# ChatGPT Fan-Out Is Smaller Than Claimed: What the Data Shows

> The 8 to 12 sub-query figure doesn’t hold up: in free ChatGPT, Ahrefs found about 1.6 in responses that had any. What that means for your content.

URL: https://llmagnet.com/blog/chatgpt-query-fan-out  
Author: Ben Cohen, Founder & CEO at LLMagnet  
Published: 2026-09-01  
Updated: 2026-11-10  
Category: Research

ChatGPT usually runs far fewer sub-queries than the 8 to 12 often quoted: Ahrefs’ study of 66.6 million free, logged-out responses found fan-outs in 14.4% of them, averaging about 1.6 in the responses that had any. So don’t build a page for a dozen imagined sub-queries. Write focused pages that answer the specific questions buyers ask, and earn mentions on the sites engines already cite.

## Contents

1. What the Fan-Out Data Shows
2. What ChatGPT Adds to Its Searches
3. What This Means for Your Content
4. Why Engines Cite Different Domains
5. What Correlates With Being Cited
6. Sources Differ by Engine and Move Quickly
7. Where Structured Data and MCP Fit
8. What to Do This Month

## What the Fan-Out Data Shows

Fan-out is when an AI system turns a prompt into one or more search queries of its own. Ahrefs looked at 15 million fan-outs from August and September 2026, on GPT-5.6 Luna. These were broad prompts drawn from real search demand, in free sessions, which search less than paid or reasoning models. Ahrefs is a vendor, and calls the result a snapshot of one model. It also corrected its own wording: 14.4% is the share of responses containing fan-out queries, not the share that searched the web.

![Four steps to a citation: the engine decides to search, it writes its own queries, it retrieves results, and it picks what to cite.](https://llmagnet.com/images/blog/charts/fanout-flow.png)

*Our simplification. OpenAI doesn’t publish its retrieval pipeline.*

Peec, another vendor, gets higher numbers. [Its analysis of over 20 million fan-outs](https://docs.peec.ai/research/chatgpt-fanout-length-trend) found about 2.3 to 2.8 per prompt from October 2025 to January 2026, with the queries growing more specific. [After GPT-5.6 launched](https://docs.peec.ai/research/gpt-56-retrieval-changes.md), it reports the share of prompts needing a single query fell from 94.0% to 43.5%, and that 7.2% of prompts now take six or more iterations. Peec doesn’t give its sample for that second analysis, and it counts retrieval iterations, so it isn’t directly comparable.

![Share of prompts by retrieval iterations in Peec’s data: one iteration fell from 94.0 percent on GPT-5.5 to 43.5 percent on GPT-5.6, and two iterations rose from about 5 percent to 33.5 percent.](https://llmagnet.com/images/blog/charts/fanout-iterations.png)

*Source: Peec, August 2026. Vendor data, sample size not disclosed.*

Peec also reports that sources retrieved per prompt doubled, from about 12 to about 24. So even where the query count is low, each search can bring back more pages to choose from.

Other measurements go higher still, and the setup explains most of the gap:

| Source | Sample | Setup | Fan-outs | Share using `site:` |
| --- | --- | --- | --- | --- |
| **Ahrefs** | 66.6 million responses | GPT-5.6 Luna, free, logged out, broad prompts | In 14.4% of responses, about 1.6 in those | 6.8% |
| **Peec** | Not disclosed | GPT-5.6, partly transactional prompts | 56.5% of prompts had two or more | About 23% |
| **[Nectiv](https://www.searchenginejournal.com/what-we-can-learn-from-evolving-chatgpt-fan-out-queries/586254/)** | About 4,000 prompts | GPT-5.6 Sol, through the API | 7.61 per prompt, up from 2.17 in 2025 | 64% |
| **Suganthan Mohanadasan** | 57 conversations | Logged in, Thinking mode, SaaS-heavy | 15 to 40 | Frequent on commercial prompts, no figure |

Ahrefs attributes the differences mainly to prompt type and session type: many other studies used commercial “best X” prompts, and free sessions search less than paid or reasoning models. The Mohanadasan row comes from Ahrefs’ comparison table. Nectiv’s 7.61 is the closest published figure to the popular claim, but Lily Ray, writing about that study, notes that free users get the cheaper Luna variant and not Sol. We couldn’t find a ChatGPT measurement behind a fixed range of 8 to 12. Count them on your own prompts, and expect the number to change with each model.

