TL;DR
- ChatGPT now runs fewer, shorter searches. It averaged 2.15 fan-out queries per response in early December 2025 and 1.0 from April 2026, while average query length fell from about 117 to about 53 characters.
- It started searching inside specific websites. On August 8, 2026, the share of ChatGPT fan-out queries using the site: operator jumped from 0.37% to 16.8% overnight, and searches per response rose from about 1.08 to 1.83.
- It cites fewer sources. After GPT-5.3 rolled out on March 4, 2026, the average ChatGPT answer cited 4.7 to 4.9 sources, down from 6.4 the week before.
- Those changes move citations fast. Six days after the site: shift, Reddit's share of ChatGPT Search citations fell 86%, from 3.83% to 0.52%.
- Intent still shapes every search. In Promptwatch's original 2025 study of 5,513 prompts, 96% of informational prompts produced generic searches and 60% of commercial prompts produced competitor comparisons.
How ChatGPT search works, step by step
- ChatGPT decides whether to search. It searches automatically when a question would benefit from current information, or when the user asks it to.
- It rewrites the question into search queries. This is query fan-out: one prompt becomes one or more short, targeted queries. For example, a question like "What's the best CRM for a 10-person agency?" might become a query like "best CRM small agency 2026."
- It sends those queries to search providers. OpenAI says it works with third-party search providers, and also runs its own search crawler, OAI-SearchBot, to build an index.
- It fetches pages live when needed. A separate agent, ChatGPT-User, retrieves specific pages in the moment a user's question calls for them.
- It reads, synthesizes, and cites. ChatGPT writes one answer from the pages it retrieved and cites about five of them, with citation links in the answer.
The practical consequence: your page has to be reachable by OpenAI's crawlers, match one of those short sub-queries, and then be one of roughly five sources ChatGPT decides to cite.
What is query fan-out?
Query fan-out is the step where an AI system splits a user's prompt into one or more sub-queries and searches for each one before writing its answer. The sub-queries decide which pages can even be considered, so they matter more than the original prompt's wording. Promptwatch records the fan-outs behind every tracked prompt and re-runs them against a customer's own site to check whether its content can answer them.
How ChatGPT's searches changed from 2025 to 2026
Promptwatch first captured ChatGPT's search queries in June 2025, when ChatGPT started exposing them in response metadata. The same measurement, repeated over the following months, shows how much the search layer has moved:
| Period | What Promptwatch observed |
|---|---|
| June 2025 | 90% of prompts produced exactly 2 search queries; queries averaged 6.4 words and 48 characters (sample of 5,513 prompts) |
| Early December 2025 | 2.15 fan-out queries per response, about 117 characters on average |
| Early March 2026 | 1.84 fan-out queries per response, high-80s characters |
| March 4, 2026 | GPT-5.3 rollout: average citations per answer fell from 6.4 to 4.7–4.9 |
| April 2026 onward | 1.0 fan-out query per response, about 53 characters |
| August 8, 2026 | site: operator share of fan-outs jumped from 0.37% to 16.8%; searches per response rose from ~1.08 to ~1.83 |
| August 14–17, 2026 | Reddit's share of ChatGPT Search citations averaged 0.52%, down from 3.83% |
The June 2025 figures come from a one-off sample, and the later ones from Promptwatch's continuous ChatGPT query fan-out tracking, so treat the early comparison as directional. The direction is clear either way: ChatGPT moved from sentence-like searches toward short, keyword-style queries, then added searches targeted at specific domains on top.
ChatGPT now searches inside specific websites
On August 8, 2026, the share of ChatGPT fan-out queries using the site: operator, which restricts a search to one domain, jumped from 0.37% to 16.8% within a day, roughly a 46-fold increase. A brief dip to 0.15% on August 3 to 5 suggests a staged rollout.
Average searches per response rose from about 1.08 to 1.83 at the same time. The site-specific searches were added on top of general web searches rather than replacing them. For brands, this means ChatGPT increasingly looks things up directly on the sites it already has in mind, including yours, your competitors', and review platforms.
