Key takeaways
- A query fan-out is the set of hidden searches an AI engine runs after someone types one prompt. Your content competes across all of them, not just the visible question.
- ChatGPT's behavior changed sharply in August 2026. On August 8, the share of fan-outs using the site: operator jumped from 0.37% to 16.8%, according to Promptwatch Data.
- Published studies disagree on fan-out counts (roughly 1 to 7.6 per prompt) because they use different collection methods, models, and dates. Treat every number as a snapshot.
- Engines differ. Perplexity mostly cleans up your wording, Grok runs a research brief, and ChatGPT invents new vocabulary.
- The practical playbook: cover sub-questions (comparisons, pricing, alternatives), make your own domain easy to retrieve, and monitor continuously instead of auditing once.
What a query fan-out actually is
When you ask ChatGPT "what are the best project management tools?", it does not simply search that sentence. It breaks the prompt into several background searches, runs them in parallel, pulls results, and blends them into one answer. Those background searches are query fan-outs.
A useful analogy comes from Peec AI's analysis of 5 million fan-outs (April 1-21, 2026): ask a friend "what's a good restaurant?" and they silently think about budget, occasion, cuisine, and distance before answering. Fan-out is the machine version of that.

Two terms get mixed up constantly, so pin them down early. Lily Ray draws the distinction in her August 2026 analysis: a page is retrieved when ChatGPT fetches it while running fan-outs, and cited when it appears as a link in the visible answer. Retrieval is growing while the number of unique cited domains per response is falling. When two fan-out studies disagree, check which one is measuring which.
Not every ChatGPT answer triggers any of this. Questions answerable from training data can skip web search entirely, and cheaper models lean on training data more often. Fan-outs only appear when the model decides it needs fresh information.
Why fan-outs decide who gets cited
The mechanism matters because of how results get combined. Metehan Yesilyurt reported that ChatGPT appears to use Reciprocal Rank Fusion, which merges rankings across subqueries. The consequence, if that holds, is simple: a page that shows up for several sub-searches scores better than a page that shows up for one. Narrow content that nails a single angle has a structural disadvantage.
There is also a brutal funnel after retrieval. In the Nectiv-based data Lily Ray cites, brands named in the fan-out query itself were cited 68.9% of the time, while pages that were merely fetched were cited 2.1% of the time. Only about 3.1% of 3,554 retrieved pages made it into an answer. Being retrieved is table stakes. Being named in the search string is the real prize.
For example, when someone asked for the best AI note-taking app, the very first fan-out already contained a list of specific brands. If you are not in the model's shortlist before it searches, you are fighting for scraps afterward.
What the data says ChatGPT searches for
The words it injects
Peec's analysis found the words ChatGPT adds most often to fan-outs, even when the user never typed them: "best" (15.33%), "what" (8.72%), "review(s)" (6.84%), "2026" (5.44%), "top" (5.24%), "comparison" (4.48%), "vs" (4.27%). For advice-style prompts like "Should I use X?", "best" showed up in 24.3% of fan-outs. This is a plausible reason listicles keep winning. Separately, Promptwatch Data on ChatGPT citation share by content type shows product pages at 22.4% of citations (Sep 28 to Oct 4, 2026), followed by landing pages and listicles.
Reviews matter too. ChatGPT searches for them unprompted, so sites like G2, Glassdoor, and Sitejabber quietly shape how AI describes you.
The site: operator shift
This is the biggest recent change, and it post-dates most published studies. According to Promptwatch Data, on August 8, 2026 the share of fan-out queries containing site: went from 0.37% to 16.8% overnight, roughly 46x, or about 1 in 6 fan-outs. Average fan-outs per response rose from about 1.08 to about 1.83 the same day and stayed there. Because the average rose rather than stayed flat, site: searches appear to have been added on top of generic searches, not swapped in for them.
