Query Fan-Outs, Prompt Volumes, and Difficulty Scores in ChatGPT and Gemini
A prompt is not one search. It expands into a fan-out of retrieval queries before the answer is written. Volumes and difficulty scores are how you decide which fan-out is worth the work.
A prompt is not one search. The string a user types into ChatGPT or Gemini is the seed, not the whole query. Before the answer reaches the screen, the model expands that seed into a set of retrieval searches, picks sources from those, and only then writes the response. That expansion is the query fan-out. It is the part of AI search that a classic keyword tool never sees, and it is the part that decides which prompts are worth the work and which are noise.
This post explains query fan-outs, prompt volumes, and difficulty scores on their merits, and then lands on the Promptwatch feature that turns them into an acting workflow. The numbers come from the named Promptwatch data report and the Promptwatch fact sheet. Where the data is directional, we say so.
What a query fan-out actually is
When a user types "best CRM for a 12-person agency" into ChatGPT, the model does not run one search for that string. It runs several. It might search for "CRM for small agencies," "CRM pricing small team," "CRM reviews 2026," "HubSpot vs Attio for agencies," and a couple more, then assemble an answer from the sources those searches return. Those background searches are the fan-out. A single prompt can trigger 3 to 8 or more separate retrieval queries, and the set changes with the model, the retrieval system, and the time of day.
The ChatGPT query fanouts report tracks this directly. It shows two things that matter for planning. The first is the fan-out depth, the number of background searches a prompt triggers. The second is the shape of those queries, and the shape has changed dramatically. Average query length dropped from roughly 117 characters in early December 2025 to the high 80s through February and March, then to roughly 53 characters once fanout volume returned in April. ChatGPT is searching more like someone typing keywords into Google and less like someone pasting in a full sentence.
That shift changes the work. If the model is running shorter, keyword-shaped queries, the pages that rank for those keyword-shaped queries are the ones that get retrieved. A page built for a long conversational prompt may not match the short query the model actually runs. The fan-out is the bridge between the prompt the user typed and the query the engine actually executed.
Why the fan-out is the unit of work, not the prompt
A prompt tracker that only stores the seed string is blind to the fan-out. It tells you whether your brand showed up for "best CRM for a 12-person agency." It does not tell you that the model actually ran "CRM reviews 2026" and "HubSpot vs Attio for agencies" and that a comparison post you do not control won the citation on the second query. The fan-out is where the citation is actually won or lost.
This is why a prompt list without fan-outs is a partial measurement. You can rank well on the seed prompt and still lose the citation, because the citation comes from a fan-out query, not the seed. The agent move is to track the fan-out, find the query in the fan-out where a competitor or a third-party listicle won the slot, and either build a page for that query or shape the offsite conversation so a different source wins it.
Prompt volumes: how many people actually ask this
The fan-out tells you how the model searches. The volume tells you whether the prompt is worth chasing. A prompt with a deep fan-out and no volume is a curiosity. A prompt with a shallow fan-out and high volume is a priority. Volume is the demand signal that lets an agent prioritize the prompt set instead of treating all prompts as equal.
Promptwatch prompt volumes are a weighted average over the keywords attached to the prompt, sourced from the prompt volumes methodology: AI search weighted at 5x, Bing at 1.5x, Google at 0.5x, refreshed quarterly at most, banded 1 to 10k-plus as 1 to 5 bars, with "No data" until a keyword has volume. The weighting matters. A prompt that shows up in AI search gets five times the weight of one that shows up in Google, because the question is AI visibility, not classic search demand. A keyword tool that weights Google at 1x and ignores AI search entirely is answering a different question.
The practical use is prioritization. An agent that reads volumes can sort the prompt set by demand and spend the content budget on the prompts that people actually ask, instead of the prompts that sound good in a brainstorm.
Difficulty scores: how hard the slot is to win
Volume tells you whether the prompt is worth chasing. Difficulty tells you whether you can win it. A high-volume prompt with a difficulty score that is out of reach is a poor use of budget. A lower-volume prompt with a difficulty score you can move is a better use of budget. Difficulty is the constraint that turns a demand list into a realistic plan.
Promptwatch difficulty scores are calculated over the past 30 days from keyword competition and the authority of the sources already cited. The logic is direct. If the sources already cited for a prompt are high-authority domains, the slot is harder to take. If the sources are mid-authority or volatile, the slot is more contestable. An agent that reads difficulty can tell a client "this prompt is worth chasing and winnable" versus "this prompt is worth chasing and not winnable this quarter," which is the distinction that separates a plan from a wish.
ChatGPT and Gemini: why the split matters
The title names ChatGPT and Gemini because they behave differently, and a fan-out that wins on one can lose on the other. ChatGPT runs roughly five sources per web-search response, per the average sources per response report. Google AI Overviews and Perplexity run around ten. Gemini's consumer search behavior largely surfaces through Google AI Overviews and AI Mode, which are the longer-inventory surfaces. The implication is that a fan-out query that misses the ChatGPT shortlist can still make the longer Google list, and a prompt that is too competitive on ChatGPT may be contestable on the Google surfaces.
An agent that runs one fan-out playbook across both engines is leaving the longer inventory on the table. The right move is to measure the fan-out per engine, and to target the engine where the fan-out query is most contestable for the client's domain authority.
The named feature: listQueryFanouts and the prompt tracking layer
The measurement that turns the fan-out from a concept into an acting workflow is the listQueryFanouts tool in the Promptwatch MCP set, backed by the prompt tracking layer that carries volumes, difficulty, and fan-outs on the same prompt row. listQueryFanouts exposes the fan-out structure around a seed query, the related prompts an answer tends to pull in. That is how an agent finds the real query behind a tracked one. The prompt tracking layer attaches the volume and the difficulty score to that row, so the agent can prioritize the fan-out by demand and by contestability, not by guess.
The acting half is the Content Agents layer. When listQueryFanouts shows that a fan-out query is won by a third-party listicle, and the prompt tracking layer shows that the query has volume and a contestable difficulty score, the content gap is the brief for the next article. The review inbox is the gate that keeps a human on the publish step. The loop is closed: the fan-out finds the query, the volume and difficulty prioritize it, the content agent drafts for it, and the human ships it.
How to use the fan-out in a weekly loop
A weekly loop that uses fan-outs well has three steps. Run listQueryFanouts on the prompts that matter, to find the fan-out queries the model actually executed. Read the volumes and difficulty scores on those fan-out queries, to sort them by demand and contestability. Pick the one or two fan-out queries where the client has a real chance, and brief the content agent or the offsite program against those, not against the seed prompt. The seed prompt is what the user typed. The fan-out query is what the engine retrieved. The work is on the second one.
The fan-out is the part of AI search that a keyword tool cannot give you, because it is not a keyword. It is the expansion of a prompt into the retrieval queries an engine runs behind the scenes. The volume and the difficulty are the parts that turn that expansion into a prioritized plan. Together they are the prompt layer that separates a tracker from a platform, and the named feature that does the work is listQueryFanouts on the Promptwatch MCP server.