Key takeaways
- Query fan-out is the step where an AI system rewrites one prompt into several searches, runs them, and merges the results. Those sub-queries decide which pages get retrieved and cited.
- Tools split into simulators (they predict likely sub-queries), real-UI measurement platforms (they record what the AI actually searched), and SEO suites that approximate fan-outs as a side feature.
- Fan-outs are smaller than most blog posts claim. Promptwatch's measured ChatGPT average was about 1.84 per response in early March 2026, and it was exactly 1.0 when the data resumed in April.
- Fan-out behavior changes overnight. On August 8, 2026, ChatGPT began using the site: operator at scale, so any fan-out analysis from before that date is stale.
- Use a free simulator for brainstorming, a measurement platform for decisions, and your own crawler logs and Search Console data to confirm the pages actually get fetched.
What query fan-out is, in plain terms
When you ask ChatGPT, Gemini, Perplexity or Google AI Mode a complex question, the model usually does not run your words as a single search. It breaks the prompt into sub-queries, fires them off in parallel, scores what comes back, and writes one answer from the winners. Similarweb's explainer describes it as a hidden reasoning process where the AI rewrites a question into multiple sub-queries and merges the results.
The practical consequence: you are no longer competing for one query. A page can get cited without ranking first for the head term, because it answers one of the sub-queries well. That is why fan-out analysis has turned into its own tool category.
How many sub-queries does AI actually run?
Here I want to push back a little on the standard advice. Plenty of articles say a prompt triggers "8 to 20 or more" background searches. Semrush, to its credit, notes that Google has not published exact counts and calls its figure an estimate. Real measurements from ChatGPT look a lot smaller.
Promptwatch Data's ChatGPT query fanout report, built on real UI monitoring, shows:
- Average fan-out queries per response fell from 2.15 in early December to about 1.84 by early March 2026.
- After a gap in the data, the average came back in April at exactly 1.0 per response.
- Average fan-out query length dropped from about 117 characters in December to the high 80s in February and March, and about 53 characters in April. That is roughly six to eight words, closer to a keyword search than a conversational question.
So the honest answer is: it depends on the platform, the prompt, and the month. A simulator that gives you 15 neat sub-queries for every prompt is showing you a plausible decomposition, not a recording of what happened. Both are useful. They are not the same thing.
The August 2026 shift that broke old analyses
On August 8, 2026, ChatGPT started using the site: operator at scale. The share of fan-outs using it jumped from 0.37% to 16.8% in a day, and average fan-outs per response rose from about 1.08 to about 1.83. Promptwatch's reading is that the site: searches were added on top of the generic ones, and that this was a platform-side rollout.
Two lessons. First, your own domain is now a direct retrieval surface, so crawlability and indexation gaps cost you answers. Second, a fan-out snapshot taken in July is out of date. Any tool that gives you a static list, or a simulation that does not track platform changes, can mislead you quietly.
The three types of fan-out tools
Before comparing products, sort them by what they actually do.
| Type | What you get | Strength | Weakness |
|---|---|---|---|
| Simulators and extractors | Predicted sub-queries for a prompt you type in | Free or cheap, fast brainstorming | Guesses, not observed behavior |
| Real-UI measurement platforms | Sub-queries recorded from actual AI responses, tied to citations | Reflects what the AI really did | Costs money, depends on prompt set |
| SEO suites with AI add-ons | Approximated sub-queries inside a broader toolkit | One login, familiar workflow | Shallower AI-specific data |
| Native platform views | Sub-queries a platform shows itself (Perplexity displays steps) | Free, real | One prompt at a time, no history |
Free ways to look at fan-outs
Perplexity's visible steps
Perplexity shows the searches it runs while it builds an answer, which makes it the easiest place to watch fan-out happen for a single prompt. It does not scale and does not keep a history, but for building intuition it is hard to beat.
Perplexity
Queryfanout.io and similar simulators
Semrush's guide points to Queryfanout.io, a free tool that replicates Google's fan-out process to suggest sub-queries for a target prompt. Treat the output as a hypothesis list. It is good for spotting angles you missed (pricing, alternatives, comparisons), not for proving what any engine ran.
ChatGPT itself
You can ask ChatGPT to reason about how it would break a topic down, but the answer is a model describing its own behavior, not a log. Use it for ideas only.
The GEO Inspector extension
Promptwatch publishes a free Chrome extension, GEO Inspector for ChatGPT, that surfaces fan-outs and the difference between fetched and cited sources inside the ChatGPT interface. That makes it a rare free option that reads real behavior rather than simulating it.
Platforms that measure real fan-outs
This is where the money decisions happen. A good measurement platform records the sub-queries from real responses, over time, and connects them to citations and traffic.
Promptwatch
Promptwatch is the platform I would start with if the goal is acting on fan-outs rather than just looking at them. It monitors the actual interfaces of ChatGPT, Google AI Overviews, AI Mode, Gemini, Perplexity and others, and its prompt intelligence shows query fan-outs, the exact web searches AI runs per response, next to monthly prompt volumes and difficulty scores. Because those fan-outs sit beside Citation Trends, IP-verified AI crawler logs and Google and Bing Search Console data, you can check whether a sub-query you care about leads to a crawl, a citation and a click, and the Content Agent can then draft pages for the gaps. Most monitoring-only trackers stop before that last step.