## What ChatGPT Adds to Its Searches

The words inside fan-outs are more useful than the count, because they tell you what the engine thinks the answer needs.

![Share of fan-out queries containing each word in September 2026: 2026 22.9 percent, official 13.1, price 8.9, best 5.3, reviews 3.6, cost 2.0.](https://llmagnet.com/images/blog/charts/fanout-words.png)

*Source: Ahrefs, updated October 2026. Vendor data. Ahrefs confirms an English-only base for “official” alone.*

Ahrefs saw “price” rise from 6.0% to 8.9% of fan-outs between August and September, and “official” from 10.1% to 13.1%. [Peec](https://docs.peec.ai/research/chatgpt-fanout-length-trend) found the average fan-out roughly doubled in length, from about 6 words in October 2025 to about 12 in January 2026, while the number per prompt stayed flat.

Brand sites benefit from this. In [Ahrefs’ data](https://ahrefs.com/blog/query-fan-out-chatgpt/), 75.9% of `site:` searches to the most-searched sites went to brand sites. In a separate test of 20,000 US prompts, 84.4% of factual questions cited at least one official source, and so did 68.3% of comparison questions. These numbers suggest ChatGPT turns to official pages to confirm facts such as prices and specifications, while comparisons draw on a wider mix.

Ahrefs’ own example shows the shape. Of 14,536 “running shoes” prompts in US ChatGPT, 1,474 responses contained fan-outs. The most common words were “best” (206), “2026” (160), “review” (117), “vs” (105) and “official” (65). That’s a list of pages a shoe brand would want: a current model page, a comparison, and something a reviewer can link to.

## What This Means for Your Content

Fewer, more specific queries change the advice. You don’t need coverage of a long list of guessed sub-queries. You need to be a strong result for the specific query the engine writes.

- **Be specific.** A page titled for one buyer’s question, with the audience, use case or year stated, is a better match for a specific query than a general overview.
- **Answer the follow-up on the same page.** If ChatGPT only searches once, one page may have to supply both the recommendation and the reason for it.
- **Cover pricing, integrations and comparisons where buyers look for them.** Separate pages are sensible if each has enough to say. We found no study showing that one structure wins.

Here’s how to apply it. The company and queries below are made up for illustration. Quillon sells payroll software and logs what ChatGPT searches for on 25 buyer prompts.

| Logged query | Closest page | Gap | Fix |
| --- | --- | --- | --- |
| contractor payroll software 2026 reviews | Homepage | No reviews or year on the page | Build a customer results page, dated and updated |
| quillon pricing official | Pricing page | Prices load by script | Render prices in the first HTML response |
| quillon vs northfield payroll | None | No comparison page | Publish an honest comparison |
| site:quillon.example contractor countries | Features page | List of countries is an image | Replace it with text |

Check the result by asking whether the page now answers the logged query in its first screen of text, then rerun the prompts after the page has been recrawled.

**When the advice changes:** Ecommerce and software teams should put the most care into price, specification and comparison pages, since those are the words engines add. If your buyers rely on paid or reasoning modes, expect more searches per prompt and more `site:` checks on your own domain. For local services or simple informational queries, the fan-out may never fire, so your off-site presence and clear facts matter more than page count.

## Why Engines Cite Different Domains

[Writesonic’s study of 161,286 prompts](https://writesonic.com/blog/ai-citation-source-overlap-study) (May and June 2026, vendor data) found ChatGPT and Perplexity shared about 13% of cited domains on a Jaccard measure. Exact URLs would overlap less. So being cited on one engine tells you little about another.

The engines don’t publish their retrieval pipelines, and outside studies can’t fully separate the reasons. Differences in index, query writing and source selection all play a part, and all of them change with new models.

## What Correlates With Being Cited

[Ahrefs’ May 2025 study of 75,000 brands](https://ahrefs.com/blog/ai-overview-brand-correlation/) found branded web mentions correlated with AI Overview brand visibility at 0.664, against 0.218 for backlinks. It covers Google AI Overviews, not ChatGPT, and only higher-authority domains. It’s a correlation, so it can’t prove that mentions cause citations.

A sensible reading is that the more independent pages discuss your brand in a relevant context, the more chances any query has to find one. That’s a hypothesis, but it costs little to act on.

