Key takeaways
- Fan-out queries are written by the model, not typed by people. Most have no search volume, so a keyword list built from volume tools will miss them.
- Promptwatch Data shows fan-outs are short (about 53 characters in April 2026) and unstable. Their count, length and operators changed in steps between December 2025 and August 2026.
- Stop treating 3-8+ searches per prompt as typical. Promptwatch's tracked averages ran from roughly 1.0 to 2.15 per response.
- Rebuild your research around prompts, sampled fan-outs and sub-question clusters, and re-sample regularly instead of building the list once.
- Keep your classic keyword research. It still tells you which topics matter. It just stops being the full map.
The mismatch nobody warns you about
Here is the experiment that convinces most people. Take a prompt your customers might type into ChatGPT. Capture the web searches it runs behind the scenes. Paste those into your keyword tool.
Most of them come back empty.
Ahrefs ran a version of this and showed a list of ChatGPT fan-out queries entered into Keywords Explorer, where only one out of eleven had search volume. Ahrefs explains that keywords not in its database are usually excluded because of extremely low search interest. That is the heart of the problem. Your keyword list is a record of what humans typed into search boxes. Fan-outs are what a model decided to search on a human's behalf. Those are two different populations that overlap less than you would hope.
I find this both obvious and slightly annoying. Obvious, because of course a model phrases things differently. Annoying, because an entire workflow (volume, difficulty, rank, repeat) quietly assumes the query is something a person typed.
What a fan-out query actually is
When you ask an AI engine something complex, it often splits the question into several searches, runs them, reads the results and writes a combined answer. Conductor describes it as the AI creating a list of requirements from your question. Ahrefs puts the human side well: searchers care about context they rarely put into words, and they mentally stack requirements without typing them.
So a prompt like "which CRM suits a 50-person B2B SaaS company" might trigger searches on pricing, integrations, comparisons and reviews. Omnius cites a Writesonic test where GPT-5.4 Thinking ran eight web searches before answering that kind of question.
The important point from Promptwatch's ChatGPT query fan-out data is the framing: your content does not need to match the user's prompt. It needs to match one of the queries ChatGPT generates from that prompt.
Why fan-outs rarely match your keyword list
Four reasons, and they stack.
1. They are model-written, not human-typed
Volume tools estimate demand from clickstream and search engine data. A model-generated query like "CRM pricing 50 users integrations 2026" may be perfectly sensible and still have zero recorded human searches. It exists because the model needed it, not because a person wanted it.
2. They are specific in ways humans rarely are
The iPullRank write-up on fan-out notes that a document might not match the user's exact phrasing but could answer a generalized version of the question, or a more specific variant the user did not think to ask. Fan-outs generate those variants on purpose.
3. They are short and keyword-like, which is deceptive
Promptwatch Data tracked average fan-out length at about 117 characters in early December 2025, the high 80s through February and March 2026, and about 53 characters once fan-outs returned in April. That is roughly six to eight words. Short enough to look like a normal keyword, which tempts you to assume your list already covers it. It usually doesn't, because the wording is the model's.
4. They move
This is the one most guides skip. In Promptwatch's tracking, average fan-outs per response were 2.15 in early December 2025, eased to about 1.84 by early March 2026, and came back in April at exactly 1.0. The report attributes the drops to steps tied to behavior changes.
Then on August 8, 2026, ChatGPT started using the site: operator at scale. The share of fan-outs using site: jumped from 0.37% to 16.8% in one day, and average fan-outs per response went from about 1.08 to about 1.83. Since then roughly one in six fan-outs is scoped to a single domain.
A keyword mapping built before that date describes a different system.
A note on the numbers you will see quoted
You will read that a prompt triggers "3-8+ searches" or that Google AI Mode produces "5 to 20 subqueries". Treat those as ceilings from specific tests, not averages. Promptwatch's own intro mentions 3-8+, yet its tracked ChatGPT averages sit around 1.0 to 2.15 per response. The 5-20 figure in the MD Marketing Digital piece is described as an industry estimate based on AI Mode behavior, and I haven't verified it against first-party data. Victorious says it plainly: the count depends on the platform and the query, so treat any single number as an example, not a target.
Classic keyword research vs fan-out research
| Dimension | Classic keyword research | Fan-out research |
|---|---|---|
| Query source | Humans typing into a search box | Model-generated searches |
| Main input | Search volume and difficulty | Sampled prompts and captured fan-outs |
| Typical length | Varies widely | About 53 characters in April 2026 (Promptwatch Data) |
| Stability | Changes slowly | Changed in steps, e.g. August 8, 2026 |
| Unit of work | A keyword and a page | A prompt, its sub-questions and a page cluster |
| Refresh cadence | Quarterly or yearly | Continuous, plus after major model updates |
How to rebuild keyword research around fan-outs
Step 1: Keep the keyword list, demote it
Your list still identifies core topics. Victorious is right that keyword research tells you where to start, and you then research the follow-up, comparison and clarifying questions. Think of the old list as the seed, not the map.
