Documentation
Query fan-out
How to see what the AI actually searched for on your prompts — and how firmly your brand sits inside its own queries
Contents
1.What fan-out is and why it matters
An AI engine almost never searches your question verbatim. Before answering, it breaks the question into a series of its own search queries — sub-queries — and gathers sources from those. The "Query Fan-Out" section shows those sub-queries: the actual wordings the AI took to search while answering your prompts.
- A single prompt expands into several sub-queries: the model translates the question, adds a year, a city, a category — and often brand names you never mentioned
- This is the closest thing to classic SEO search queries, except the wording comes from the model, not the user — and it decides whose pages make it into the answer
- Sub-queries are captured automatically on every monitoring run for the providers that expose their search: Yandex (Search with Alice), Alice AI, ChatGPT, Perplexity, Gemini, Claude, DeepSeek
- The section obeys the global filters up top — period, providers, prompts: look at fan-out across the whole brand or narrow it to a single topic
- A specific prompt's sub-queries also live in its own card, on a separate fan-out tab — handy when you're digging into why you weren't shown on that topic
When a sub-query repeats your prompt word for word, the model simply searched it as-is. The valuable rows are the ones where it added something of its own: a year, a region, "reviews", "comparison", or a competitor's name. Those are the queries that actually decide the answer.
2.Brands in sub-queries: how well the model knows you and your competitors
The "Brands in sub-queries" card counts how often the AI, unprompted, puts a brand name into its own search query. This isn't about being mentioned in the answer or cited as a source — it's about what the model pulls from its own memory when deciding who is even worth checking on your topic. In effect, a measurable LLMO signal: how firmly the brand is embedded in the model's knowledge.
- On the left, "Frequent phrases": recurring word combinations from sub-queries, with the number showing how many of your prompts contain the phrase. It's the topic's vocabulary as the model sees it
- On the right, you against your competitors. In the example the brand name landed in 36 of 3,190 sub-queries (1%), competitor names in 620 (19%). A 19x gap means the model treats other names as the anchors of this niche
- "Top competitors in sub-queries" ranks who the AI substitutes most often (Alfa-Capital — 161, VTB — 93, VIM Investments and UK Pervaya — 54 each). These are the brands the model starts its search from on your topic
- Watch your share relative to competitors and how it moves, not the absolute number. Zero sub-queries with your name while competitors are active is a direct signal that the model doesn't know you in this category
- "Brands in sub-queries over time" shows the same data by day in two modes — "Share, %" and "Occurrences" — for you and the top 5 competitors on one chart: you can see whether you're gaining weight in the model's memory or losing it
- There's a warning icon next to the cards for good reason: branded prompts inflate the metric (you named the brand yourself, so the model echoed it into search), and names that double as common words sometimes match ordinary text. The cleanest read comes from non-branded prompts
The goal isn't to appear in a sub-query once — it's to close the gap with the top competitors. The same thing works here as with human brand awareness: reviews, comparisons and roundups on third-party sites where your name sits next to the category leaders — that's exactly the text the model learns the niche from.
3.The sub-query table: what to do with it
Below the charts is a table of every sub-query for the period: the wording itself, the provider that generated it, and how many times it came up. It's a ready-made content backlog written by the AI engines rather than by you or a keyword tool.
- "Runs" is how many times the wording appeared in the period. Sorting by that column lifts the durable topics to the top and pushes one-offs down
- The provider icon shows whose model phrases it that way: the same topic sounds different across providers — one drills in with follow-up queries, another searches almost verbatim from your prompt
- The table search is the fastest way to pull every wording containing "reviews", "comparison", "price", or a specific competitor's name
- You can lift the wordings as they are into headings, H2s and FAQ blocks: these are literally the strings the model uses to look for sources
- The filters up top apply here too: pick a single prompt to see only its fan-out; the same table is available inside the prompt's own card
A simple routine: sort by "Runs", walk the top 20–30 wordings, and honestly ask whether you have a page that answers exactly that. Every frequent wording you don't cover is a topic the model is already asking about — and finding the answer somewhere else.