## Sources Differ by Engine and Move Quickly

Don’t plan around a fixed list of favorite sources for each engine. In [Ahrefs’ November 2026 AI Overviews ranking](https://ahrefs.com/blog/most-cited-domains-ai-overviews/), YouTube held a 24.5% mention share and Reddit 18.1%, while the number of answers citing them fell 11% and 21% from the month before. [Semrush reported](https://www.semrush.com/blog/reddits-citations-in-chatgpt-fall/) Promptwatch data showing Reddit’s share of ChatGPT citations falling from 3.8% to 0.5% in August 2026, a drop Promptwatch calls provisional. [Ahrefs’ US ChatGPT data](https://ahrefs.com/blog/query-fan-out-chatgpt/) shows the longer trend: Reddit went from about 1 in 5 answers in March 2026 to about 1 in 43 in September, and Wikipedia from 1 in 7 to 1 in 35.

Check who is cited for your prompts each month. A snapshot from last quarter may already be wrong.

## Where Structured Data and MCP Fit

**Schema:** Evidence is mixed and depends on type and platform. In [Ahrefs’ controlled test](https://ahrefs.com/blog/schema-ai-citations/) of 1,885 pages that added JSON-LD, citations didn’t change significantly on AI Mode or ChatGPT and fell 4.6% on AI Overviews relative to matched controls, a decline Ahrefs couldn’t explain. Those pages were already heavily cited, so it doesn’t settle the question for pages that aren’t. Add schema where it describes the page accurately, but don’t expect it to earn citations alone.

**MCP:** [Introduced by Anthropic in November 2024](https://stacklok.com/wp-content/uploads/2026/01/State-of-MCP-in-Software-2026_FINAL.pdf), MCP lets an AI application call tools on your system. It doesn’t replace retrieval. An ordinary ChatGPT answer won’t query your server unless that connection exists, so treat MCP as a way to serve customers and partners who connect to it, not as a citation tactic.

## What to Do This Month

1. **Collect real queries.** Choose 25 prompts buyers would ask, and use a tool that logs ChatGPT’s fan-outs. Note the words it adds, such as years, prices or reviews.
2. **Match your pages to them.** Where a real query has no good page on your site, write one, or add the missing answer to the closest page.
3. **List the cited domains.** Record which sites each engine cites for those prompts. Pick the ten your buyers actually read.
4. **Earn mentions there.** Contribute articles, complete listings and answer questions in the open, with any affiliation stated.
5. **Don’t count on llms.txt for citations.** [SE Ranking’s study of about 300,000 domains](https://seranking.com/blog/llms-txt/) found no measurable link between having the file and being cited. If you publish one, treat it as a map for agents.

LLMagnet measures AI visibility across seven engines, and its agent prepares fixes for you to approve. To see where you stand, run the free audit at [ai-visibility.llmagnet.com](https://ai-visibility.llmagnet.com).

## FAQ

### How many searches does ChatGPT run per question?

It varies by model and prompt. Ahrefs found fan-outs in 14.4% of 66.6 million free, logged-out responses, averaging about 1.6 in those responses. Nectiv measured 7.61 per prompt through the API on a different model. We found no ChatGPT measurement behind a fixed range of 8 to 12.

### What is query fan-out?

It’s when an AI system turns a user’s prompt into one or more search queries of its own, then uses the results to write an answer. In ChatGPT, these queries can add specifics such as a year, price or reviews.

### Do I need a separate page for every possible sub-query?

No. Fan-out counts are lower than often claimed, so focus on the specific questions buyers ask. Separate pages for pricing, integrations and comparisons make sense where each has enough to say, but no study shows one structure wins.

### How much do ChatGPT and Perplexity overlap?

Writesonic’s study of 161,286 prompts from May and June 2026, which is vendor data, found the two shared about 13% of cited domains on a Jaccard measure. Exact-URL overlap would be lower.

### Do brand mentions beat backlinks for AI visibility?

They correlate more strongly in one study. Ahrefs found branded web mentions correlated with AI Overview visibility at 0.664, against 0.218 for backlinks, across 75,000 brands. It covers AI Overviews and doesn’t show cause.

### Will an MCP server get me cited in ChatGPT?

Not by itself. MCP lets an AI application call tools on your system once a connection exists. An ordinary ChatGPT answer won’t use your server without one, so treat MCP as a service for connected customers and partners.

### What words does ChatGPT add to its searches?

In Ahrefs’ September 2026 data for free ChatGPT, 22.9% of fan-outs contained “2026”, 13.1% “official” and 8.9% “price”. That points to current, official pages with prices and specifications in plain text.