--- title: "Reading Query Fan-Out Like a Map" description: "When someone asks an AI engine a question, the engine fans it out into many sub-queries and stitches the answer from whatever sources win each one. The practical unit of AI visibility is that map of sub-queries, and how ClarAI's 12-axis Query Fusion Model plus per-prompt reverse-engineering make it operable." published: 2026-08-15 reading_minutes: 8 tags: ["How AI search works", "Measurement"] canonical: https://myclarai.com/blog/reading-query-fan-out-like-a-map source: ClarAI blog --- # Reading Query Fan-Out Like a Map _When someone asks an AI engine a question, the engine fans it out into many sub-queries and stitches the answer from whatever sources win each one. The practical unit of AI visibility is that map of sub-queries, and how ClarAI's 12-axis Query Fusion Model plus per-prompt reverse-engineering make it operable._ When a buyer asks an AI engine a question, the engine does not run that one query. Google's own guidance describes AI Overviews and AI Mode as running a query fan-out: they decompose the single question into many related sub-queries, run those against the Search index, then synthesise one answer from whatever sources win each branch. So the practical unit of AI visibility is not the keyword, it is the map of sub-queries your buyer's question fans into. Read that map and you can see which sub-queries you already win, which a competitor owns, and which are unclaimed, then aim content at the gaps. This post covers the mechanism Google published, why the map beats the keyword, how ClarAI draws and reads that map, and where the map stops being able to promise anything. ## What Google means by query fan-out Google's official guide, Optimizing your website for generative AI features on Google Search, describes AI Overviews and AI Mode as retrieval-augmented generation over the ordinary Search index, expanded by query fan-out. The engine does not answer from memory. It decomposes the buyer's single question into many related sub-queries, issues them concurrently against the ranked index, and assembles one answer from the pages that win those sub-queries, citing the sources it leaned on. One question quietly becomes many: comparisons, alternatives, follow-ups, persona variants, regional phrasings, and format-specific asks like 'checklist' or 'template'. None of those hidden queries is visible to the buyer, and none of them is the head keyword you were optimising for. Google is equally clear that the internal machinery stays internal. That single sentence settles what any honest tool can be in this category: an observer of the output, never an insider to the ranking. > No third-party tool has access to our internal ranking or AI systems. > > Google Search Central, Optimizing your website for generative AI features on Google Search (2026) So the object of study is the observable layer: the answers engines actually give and the pages they actually cite. The fan-out is the shape of that layer. You cannot read Google's decomposition from the inside, but you can model the sub-queries a question is likely to fan into, and you can measure which sources keep winning them across runs. That measurable shape is the map. ## Why the map beats the keyword Two consequences follow from fan-out, and both break keyword thinking. First, a page that would never rank position one for the head query can still be pulled into a cited answer through a fan-out branch, which is why long-tail and comparison content earns citations it never earned clicks for. Second, optimising for a single keyword misses the mechanism entirely: the unit of competition is the whole fan-out set around a buyer question, not one query. Win the head term and you can still lose every branch to competitors, or lose the head term and still get cited through a comparison or a persona variant. Think of each buyer question as a territory rather than a target. The territory is made of tiles, one per sub-query, and different sources win different tiles: a listicle owns the 'best X' tile, a Reddit thread owns the 'is X worth it' tile, a vendor page owns the 'X pricing' tile. A keyword report tells you your rank on one square. A map tells you the whole territory: which tiles you hold, which a rival holds, and which nobody has claimed. The gaps are where the next piece of content should go, because an unclaimed tile is cheaper to take than a contested one. ## Drawing the map: the 12-axis Query Fusion Model To read a map you first have to draw it, and drawing it means turning one buyer question into the spread of sub-queries it fans into. ClarAI ships a Query Fusion Model that expands a seed question along twelve axes of variation, so a single topic becomes a structured set of prompts instead of one keyword. - Intent stage: awareness, consideration, decision, or support. - Modality: informational, transactional, navigational, or commercial. - Buyer persona: the role inside the buying group, from founder to individual contributor to buying committee. - Tense: present versus temporal framing like 'best X in 2026'. - Comparison axis: single-brand, head-to-head, or market-map framing. - Geographic specificity: global, regional, or city-level