# Query fan-out generator > For the complete documentation index, see [llms.txt](https://attensira.com/llms.txt). > Every page on this site is also served as Markdown: append `.md` to any URL, or send `Accept: text/markdown`. > Type one prompt and see the sub-queries a model would expand it into, grouped by facet: pricing, comparison, use case and more. Modelled. Free. - Canonical page: https://attensira.com/tools/query-fan-out - Source: converted from this page's own HTML. It carries no claim records, because the page is not a content-pipeline entity. The query fan-out generator takes one prompt and shows the sub-queries a model says it would expand it into before it retrieves anything, grouped by the facet each explores: pricing, comparison, use case, integration and the rest. The sub-queries are modelled, not intercepted, so read them as likely, not certain. Free, no sign-up. Fan-out 10 sub-queries · 9 of 9 facets Prompt best AI visibility tool for a small marketing team - Evaluation · 2: top ai visibility tools for small marketing teams customer reviews of ai visibility tools for marketing - Definition: ai visibility tools definition and features - Capability: capabilities of ai visibility tools for marketing teams - Comparison: comparison of ai visibility tools for small marketing teams - Alternatives: alternatives to ai visibility tools for marketing teams - Pricing: pricing of ai visibility tools for small teams - Use case: use cases of ai visibility tools in marketing - Integration: integration of ai visibility tools with marketing software - Implementation: implementation steps for ai visibility tools in small teams Show all 9 facets Show the first 4 Query fan-out generator, production, n=1 run, 10 sub-queries · 4 Oct 2026 Or use the free tool below ## Fan out a question Seconds: one model call on our server · Free, no sign-up The prompt to expandShow the fan-out One prompt, up to 24 words. This expands a question; it does not answer one. **These sub-queries are modelled, not intercepted.** We ask a model, under a fixed response schema, what it would expand your prompt into. Some developer APIs do return the searches their own model ran for that call, but none gives us what the ChatGPT, Google, Claude or Perplexity apps searched for a person, so we cannot show you those and do not claim to. Or start from an example: best project management toolnotion vs airtable for a small teamcheapest way to add search to a docs sitehipaa compliant video calls for therapists Your result appears here - The sub-queries: the searches it expands into - the facet each one serves: pricing, reviews… - the facets it missed: out of nine See our own run Share ## What does it return? 1. ### The sub-questions Your question split into the separate lookups an answer engine can run before it writes one answer. 2. ### A fixed set of facets Definition, capability, comparison, alternatives, pricing, use case, integration, implementation and evaluation, so two runs compare. 3. ### Anything outside the list Shown under “Other”, labelled with what the model actually said. ## What does one buyer prompt fan out into? We gave it a prompt our own buyers might type: “best AI visibility tool for a small marketing team”. One question about “the best tool” became ten questions about price, reviews, integrations and setup. A page that only answers “which is best” covers one facet of nine. This is one modelled sample, not an engine's real expansion, which nobody outside the engine can read. Treat the spread of facets as the finding and the exact wording as noise. The fan-out is the panel at the top of this page. Read on 4 Oct 2026. About this tool - You give it: One prompt, up to 24 words - It runs: Seconds: one model call on our server - It keeps: About a day in a shared cache keyed by the prompt; nothing tied to you - Cost: Free, no sign-up - Our own run: best AI visibility tool for a small marketing team · 4 Oct 2026 Details ## What is query fan-out? An answer engine does not have to retrieve against the sentence you typed. It can decompose the question first (into a definition lookup, a comparison, a price check, a “does it work with X”), retrieve for each, and write one answer over the results. That decomposition is the fan-out, and it is one reason a page that ranks for your headline question can still be absent from the answer: the retrieval that mattered was for a sub-question your page never addresses. The facet names are ours, not any engine’s. We hold the model to a closed list (definition, capability, comparison, alternatives, pricing, use case, integration, implementation, evaluation) so that two runs are comparable to each other. An open facet field would let the model invent a new grouping every time, which reads as insight and is noise. Anything it returns outside the list is shown under “Other” and labelled with what it actually said. The facets - Facets: Nine, from a closed list - Anything outside the list: Shown under “Other” - The sub-queries are: Modelled, not intercepted ## Who writes the page for a facet nobody answers? Inside Attensira, the prompts you track are read in every engine your plan includes, and a page that is missing an answer is drafted for you to approve. [Tracked prompts](https://docs.attensira.com/tracking/prompts). Software