When I talk to business owners about ChatGPT, most of them picture it like this: you type a question, ChatGPT “looks it up,” and it hands back an answer. Many also believe that if their page ranks for the exact question a customer types, ChatGPT will find it. That is only half true.
What actually happens is more interesting. When ChatGPT decides it needs fresh information, it rewrites your prompt into one or more short web searches, runs them, reads the results and then writes its answer. Those hidden searches are called fan-out queries. They are often shorter, plainer and more “keyword-like” than the prompt itself, and they often include extra words you never typed, such as the current year, a brand name or the word “official.”
In this guide, I will explain how ChatGPT’s fan-out works, what new data from Ahrefs says about it, how you can see those queries yourself, and how to turn them into a practical content list. If you want the Google side of the story, I cover sub-questions in AI Mode in my companion piece on writing content for query fan-out in Google. Here, I stay with ChatGPT only.
What happens between your prompt and ChatGPT’s answer
Let me explain the steps in plain words. I treat this as a simplified model, because OpenAI does not publish the full details.
- ChatGPT reads your prompt and decides whether it needs the web. Many questions, like “explain compound interest,” can be answered from what the model already knows. No search happens.
- If it needs fresh or specific facts, it writes search queries. These are the fan-out queries. A long, chatty prompt becomes one or more short searches.
- It sends those queries to its search systems and gets back a list of pages.
- It fetches and reads some of those pages. OpenAI’s documentation says its
OAI-SearchBotcrawler is the one that surfaces sites in ChatGPT search, whileChatGPT-Uservisits pages when a user asks something. - It writes the answer and shows some of the pages it used as citations.
The important point for you is step 2. Your page doesn’t compete for the user’s prompt. It competes for the rewritten queries. If your page doesn’t match those queries, it may never reach step 4.
OpenAI’s own crawler page is clear on the first condition:
“OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features.”
— OpenAI, Overview of OpenAI Crawlers
So before anything else, check that your robots.txt does not block OAI-SearchBot. OpenAI says search-related robots.txt changes can take about a day to apply. If you blocked it by accident, no amount of content work will help. I explain the full index side in my post on how ChatGPT’s search index finds pages.
How often ChatGPT searches, and how many queries it runs
The biggest public dataset I have seen on this comes from Ahrefs. In How ChatGPT Does Query Fan-Out (15 Million Queries Studied), published in October 2026, Louise Linehan’s team looked at 66.6 million free, logged-out ChatGPT responses collected in August and September 2026. Logged-out means no chat history or memory, so it shows ChatGPT’s default behavior.
Here is the headline number that surprised many marketers:
“Only 14.4% of the 66.6 million ChatGPT responses we tracked triggered a web search.”
— Ahrefs, How ChatGPT Does Query Fan-Out
In other words, most answers in that sample came from the model’s own knowledge, not from a live search. When a search did happen, Ahrefs counted about 15 million fan-out queries from 9.6 million searching responses. That is roughly 1.6 queries per searching answer on average.
Other studies report very different counts, and that tells you something important. Ahrefs summarizes several of them:
| Study (as reported by Ahrefs) | Setup | Fan-outs per prompt | Share of site: queries |
|---|---|---|---|
| Ahrefs Brand Radar | 66.6M free, logged-out responses, Aug–Sep 2026 | ~1.6 per searching response | 6.8% |
| Nectiv (Chris Long) | ~4,000 prompts via API, a different model tier | 7.61 | 64% |
| Peec | Sample size not disclosed | 56.5% of prompts had 2+ | ~23% |
| Wil Reynolds | Small tracked subset | Not stated | 76% |
Why the big spread? Different models, paid versus free access, API versus the app, and different prompt types. A complex comparison prompt fans out much more than a simple fact question. NoGood’s query fan-out guide says complex prompts can produce “8-15 sub-queries,” though it doesn’t share its sample size. Ahrefs’ help center says most prompts it tracks return about two fan-out queries, with simple prompts returning one or none.
In brief: expect a handful of short searches for commercial or comparison questions, and none at all for many simple ones. I treat any single average as direction, not hard science.