Tools like Semrush are still useful here for seed topics and People Also Ask style questions.
Step 2: Write the prompts, not just the keywords
For each core topic, list the prompts a real buyer would type into ChatGPT or AI Mode. Include the messy, multi-constraint ones ("for a 50-person team, under $X, integrates with Y"). Pull from sales calls, support tickets, Reddit threads and Search Console queries. Prompt volume and difficulty data, where available, help you prioritise.
Step 3: Sample the actual fan-outs
Run those prompts and capture the searches the model generates. You can do it by hand with a browser extension or a manual test, but at any scale you want a tool. Promptwatch tracks prompts across ChatGPT, AI Overviews, AI Mode, Gemini and Perplexity from the real interfaces and shows query fan-outs per response, alongside volumes and difficulty.
Other trackers cover parts of this. Peec AI and Otterly.AI are prompt-level trackers, and Profound is a heavier monitoring platform. Check what each exposes about fan-outs before you buy, since I did not verify their feature sets for this guide.
Step 4: Cluster the fan-outs into sub-questions
Group what you captured into the repeating shapes Promptwatch recommends covering: comparisons ("X vs Y"), pricing, alternatives, and how-tos. Omnius suggests ranking recurring gaps by how much each matters to product evaluation, how weak your current answer is, and whether you have real evidence to add. That last filter is the useful one. Don't write a page just because a fan-out exists.
Step 5: Map clusters to pages, then to headings
Promptwatch's recommendations from its fan-out report are practical:
- Write headings that read like search queries, not conversational questions. "Best CRM for small agencies 2026" beats "What should you look for when choosing a CRM?"
- Put the entity and category first in titles and H2s.
- Favor a cluster of focused pages over one mega-page.
This fits the pattern in the MD Marketing Digital piece, which cites a Surfer SEO analysis: ranking for the main query plus at least one subquery accounted for 51.2% of AI Overview citations, versus 19.6% for the main query alone. One caveat: that is a Google surface, not ChatGPT.
Step 6: Make your own domain a retrieval surface
Since the August 8 site: shift, about one in six ChatGPT fan-outs targets a single domain. Promptwatch's advice: thin category pages, broken internal search and unindexed content now cost you answers. Check crawlability and indexation of the pages you want cited. Crawlers like Screaming Frog and Sitebulb help with that audit.
Step 7: Re-sample on a schedule
A one-off audit is stale quickly. Promptwatch suggests re-running fan-out analyses for tracked prompts and re-auditing your keyword-to-heading mapping after major ChatGPT updates. Practically, that means a monthly sample at minimum, plus a rerun any time you see a sudden swing in citations.
What to measure afterwards
A fan-out-driven plan needs different scorecards than rank tracking:
- Fan-out coverage: of the sub-queries captured for your priority prompts, how many do you have a focused page for?
- Citation presence: is your domain cited for those prompts, and where in the source list?
- Crawl and traffic signals: are AI crawlers hitting the pages, and does anything click through?
For the citation side, remember how the sources behave. Promptwatch's content-type data shows product pages as the most cited type in ChatGPT (22.4% of citations, Sep 28 to Oct 4, 2026), and its half-life study found landing pages and docs stayed cited longest while news lost half its pages within 36 days. Fan-outs that favor commercial intent will pull in the page types that do well there.
Common mistakes
- Treating the fan-out list as a fixed keyword list. It's a sample from a moving system.
- Discarding zero-volume fan-outs. Zero volume doesn't mean zero value, since the model is deciding what to cite from them.
- Writing conversational question headings when the fan-outs are terse, keyword-like strings.
- Building one giant page to cover everything instead of a cluster of focused ones.
- Assuming ChatGPT's behavior applies to Google AI Mode or Perplexity. Sample each surface.
Where this leaves your keyword research
Keyword volume tools aren't obsolete. They are measuring human demand, which still matters for Google. But they were never built to describe what a model searches for on someone's behalf, and the numbers show that behavior changes in steps. Keep the seed list, sample the fan-outs, cluster them into sub-questions, and put a recurring re-sample on the calendar. If you want to compare trackers that capture this data, the GEO software directory at bestgeosoftware.com is a good place to start, and the full set of citation research is on the Promptwatch Data hub.