that does your AI-search visibility work, with you in control of every change. ## Why does fan-out matter for which page gets cited? Classic search rewarded one page for one query. An engine that splits one question across several retrievals can cite whichever source answered each part best, so the page that gets cited is not always the page that best answers the whole question. It can be the page that answers one slice of it cleanly enough to be lifted. No engine publishes how it does this, so treat it as a working model, not a mechanism. That has a practical consequence for how you write. A long page that covers nine facets shallowly gives each sub-query a weak match; a page that answers one facet completely gives one sub-query a strong one. The fan-out is the closest thing to an outline of what those pages should be. Once they exist, the question stops being “do we rank” and becomes [which prompts name us and which name someone else](https://attensira.com/features/prompt-tracking). In practice - A page cited for a slice: Answers one facet completely - Engines publish the method: No: treat it as a working model - The next question: [Which prompts name you](https://attensira.com/features/prompt-tracking) ## What do you do with the spread? The useful reading is rarely the wording of one sub-query. It is the spread across facets. | What it shows | Why | What to do | | --- | --- | --- | | A facet with several sub-queries and no page of yours | Why: The expansion keeps asking about it, so that retrieval is one you cannot win yet. | What to do: Write the page. The sub-query is the brief. | | A comparison you did not choose | Why: The model considers that competitor adjacent to you. Being named beside someone is not a ranking. | What to do: Treat it as the comparison page a reader is one hop away from. | | A facet the model never reached | Why: One run is one sample. Missing from every run, it says the prompt does not carry that dimension. | What to do: Run it two or three times. Then ask whether the prompt is narrower than the market it stands for. | | A sub-query you disagree with | Why: The phrasing of the prompt pulled the model somewhere you did not intend. | What to do: Keep it as information: a real person typing that prompt would be pulled the same way. | ## Questions people ask about query fan-out What it will not claim - A modelled fan-out from one general model, not any engine's real expansion. - The facet names are ours, held to a closed list so runs compare. - No search volume: nothing here says how often a sub-query is asked. ### Are these the actual queries ChatGPT runs for my prompt? No. The searches the ChatGPT, Google, Claude and Perplexity apps run for a person are not available to us through any API, log or header. Some developer APIs do return the searches their model ran for one API call: [Gemini’s grounding metadata](https://ai.google.dev/gemini-api/docs/google-search) lists its web search queries, and [Claude’s web search tool](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool) returns each query it ran (both read 5 Oct 2026). Those are API runs, not the consumer apps. What this tool shows is one model’s answer to “what would you search for before answering this”: a proxy, labelled as one. ### Why do I get a different list when I run it again? Because the call is non-deterministic. The same prompt can come back with different wording, and sometimes with a different facet spread. We will not present one run as a stable fact, and you should not either: treat it as one sample. The facets that appear every time are the finding; an exact phrase that appeared once is wording, not evidence. ### How many sub-queries should a prompt produce? We ask for between five and twenty and render at most twenty. That is a bound we chose so the output stays readable and one call stays cheap, not a measurement of how many retrievals any engine performs. We have no way to measure that, so we do not report it. Duplicates and unusable items are dropped and counted separately, so a run can show fewer than five. Reading a count here as “ChatGPT runs fourteen searches” would be reading our own bound back as a fact about someone else’s system. ### Can I use these as the prompts I track? That is the intended use, with one filter: keep the ones a real buyer would type. A modelled expansion produces retrieval-shaped fragments, and some of them are internal steps rather than things a person would ask. The ones worth tracking are the ones you would be glad to be the answer to. [Topics group related prompts together](https://attensira.com/features/prompt-tracking) so a facet is tracked as a set rather than as fifteen unrelated strings. ### Do you store the prompts I type here? Not against you. There is no account, no email field and nothing here tied to an identity. The endpoint is rate limited by the address the request arrived from, and that is the only thing about you it reads. The expansion is held for about a day in a shared cache keyed by the prompt, which is what keeps a popular prompt costing one model call rather than one per visitor. So a prompt somebody else already ran comes back from the cache, and the shareable link may replay that rather than expanding again. ## This shows the facets. Attensira's crew writes the missing pages.