How long are the fan-out queries?
None of the sources I read publish a reliable average length. What they all show, through their examples, is that fan-out queries look like classic search keywords: three to eight words, no greetings, no “can you help me.” A prompt like “I run a small bakery and need simple accounting software that handles payroll, what should I use?” tends to become something like “best accounting software small business payroll 2026.” That example is mine, for illustration, but it matches the pattern the studies describe.
The words ChatGPT adds: years, “official,” prices, brands and places
This is the most useful part of the Ahrefs data for content teams. They tracked which words appeared in fan-out queries from August to September 2026. Here is a summary of the shifts they reported:
| Word or pattern | August | September | What I take from it |
|---|---|---|---|
| “2026” (current year) | 17.6% | 22.9% | ChatGPT wants fresh pages. Dated, updated content has an edge. |
| “official” | 10.1% | 13.1% | It looks for the source of truth, usually the brand’s own site. |
| “price” | 6.0% | 8.9% | Pricing pages and clear cost info matter more. |
| “reviews” | 2.7% | 3.6% | Review pages and third-party reviews still get searched. |
| “best” | 5.0% | 5.3% | Steady. “Best of” lists remain part of the mix. |
| Country names (e.g. India 3.0% → 6.2%) | — | — | ChatGPT localizes searches more often. |
Ahrefs also found that 6.8% of fan-out queries used the site: operator, which restricts a search to one website. About three in four of the site: searches to the most-searched sites pointed at brand websites, like apple.com or nike.com, rather than retailers or forums. In a separate test of 20,000 US prompts, 84.4% of fact-based answers cited at least one official source.
Let me translate that into plain advice. When someone asks ChatGPT about your product, there is a good chance it will go straight to your website to check facts like price, features and specs. If those facts are missing, buried in a PDF or out of date, ChatGPT will take them from someone else. I wrote about keeping these facts in one place in my guide to a brand facts page.
Why definition-style searches are falling
Ahrefs also reported that searches containing “meaning” and other definition-style queries dropped by about a fifth. My reading is simple: the model already knows definitions, so it doesn’t need to search for them. If your content strategy depends on “what is X” articles, those pages are less likely to be fetched by ChatGPT. Pages with current facts, prices, comparisons and first-hand detail are more likely to be.
How to see ChatGPT’s fan-out queries yourself
You don’t have to take anyone’s word for this. You can look at the queries for your own prompts. Here are three ways, from free to paid.
Method 1: browser developer tools (free)
This sounds technical, but it takes about five minutes. NoGood’s guide describes these steps, and they work in Chrome:
- Open ChatGPT in Chrome. Right-click anywhere on the page and choose Inspect.
- Click the Network tab in the panel that opens.
- Type a prompt that is likely to trigger a search, such as “best CRM for a 10-person real estate team in 2026.”
- In the Network panel, filter by Fetch/XHR.
- Click the requests that appear while the answer loads, and open the Response tab.
- Press Ctrl+F (Cmd+F on Mac) and search for
search_model_queries. The text next to it lists the searches ChatGPT ran.
A word of caution: OpenAI changes its interface often, so the field name can change. If search_model_queries doesn’t show up, try searching the response for queries or for a word from your prompt.
Method 2: Chrome extensions
Several free extensions in the Chrome Web Store claim to pull these queries out for you. They save time, but they read your ChatGPT session, so check the permissions and the publisher before installing one. I would not use one on an account that holds client data.
Method 3: AI visibility tools
Paid tools collect fan-out queries at scale. Ahrefs’ help article on how to view fan-out queries explains that Brand Radar’s AI Responses report shows a fan-out column for ChatGPT and Perplexity. This is useful when you want hundreds of prompts, not five. If you are choosing a tool, my post on AI visibility tools compares the main options.
| Method | Cost | Scale | Best for |
|---|---|---|---|
| Developer tools | Free | One prompt at a time | Learning the pattern, spot checks |
| Chrome extension | Usually free | Tens of prompts | Quick research sessions |
| AI visibility platform | Paid | Hundreds to millions | Ongoing tracking, trend data |
How ChatGPT fan-out differs from Google AI Mode
Google uses the same idea. Its documentation on AI features and your website says:
“Both AI Overviews and AI Mode may use a “query fan-out” technique.”