Six days later, Reddit's share of ChatGPT Search citations collapsed from a 3.83% average (July 18 to August 7) to 0.52% (August 14 to 17), an 86% drop. Google AI Mode (−30.5%) and AI Overviews (−11.3%) moved far less over the same weeks. The full story is in why ChatGPT stopped citing Reddit.
ChatGPT cites fewer sources per answer
ChatGPT cited 6.4 sources per web-search answer in the week of February 26, 2026. After GPT-5.3 rolled out on March 4, that fell to 4.7 to 4.9, about 27% fewer citation slots. The drop hit GPT-5.3, GPT-5.4, and GPT-5-Mini at once and had not recovered a month later.
That makes ChatGPT one of the tightest citation surfaces. In Promptwatch's average sources per response tracking, Google AI Overviews and Perplexity cite about 10 sources per answer, roughly twice as many.
What Promptwatch's original 2025 study found
The first dataset, collected in June 2025 in the days after ChatGPT started exposing its search queries, covered 5,513 real prompts that triggered a web search. It remains useful for one thing the later data confirms: the intent behind a prompt shapes the searches ChatGPT runs.
How many searches ChatGPT ran per prompt (June 2025):
- 2 queries: 4,981 prompts (90%)
- 1 query: 524 prompts (9%)
- 3 queries: 8 prompts (under 1%)
How long the queries were:
- Short (3 words or fewer): 2.6%
- Medium (4 to 7 words): 72.6%
- Long (8 words or more): 24.8%
- Range: 1 to 16 words, 7 to 117 characters
How intent shaped the query type:
| Prompt intent | Most common query type | Share |
|---|---|---|
| Informational | Generic (organic) searches | 96% |
| Branded | Brand-specific searches | 80% |
| Commercial | Competitor comparisons | 60% |
Informational prompts with generic searches made up about half of the dataset. Informational queries also ran slightly longer, about 7.1 words against 6.1 to 6.5 for other intents.
Brands get searched in context, not alone
ChatGPT rarely searched a brand name on its own. It embedded brands in context:
- Not "Tesla", but "Tesla Model 3 safety ratings" or "Tesla vs BMW electric vehicle comparison"
- For commercial prompts, patterns like "[Brand A] vs [Brand B] pricing comparison", "[Product] alternatives", and "[Service] reviews compared to [Competitor]"
If your brand doesn't appear in comparison and use-case content, it's missing from the searches ChatGPT runs for commercial questions.
"LLMs basically give you a summary of the whole online presence of your brand." Ilonka Karoly, Media Performance Director, Total Search at WPP Media
Example prompts and the searches ChatGPT ran
| Prompt | Search query 1 | Search query 2 |
|---|---|---|
| What's the best place in San Antonio to buy a certified pre-owned Mercedes-Benz? | certified pre-owned Mercedes-Benz dealership San Antonio | best place to buy certified pre-owned Mercedes Benz San Antonio |
| I'm looking for a high-tech SUV with great driver-assistance features. | best high tech SUVs driver assistance 2025 | high tech suv with advanced driver assistance features review |
| What are the best project management tools for small teams? | best project management tools small teams 2024 | project management software comparison small business |
| What are the key features to look for in accounting software? | best accounting software features small business | important features accounting software comparison |
| How do cloud storage services compare in terms of security? | cloud storage security comparison 2024 | best secure cloud storage services comparison |
| What's the difference between different types of business insurance? | types of business insurance explained | business insurance coverage comparison guide |
Notice the added year and the words "best", "comparison", and "review". ChatGPT reformulates a conversational question into the kind of query a search engine answers well.
Which crawlers ChatGPT uses to find pages
OpenAI runs three separate agents, and each one does a different job:
- GPTBot collects content to train future models. Blocking it doesn't remove you from ChatGPT search.
- OAI-SearchBot builds the index ChatGPT search draws on. OpenAI says a site must allow it, and allow traffic from OpenAI's published IP ranges, to be included.
- ChatGPT-User fetches a specific page live when a user's question calls for it.