scope. - Industry vertical: the sector the question sits in. - Company size band: SMB, mid-market, or enterprise, which shifts feature priority. - Use-case granularity: a broad workflow versus a specific job to be done. - Pricing dimension: free, freemium, paid, or a specific tier. - Negation: alternatives-to, replacing, or leaving a named tool. - Format intent: the surface the searcher expects, such as guide, checklist, template, comparison, or review. Those twelve axes are a coordinate system, not a claim about Google's internal decomposition. ClarAI uses a small model to expand a seed plus a chosen set of axes into natural-language prompts a real person would type, and falls back to a deterministic local template when the model is unavailable, labelling the output honestly when it does. The point is not to guess Google's exact fan-out, which nobody outside Google can see, but to enumerate the plausible sub-queries a buyer question covers so you have a map to check yourself against. ## Reading the map: who owns each tile Drawing the map is half the work. The other half is reading which tiles you actually win, and that is where ClarAI's per-prompt reverse-engineering comes in. For a given prompt it models the fan-out across the major engines a few times, three by default and up to five, and separately reads the citations that actually came back from the engines it monitors: ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. Sampling more than once matters because the answers move between runs, so a single reading proves nothing and a repeated one starts to show a stable signal. It then classifies the kind of source winning each branch, and when one source type crosses a 70 per cent share of the citations, it treats that as the dominant pattern and maps it to a concrete action: - Listicles dominate: earn mentions in the lists that already rank, rather than writing your own. - Vendor pages dominate and you do not have one: create the missing page. - Reddit dominates: engage in the threads, since no page you publish will win that tile. - YouTube dominates: earn video coverage by pitching or sponsoring creators. - Review sites dominate: drive customer reviews on the platform being cited. That is what reading the map looks like in practice: not the flat question 'am I mentioned', but 'who owns this tile, and what does taking it actually require'. The 70 per cent threshold is deliberately blunt because the answer to 'what should I do' should not flip on a single citation. When no source type dominates, the recommendation defaults to writing a new, format-correct page tuned to that prompt, because a scattered citation pattern is an opening rather than a wall. ## What the map cannot promise A map is only as honest as its survey, and this one is drawn on moving ground. AI answers are nondeterministic: the same prompt sent to the same engine returns different answers on different runs, so every tile ClarAI colours in is a sample of a target that shifts. That is exactly why the sampling is repeated and why the interface discloses coverage in the '4 of 72 prompts measured' form, so a thin reading can never quietly impersonate a full survey. A small movement between two runs is usually noise, not a win or a loss, and should be read that way. Two honesty lines matter most. The fan-out ClarAI shows you is a model of how engines expand a question, not a readout of Google's internal decomposition, because Google has stated plainly that no third party has that access. The citation side, by contrast, is measured from real answers ClarAI captured across the engines it monitors, but it is still a sample: capped, cached, and only meaningful once enough answers have been measured for a prompt. The map is a tool for aiming, not a guarantee of what any engine will say tonight. ## How to read your own map this week 1. Pick one buyer question that matters to revenue, not one head keyword. 2. Fan it out along the axes: comparisons, alternatives, persona variants, pricing questions, and region variants, so you have the whole territory in front of you. 3. For each sub-query, check who is actually cited in the AI answers, and note the source type: a listicle, a vendor page, a Reddit thread, a video, or a review site. 4. Mark the tiles you already win, the tiles a competitor owns, and the tiles nobody has claimed. 5. Match the action to the dominant source type rather than defaulting to writing another blog post: earn a mention, engage a thread, or drive reviews when that is what the tile rewards. 6. Take the unclaimed tiles first, and re-sample after a couple of weeks to confirm a real shift rather than a single lucky run. The frame that survives all of this is simple. Fan-out means your buyer's question is a territory, not a target, and the sources that win its tiles are visible if you sample them honestly. Draw the map, read who owns it, and put your next piece of content where the map says the gap is. **See the map for your own brand** Run a free ClarAI check on your domain and get real AI answers and real citations, so you can see which sub-queries you already win before you write another word. [Check your AI visibility](/free-ai-visibility-checker) [Read the blog](/blog)