— Google Search Central, AI features and your website
But the two systems behave differently in ways that matter for your content plan:
- Index. Google fans out into its own index, so normal Google rankings carry straight over. ChatGPT relies on its own crawlers and search partners, so you need to be reachable by
OAI-SearchBotas well. - When it searches. In the Ahrefs sample, ChatGPT searched on only 14.4% of responses. AI Mode is a search product, so it is built to retrieve.
- Brand checks. ChatGPT’s use of
site:searches to verify facts on brand websites is very visible in the data. That makes your own site a direct input. - Reporting. Google reports AI Overviews and AI Mode traffic inside Search Console’s Performance report under the “Web” search type. ChatGPT gives you no query report at all. You only see referral visits, which GA4 now groups in an AI Assistant default channel.
For the Google side, including how to map sub-questions to page sections, read my Google query fan-out guide.
Turning fan-out queries into a content gap list
This is where the research pays off. Here is how I would do it for a business, step by step.
- Write 20–40 real buyer prompts. Use the words your sales team hears. Mix “best X for Y,” “X vs Y,” “how much does X cost,” and “is X good for Y.” My guide to prompt research shows how to collect them.
- Run each prompt and record the fan-out queries. Use developer tools or a tool. Run important prompts more than once, because answers vary.
- Put every query in a spreadsheet with columns for prompt, query, added words (year, brand, “price,” “reviews,” location) and whether it used
site:. - Check who ranks for each query. Search the query yourself. Note whether your page appears, a competitor’s page appears, or nobody covers it well.
- Tag each query as “covered,” “weak” (you have a page, but it’s thin or outdated) or “missing.”
- Group the gaps by page type: pricing, comparison, review roundup, product facts, local page.
- Prioritize by money. Gaps tied to buying prompts go first.
Example: a hypothetical project-management SaaS
Let me show what this looks like with an example. These numbers are made up for illustration.
Say a small project-management software company runs 30 buyer prompts. ChatGPT searches on 18 of them and produces 41 fan-out queries. After tagging:
- 12 queries include “2026.” The company’s comparison page still says “2024” in the title.
- 9 queries are
site:searches on the company’s own domain, mostly for “pricing” and “integrations.” The pricing page shows a “Contact sales” button and no numbers. - 7 queries include “reviews” or “vs.” Two competitor comparison pages exist on review sites, but the company has none of its own.
- 13 queries are already covered by decent pages.
From 41 queries, the action list is short: update the comparison page and its year, publish clear pricing or at least starting prices, write two honest “vs” pages, and get listed on one or two review sites. That is four or five pieces of work, not forty.
Matching titles and headings to the rewritten queries
Once you know the fan-out queries, you can shape pages to match them. This is not keyword stuffing. It is making sure the page clearly answers the search ChatGPT actually ran.
| Fan-out pattern | What to put on the page |
|---|---|
| “best [category] for [use case] 2026” | A dated title and a visible “last updated” date. A short verdict near the top for each use case. |
“[brand] pricing” / site: brand + “price” | Real numbers or ranges in HTML text, not just in an image or PDF. Plan names that match how people say them. |
| “[brand] vs [brand]” | A fair comparison table with the facts a buyer checks: price, key features, limits, who it suits. |
| “[product] reviews” | Review snippets on your site, plus profiles on the review platforms your buyers use. |
| “[service] [city]” | Local pages with real local detail, address, service area and proof. |
A few rules I follow:
- Put the answer in the first lines of each section. If a heading says “How much does X cost?”, the next sentence should give the number.
- Use the buyer’s words in headings, the same words you saw in fan-out queries.
- Keep facts in plain text. Crawlers read text far more reliably than images or scripts.
