Fan-outs
The real web searches an AI assistant ran while answering your prompts, lifted word for word from the stored payload, ranked by estimated demand and trackable in one click.
When an AI assistant answers a question, it usually runs its own web searches first. Those searches are demand you can't see anywhere else.
What it does
Fan-outs shows you the queries an assistant actually typed while answering your prompts, lifted word for word from the stored payload and never simulated. Each one is a real question people's assistants ask around your topic. Each one is also a ready-made prompt you can start tracking in one click.
How to use it
Open Fan-outs to see a table of the captured queries. Each row shows:
- The query as the assistant ran it. A verbatim badge means the assistant searched your prompt word for word, an echo of the prompt rather than an expansion of it.
- An estimated AI volume for the query.
- The engines that ran it and the answer count it appeared in.
- The triggering prompts: which of your prompts caused the search.
- A one-click Track, which creates a tracked prompt from the query.
Tracking a fan-out creates a normal prompt, so it counts against your pooled prompt quota like any other.
How it's computed
Queries are grouped by their lowercased key over the selected window. Echoes sort after true expansions. The list is capped at the top 100, and each run contributes up to 20 queries.
Estimated volume comes from a DataForSEO AI-search-volume lookup backed by a three-tier cache. A string that can't be measured becomes an em-dash: "no data", not zero. That follows the data-honesty rule.
Before the lookup, each prompt is turned into candidate keywords to measure. Candidates come first from a deterministic rule: the prompt's own text and head keywords, plus any fan-out queries. Prompts whose deterministic candidates all fail to anchor are the "null-head" residue. They get an LLM fallback tier, a batched temperature-0 model call that proposes extra head keywords. It is gated to the weekly retry lane and cached once per prompt text, so the same prompt is never re-asked. The model only ever proposes candidates. It never invents a volume, band, or confidence. The volume provider decides which candidate anchors, and a prompt whose candidates all fail to anchor stays an em-dash.
Untracked fan-outs also feed a fan-out-opportunity recommendation. A query qualifies when it has at least 2 runs and either an estimated volume of at least 100 or at least 4 runs.
Limits
Only engines that emit the search-query field in their payload fill this screen. Today that is ChatGPT and Perplexity. Volume estimates need DataForSEO credentials. Without them, every volume falls back to an em-dash rather than a guess.
Related
- Shopping demand: the same fan-outs, narrowed to shopping carousels and ranked Top / Trending / Losing / New.
- Prompts: where a tracked fan-out becomes a monitored question.
- Research: demand research that feeds the same suggestion pipeline.
- Actions: where the fan-out-opportunity recommendation appears.