Other teams measured the same direction with different numbers. David Konitzny at Peec AI saw site: go from about 0.3% to about 23% of fan-outs when GPT-5.6 became the default, with single-fan-out prompts dropping from 94.0% to 43.5%. Chris Long at Nectiv, pulling from the API on 5.6 Sol, found site: in 64% of queries and average fan-outs per prompt rising from 2.17 to 7.61. The gap between these figures likely comes from collection: API versus consumer UI, plus different models and tiers.

Lily Ray's theory is that site: scoping is a spam filter: the model restricts searches to domains it already trusts, and words like "official" and "gov" climbing the unigram list fits that reading. It is a theory, not a confirmed OpenAI statement, but it matches the data.
Why the numbers don't agree
Peec reports 2.1 fan-outs per ChatGPT prompt for April 2026. Promptwatch's own time series shows about 2.15 in early December 2025, easing to about 1.84 by early March 2026, then 1.0 per response when fan-outs returned in April. Different datasets, different methods. And everything before August 8 is now partly stale. My advice: don't quote one average. Quote a range with a date.
| Source | Period | Collection | Headline finding |
|---|---|---|---|
| Promptwatch Data | Dec 2025 to Aug 2026 | Real-UI tracking | ~2.15 to ~1.0 per response, then ~1.83 after Aug 8 |
| Peec AI | Apr 1-21, 2026 | 5M fan-outs, 3 engines | 2.1 per ChatGPT prompt, "best" most injected |
| Nectiv (Chris Long) | 2026 | OpenAI API, ~4,000 prompts | 2.17 to 7.61 per prompt, site: in 64% |
| Peec (Konitzny) | GPT-5.6 default day | UI extraction | Single fan-out prompts 94.0% to 43.5% |
| Profound | 14 days, 10,000 prompts | 3 engines | 91% of ChatGPT queries unique |
How engines differ
Don't assume one strategy covers everything.
- Perplexity stays closest to what you typed. Peec measured about 1.4 fan-outs per prompt; Profound measured 88% word overlap with the original prompt and only 14% unique queries across daily runs. Classic keyword coverage still works here.
- Copilot compresses prompts into keyword-style queries, with about 50% word overlap and 47% uniqueness in Profound's data.
- Grok behaves like a research brief. Peec measured 6.8 fan-outs per prompt, narrowing by year, brand-vs-brand comparisons, and targeted source searches such as site:reddit.com or site:g2.com.
- ChatGPT is the wild one. Profound found 91% of its queries are unique run to run and typical worst-case word overlap with the prompt is just 13%. A prompt about eco-friendly protein powder for endurance athletes can fan out into plant-based supplements and sports nutrition research that share almost no words with the original.
The Profound finding has a monitoring implication that I think is under-appreciated: because ChatGPT rarely repeats itself, a single check captures one of many possible retrieval paths. One-off audits mislead. You need ongoing tracking.
How to see the fan-outs for your category
You have a few options, from free to automated.
- Read them in your browser. In ChatGPT, choose a search-enabled answer and inspect the conversation's network response for fields like
search_model_queriesorsearch_queries, as walked through in Aiso's guide. Caveat: if the field is missing, that does not prove ChatGPT didn't search. Never treat ChatGPT's own reconstruction of its searches as observed data. - Use a Chrome extension. The free Resoneo plugin and FanoutFox surface queries and sources as you chat. Promptwatch also publishes a free GEO Inspector extension showing fan-outs, fetched versus cited sources, and the reasoning trail locally.
- Generate candidates. The free ChatGPT query fan-out generator gives you plausible sub-queries for brainstorming. Label these as generated, not observed.
- Track them at scale. Peec AI and Profound both expose fan-outs for tracked prompts, and Similarweb covers AI brand visibility more broadly.
Promptwatch shows query fan-outs per tracked prompt, compared per model, collected from the real interfaces of ChatGPT, Google AI Overviews, Perplexity, and others, then ties them to citations and AI crawler activity. That connection matters, because a fan-out list on its own is just trivia until you know which of your pages answer each branch.