OpenAI is still the largest source of AI crawler traffic: its crawlers made 79.8% of verified AI crawler requests in the week of August 31 to September 6, 2026, down from 94.8% in June. On one Promptwatch customer site, a real estate marketplace, ChatGPT-User made 28.5% of all AI crawler requests over 30 days and requested each property page about 2.5 times on average, a sign it keeps returning to the pages people actually ask about. There's more detail in what AI crawler bots do.
What ChatGPT cites once it has searched
The search queries decide what gets considered; the citation mix shows what wins. In Promptwatch's data:
- Product pages lead. They made up 32.8% of ChatGPT citations in July 2026, ahead of listicles (9.7%), news articles (5.2%), and how-to guides (4.1%).
- No domain dominates. Reddit was the most-cited domain in ChatGPT Search in June 2026 with 3.71% of citations, and no domain passed 4%.
- Video barely registers. Video made up 0.1% of ChatGPT citations in July 2026, far below Google AI Overviews.
"The amazing thing about AI is that it amplifies the mistakes. In the past, your UK website may not have been as good as your Dutch website, and you could get away with it. But because AI pulls from so many sources, and you have little control over where it pulls from, it became a serious issue." Iain Davenport, Director of Business Solutions at Monks
For more figures, see AI SEO statistics from Promptwatch's data.
Methodology
2026 data: Promptwatch's dataset of over 26 billion citations, prompts, and responses, collected from the live ChatGPT interface rather than the API, so it reflects what real users see. Fan-out and citation figures only include responses where ChatGPT ran a web search. Crawler figures only count requests whose IP matches the operator's published ranges.
2025 study: 5,513 real prompts that triggered a ChatGPT web search, collected in June 2025 in the first days after ChatGPT exposed its search queries in response metadata. Of 16,539 possible query slots, 10,510 (about 64%) were filled.
Updates: Promptwatch's data reports update continuously, so the live pages may show newer numbers than this article.
Frequently asked questions
How does ChatGPT search the web?
ChatGPT rewrites the user's question into one or more short search queries (query fan-out), sends them to its search providers and its own OAI-SearchBot index, fetches relevant pages, and writes an answer that cites a few of them, about 4.7 to 4.9 sources on average since March 2026.
Does ChatGPT use Bing?
OpenAI says ChatGPT search works with third-party search providers, including Microsoft, alongside its own crawler, OAI-SearchBot. Its public documentation doesn't describe the full provider mix, so the safest approach is to be crawlable by OAI-SearchBot and well indexed in the major search engines.
How many searches does ChatGPT run per answer?
About one to two. Promptwatch measured 2.15 fan-out queries per response in early December 2025, 1.0 from April 2026, and about 1.83 after ChatGPT added site-specific searches on August 8, 2026.
What is query fan-out in ChatGPT?
Query fan-out is how ChatGPT turns one prompt into one or more search queries before answering. Since April 2026 those queries have averaged about 53 characters, much shorter and more keyword-like than in 2025.
What does ChatGPT's use of the site: operator mean for SEO?
It means ChatGPT increasingly searches inside specific domains, not just the open web. Since August 8, 2026, about 16.8% of its fan-out queries use site:. Brands need clear, crawlable pages for every product, use case, and comparison, because ChatGPT may look for the answer on their own site.
How many sources does ChatGPT cite?
About 4.7 to 4.9 per web-search answer since GPT-5.3 rolled out on March 4, 2026, down from 6.4 the week before. Google AI Overviews and Perplexity cite about 10.
How do I make sure ChatGPT can find my website?
Allow OAI-SearchBot in robots.txt, don't block ChatGPT-User or OpenAI's IP ranges at your CDN, serve key content in the HTML, and give each topic a clear page that answers the short queries ChatGPT runs. Then track whether your pages get cited. The full step-by-step method is in our ChatGPT SEO guide.
Can I see the search queries ChatGPT runs?
Not reliably from the ChatGPT interface. Promptwatch captures the fan-out queries behind each tracked prompt, so you can see exactly what ChatGPT searched for before it answered, and whether your site answers those queries.
