct brands named, 7,420 got at least one lookup. That is 21.1%.

The lookup rate is the share of answers naming a brand that also looked up its website. Datadog was looked up in 692 answers, 34.2% of the answers that named it. Its own site was cited in 81.7% of those answers and in 48.8% of the answers that named it without a lookup. The pattern repeats for every brand in the table. Most of the top 15 are software brands. The tracked prompts lean toward software buyers.

A lookup without a citation is a site: search whose target site is not cited in the answer. This happened in 30.8% of the 51,029 lookups.

Failure depends on whose site it is. 76.7% of lookups targeted the website of a brand the answer names, and 16.5% of those failed. The other 23.3% targeted sites the answer does not name as a brand, and 77.8% of those failed.

We expected pricing and docs lookups to fail more often, because many sites hide prices and split docs onto other domains. The data does not support that. On the sites of named brands, pricing lookups failed 17.5% of the time and docs lookups 16.3%, close to the 16.3% for lookups that name no page type.

Product spec lookups failed most, at 24.9% of 354. Reviews failed 20.6% of 175 and features 19.3%. Shipping and policy lookups failed least, at 8.8%. Lookups on docs, help, and developer subdomains were cited 82.2% of the time, about the same as lookups on a brand's main domain (83.4%).

184 brands got 40 or more lookups. For half of them, their own site was cited in at least 89.4% of the answers that looked them up. Only 3 were cited in under half.

ChatGPT searched these sites but rarely cited them. Check whether the pages answer the question clearly and whether important details are easy to find.

11,912 lookups targeted a website that does not belong to any brand named in that answer. Some are review and community sites: reddit.com (123 lookups, 86% cited), g2.com (99, 88%), and github.com (81, 94%). Others are brands that ChatGPT checked and then left out of the answer, such as oracle.com and esri.com.

betteruptime.com got 80 lookups and 0 citations. It is the former address of BetterStack, which now uses betterstack.com. A lookup on a retired domain returns nothing ChatGPT will cite.

An official query names a brand and adds the word official, such as "Cloudflare CDN WAF DDoS protection official". 40,026 fan-out queries contain official. That is 25.7%, about one in four. Only 1,413 of them also use site:.

When an official query names a brand from the answer, that brand's site is cited 84.7% of the time, against 74.2% for queries that name the brand without official or site:. At the answer level, 85.9% of citations in answers with official queries point to a named brand's own site, against 65.4% in answers with neither query type.

Answers with these queries are mostly vendor shortlists, so part of the gap comes from the prompt, not the word.

We placed each prompt in its market on the public Parse index. Only 10.4% of answers in this window belong to a public market, so this cut covers 4,455 answers, and each market below has at least 44.

Uptime & Synthetic Monitoring Tools answers ran a lookup 65.9% of the time. Skincare and Face Makeup Products answers ran one 6.5% of the time. Software markets with pricing questions, such as Cloud FinOps Management Platforms (47.6%), sit near the top. Consumer product markets sit near the bottom.

Parse records ChatGPT Search's own fan-out queries from September 29, 2026. Before that date, the record held the prompt text, so earlier site: lookups are not visible to us. On October 2, the record held the prompt again, and we drop that day from the chart.

On a panel of 7,016 prompts answered both from September 29 to October 1 and from October 3 to October 6, the share of answers with a lookup rose from 27.2% to 31.5%. The share of lookups followed by a citation from the site held at 69.2% and 69.4%.

Earlier ChatGPT answers we recorded from January to April 2026 (gpt-5-2, gpt-5-mini, gpt-5-3) ran 2,477,104 fan-out queries, and 299 of them used site:. That is 0.01%. Those answers came from a different ChatGPT surface, so the jump does not prove a change of model behavior.

Two outside sources disagree on when this started. Promptwatch reported that ChatGPT site: fan-outs rose from 0.37% to 16.8% on August 8, 2026, and DEJAN found no rise when sampling the OpenAI API. On the ChatGPT Search answers Parse records, the share was 32.5% from September 29, about twice the Promptwatch figure.

The corpus is 48,251 ChatGPT Search answers from September 29 to October 6, 2026, all on gpt-5-6. We dropped every fan-out query that equals the prompt after lowercasing and removing punctuation. 5,215 answers had no other query and were excluded, leaving 43,036 answers and 155,899 fan-out queries.

A site: lookup is counted once per website in the query, so 50,701 queries give 51,029 lookups. Websites are compared at the registrable domain (docs.datadoghq.com counts as datadoghq.com). A brand's website is its primary site on the public Parse index, plus any domain in this corpus whose name equals the brand name (slack.com for Slack). We could match 96.7% of brand mentions to a website.

Page types come from keyword rules on the words after the address, so they undercount pages named in other words. The headline was recomputed independently in SQL with primary sites only: 85.8% for looked-up brands and 44.2% for brands named without a lookup, a ratio of 1.94. The main figures use the broader website match and give 85.4% and 45.7%, a ratio of 1.87.

The window is eight days, because ChatGPT Search fan-out queries were not recorded before September 29. Prompts are Parse's tracked prompts, which lean toward software buyers. Figures describe this data cut, not live ChatGPT behavior.

ChatGPT Search checks the shortlist on the brands' own websites before it answers. A lookup on your site is usually followed by a citation of your site, and the brands it looks up are the brands it names first.

Write pages that answer the buyer's need in the buyer's words. Most lookups pair your domain with the need itself, not with a page type. A page that states plainly that your product does the thing asked for is the page the lookup finds.

Cover the pages ChatGPT searches for: documentation, pricing, security, compliance, integrations and API details. Keep dated facts current and redirect retired domains so readers can reach the right page.