| Approach | Cost | Observed or generated | Scale |
|---|---|---|---|
| Browser network inspection | Free | Observed | One conversation at a time |
| Chrome extensions (GEO Inspector, Resoneo, FanoutFox) | Free | Observed | One conversation at a time |
| Fan-out generator | Free | Generated | Brainstorming only |
| Platforms (Promptwatch, Peec, Profound) | Paid | Observed | Tracked prompt sets over time |
What to do with the fan-outs
1. Map your content to sub-questions
Pull the fan-outs for your ten most valuable prompts and sort them: comparisons ("X vs Y"), pricing, alternatives, reviews, how-tos, location modifiers. Then check which you have a page for. Promptwatch's own recommendations from its fan-out data are to build clusters of focused pages instead of one mega-page, and write headings that read like short keyword queries, for example "Best CRM for small agencies 2026". Its report also notes average fan-out query length fell from about 117 characters in December to about 53 in April, so short, entity-first headings fit the pattern. Front-load the entity and category in titles and H2s.
A gap analysis helps here. Tools like Promptwatch's content gap analysis map your pages against the prompts AI answers, and the Academy's content gap guide explains the method.
2. Treat your own domain as a retrieval surface
With 1 in 6 fan-outs now scoped by site:, your domain is searched directly. Thin category pages, broken internal search, and unindexed content cost you answers in a way they did not before. Check that key pages are crawlable and indexed, and look at what AI bots actually hit. AI crawler logs show whether bots read a page and where they fail, which separates an access problem from a content problem. The free Agentic Readiness Check is a quick first pass.
3. Make it obvious which brand you are
If ChatGPT is hunting for "official" sources, ambiguity hurts. Lily Ray suggests reconsidering whether "Official" or "Official Site" belongs in your homepage title or meta description, especially if you share a name with other entities. It is a modest test, easy to try, and easy to revert.
4. Be where reviews and comparisons live
Fan-outs inject "reviews" and "vs" constantly. That means third-party pages matter: listicles, G2-type directories, and comparison posts. Promptwatch's software citation data shows GitHub, Microsoft's docs, and G2 leading ChatGPT citations for software prompts (Sep 28 to Oct 4, 2026). One caution on social sources: Promptwatch's social citations report shows ChatGPT drew only 0.46% of citations from social platforms in that week while Google AI Overviews drew 12.2%, so Reddit strategy should be weighted by engine.
5. Monitor continuously
Keyword mappings built before August 8 are stale. Re-run fan-out analysis for tracked prompts after major model changes, and treat visibility as a trend line. Platform directories like bestgeosoftware.com list other tools that can track this.
Common mistakes
- Quoting one average. "ChatGPT runs 2.1 searches" was roughly true for one dataset in April and is wrong for most of the current picture.
- Chasing every fan-out. The Reddit debate on this topic has a point: the possible questions are effectively infinite. Cover the clusters (comparison, pricing, alternatives, reviews, location), not each phrase.
- Optimizing only for the typed prompt. ChatGPT's queries share about 13% of words with the prompt in the worst case. Keyword matching alone will miss branches.
- Ignoring Perplexity's different behavior. There, clean keyword-dense pages still do well.
- Mixing up retrieved and cited. More retrieval does not mean more citations.
A simple weekly workflow
- Pick 10 to 20 prompts that map to revenue.
- Capture fan-outs for each (extension or tracking platform).
- Tag each fan-out by type: comparison, pricing, alternatives, review, official-source, local.
- Check your coverage and crawl access for each cluster.
- Publish or refresh one page per gap, with an entity-first heading.
- Re-check citations after a few weeks and repeat.
If you want the broader program around this, the 90-day GEO program in the Promptwatch Academy covers sequencing. And if you would rather have an agency run it, 1001 SEO Media handles GEO as part of its SEO services.
The bottom line
Fan-outs are the closest thing we have to reading ChatGPT's mind, and the reading is changing fast. The August 8 site: shift alone made many pre-August keyword maps obsolete. Collect your own fan-outs, date every number, cover the sub-questions that keep repeating, and keep your own domain easy to retrieve. The browsing model will keep changing; the habit of checking is what holds up.