- Update the year only when you update the content. Changing “2025” to “2026” without new information is the kind of shortcut that hurts trust.
Google’s guidance supports writing for people first. Its guide to AI features warns against recycling “what others on the internet have already said.” The same logic works for ChatGPT. Ahrefs found that pages that only summarize other people’s work were the ones losing citations in its data.
What fan-out data can’t tell you
I want to be honest about the limits here, because fan-out queries are easy to over-read.
They change with the model. Ahrefs put it simply:
“OpenAI changes how ChatGPT retrieves information constantly.”
— Ahrefs, How ChatGPT Does Query Fan-Out
The Ahrefs data covers one model over two months. Patterns you capture today may not hold next quarter, so I re-check key prompts every month or two.
They change with the user. Logged-in users have memory and chat history, and those shape what ChatGPT searches for. Location changes the query too, as the jump in country names shows. A logged-out test from your office is one view, not the view. My post on ChatGPT memory and personalization goes deeper on this.
They change between runs. Answers to the same prompt vary. A SparkToro and Gumshoe study reported by Search Engine Land in January 2026 found the odds of getting the same list of brands twice were under 1 in 100. Fan-out queries vary too. Run important prompts several times before you draw conclusions.
They don’t show the whole decision. The query is only the first step. Which pages ChatGPT reads, and which it cites, depends on other things, like the authority of the site and whether OpenAI has a content deal with a publisher. Ahrefs found partner publishers’ share of citations moved between 7.5% and 15.8% across 2026.
Questions marketers ask about ChatGPT fan-out queries
Does ChatGPT search the web for every question?
No. In Ahrefs’ study of 66.6 million logged-out responses, only 14.4% triggered a web search. Simple, well-known topics are usually answered from the model’s own knowledge.
How many searches does ChatGPT run per prompt?
It depends on the prompt and the model. Ahrefs measured about 1.6 per searching answer across a huge free-tier sample, while smaller studies on other setups reported up to 7.61. Comparison and buying prompts fan out the most.
Can I see the fan-out queries without paying for a tool?
Yes. Open Chrome developer tools, go to the Network tab, run a prompt and search the responses for search_model_queries. The field name may change when OpenAI updates the app.
Should I add the current year to my titles?
Only when the page is genuinely current. The share of fan-out queries containing “2026” rose from 17.6% to 22.9% in Ahrefs’ data, so freshness matters. But update the content first, then the year.
Why does ChatGPT search my own website with site:?
It is usually checking facts like price, specs or policies at the official source. Make sure those facts are on your site, in plain text, and up to date.
Where to start with ChatGPT fan-out this month
If you only do three things, do these. First, confirm OAI-SearchBot can crawl your site. Second, run 20 buyer prompts and write down every fan-out query. Third, fix the pages that ChatGPT goes looking for on your own domain, especially pricing, comparisons and product facts.
Fan-out data is one input into a bigger plan. To see how it fits with tracking and reporting, read my framework for deciding what AI visibility data to act on, and for the wider picture, my roundup of AI search trends shaping visibility right now. If you want to prove the work is paying off, my guide to holdout and pre/post tests for AI search shows how. And if your team wants to automate the query collection, see Claude Code for SEO.
If you would like us to run this analysis for your brand and turn it into a content plan, our SEO team and content writers can help. Get in touch for a quote and tell us which prompts matter most to your business.
Further reading on ChatGPT query fan-out
- Too Much AI Visibility Data? A Framework for Deciding What to Act On
- AI Search Trends Shaping Visibility Right Now
- Proving AI Search Impact With Holdout and Pre/Post Tests
Sources for this ChatGPT fan-out guide
- Ahrefs – How ChatGPT Does Query Fan-Out (15 Million Queries Studied)
- Ahrefs Help Center – How to view fan-out queries generated by AI
- NoGood – Query Fan-Out: The Complete Guide
- OpenAI – Overview of OpenAI Crawlers
- Google Search Central – AI features and your website
- Google Search Central – Guidance on generative AI features in Search
- Search Engine Land – AI recommendation lists repeat less than 1% of the time