It also practices what the data suggests. Its fan-out research recommends headings written like search queries ("Best CRM for small agencies 2026" rather than a chatty question), entity and category first in titles and H2s, and separate focused pages for comparisons, pricing, alternatives and how-tos instead of one mega-page. Pricing starts at $95/month on Essential with a free Explore tier; confirm current plans at promptwatch.com/pricing.
Profound
Profound is an enterprise visibility platform with strong monitoring. It is a reasonable pick for large brands that want deep tracking and have the budget, though it has no Reddit tracking or agentic content execution that Promptwatch includes.
Scrunch
Scrunch monitors how assistants like ChatGPT and Claude describe and cite brands. Worth a demo if your concern is AI-facing site content and monitoring.
Peec AI and Otterly.AI
Both are lighter prompt trackers. They answer "was my brand mentioned?" reasonably well and cost less. Neither gives you crawler logs or content generation, so fan-out insight stays at the surface.
Radarkit, Finseo and the listicle favorites
A few of the 2026 roundups name Radarkit as the strongest overall pick, and one names Finseo for fan-out analysis plus bot traffic and white-label reporting. I would read those as vendor-adjacent rankings and verify claims yourself on a trial. Radarkit is in our catalog if you want to try it.
SEO suites with fan-out features
Semrush
Semrush's AI Visibility Toolkit approximates the sub-queries AI systems may generate for your prompts, and its enterprise AI optimization product claims to show Google's actual fan-out queries at scale. If you already live in Semrush, that is convenient. Verify what is observed and what is modeled before you build a strategy on it.
Ahrefs Brand Radar
Brand Radar tracks brand mentions in AI search. It is useful for share-of-voice context, less so as a dedicated fan-out analyzer.
Similarweb
Similarweb's AI search content is educational, and its guidance suggests combining fan-out extractors with content analysis tools to find gaps. Handy for market-level context.
Comparison table
| Tool | Fan-out source | Ties to citations | Crawler logs | Content action |
|---|---|---|---|---|
| Promptwatch | Real UI responses | Yes | Yes (IP-verified) | Content Agent, Action Board |
| Profound | Monitoring data | Yes | Limited | No |
| Peec AI | Prompt tracking | Basic | No | No |
| Otterly.AI | Prompt tracking | Basic | No | No |
| Semrush | Approximated, plus enterprise tier | Partial | No | Via suite |
| Queryfanout.io | Simulated | No | No | No |
| Perplexity (native) | Shown live | Per answer | No | No |
A practical workflow
Step 1: brainstorm with a simulator
Take your five most valuable prompts and run them through a free simulator. You are looking for sub-topics: comparisons, pricing, alternatives, how-tos, integrations.
Step 2: check against reality
Run the same prompts through a measurement platform or the GEO Inspector extension. Note which sub-queries really appear and how short they are. Expect brevity. If real fan-outs are six to eight words, your headings should be too.
Step 3: map sub-queries to pages
Build a simple sheet: sub-query, page that should answer it, current heading, status. Gaps are your content roadmap. Promptwatch Data's advice to build focused pages per sub-question applies here.
Step 4: confirm retrieval
Check crawler logs and Search Console. If a bot never fetches the page, content quality is not the problem. The tools covered in the GEO software directory at bestgeosoftware.com can help you compare options for this.
Step 5: re-run after platform changes
The August 8 site: change shows why. Re-audit your keyword-to-heading mapping after any major ChatGPT update, and prefer tools that monitor continuously over one-off snapshots.
Mistakes I see teams make
- Treating a simulation as measurement. A list of 15 neat sub-queries feels authoritative. It is a guess.
- Optimizing for the 8 to 20 figure. Measured ChatGPT averages in Promptwatch's data are closer to 1 to 2 per response, though other engines may differ and I have not seen first-party data on AI Mode or Gemini.
- Writing conversational headings. Real fan-outs read like search queries.
- Ignoring the site: effect. Since August, thin or poorly indexed pages on your own domain hurt more.
- Skipping the link between fan-out and traffic. A sub-query you cover but never get clicked from is a finding, not a win.
Which tool should you pick?
- Solo marketer or just curious: Perplexity's visible steps, GEO Inspector and a free simulator cost nothing.
- Small team that needs to act: Promptwatch Essential or Professional gives real fan-outs plus content workflows.
- Agency: Promptwatch's agency plans (Kick-off from $199/month) include unlimited projects and white-label reporting.
- Already on Semrush: try the AI Visibility Toolkit first, then decide if you need deeper data.
- Enterprise with big budget: compare Promptwatch against Profound and Scrunch on a real prompt set.
For wider tool research, the AI rank tracking directory at ai-rank-tools.com lists many more options, and Promptwatch's own research hub is worth bookmarking because the numbers move month to month.
FAQ
Can I see exactly what sub-queries ChatGPT ran?
Not natively in a clean log. Browser extensions and platforms that monitor the real UI can capture them. Perplexity shows its steps directly.
Do all AI engines fan out the same way?
No. Claude, ChatGPT with browsing, Perplexity and Gemini all decompose complex prompts, but depth and style differ, and an SSRN study of 1,323 fan-out queries across ChatGPT, Gemini and Perplexity found behavior varies by platform and intent type.
How often should I re-run fan-out analysis?
Monthly at minimum, and immediately after notable platform updates.