AI & Searchgrowth

Fanout Query Extractor

Reverse-engineer ChatGPTs grounding questions

What it does

The Ahrefs study showed that cited URLs consistently have higher semantic similarity to ChatGPTs internal fanout queries than to the original user prompt. That means if you know which sub-questions the AI asks itself, you can write H2s, slugs, and titles that match them verbatim. Fanout Query Extractor simulates the fanout process via Perplexity Sonar Pro + Gemini in parallel, cross-references the results for higher confidence, and computes cosine similarity between each fanout query and your URL or title. For each gap query, Gemini Lite generates 2-3 editor-ready H2 candidates with rationale. Output: ranked list with smart H2 export (gap queries replaced by AI candidates), CSV, and AI-generated insight on where you cover well and where you have gaps.

The Problem

You optimize titles and H2s against keywords — but ChatGPT matches against fanout queries you can't see.

• AI models decompose the prompt into internal sub-questions before retrieval
• Ahrefs: cited URLs have 89.78% citation rate when the slug matches the fanout (vs 81.11%)
• Traditional keyword research targets the seed prompt, not the sub-questions
• Competitors who anticipate fanout queries get cited more often — you get skipped

The Solution

Simulate the fanout process, semantically score your content, and get AI-written H2s for each gap.

• Perplexity Sonar Pro + Gemini propose fanout queries in parallel
• Cross-referencing yields high/medium confidence scoring per query
• Cosine similarity 0-1 shows how well your URL or title matches each query
• AI generates 2-3 editor-ready H2 candidates per gap query — not verbatim "## query"
• Export: smart H2 markdown (gap queries replaced by AI candidates), or CSV for tracking

Features

• Dual-model fanout: Perplexity and Gemini propose sub-questions in parallel. Queries both models propose get high confidence.
• Cosine similarity scoring: OpenAI text-embedding-3-small embeds each query and your reference — you see exactly which you cover (≥0.65) and where you have gaps (<0.40).
• AI-generated H2 suggestions: For each gap and partial query, Gemini Lite produces 2-3 editor-ready H2 candidates (max 75 chars) with rationale — you get finished headlines, not just questions.
• Smart H2 + CSV export: Markdown export automatically replaces gap queries with AI candidates and preserves the original question as a comment. CSV to track changes over time.
• Market-aware: Choose Norway/Sweden/Denmark/Finland/English for geo-targeted fanout suggestions based on local SERP context.

How it works

1. Write the seed prompt: The question a user would actually ask an AI assistant about your topic.
2. Add a reference (optional): URL or title to compute semantic similarity AND activate AI-generated H2 suggestions for gap queries.
3. We simulate the fanout process: Perplexity Sonar Pro + Gemini propose sub-questions in parallel, cross-reference results, and Gemini Lite generates H2 candidates for worst-similarity queries.
4. Use ranked queries and H2s: Copy finished H2 headings with one click, export smart markdown, or open each query directly in Google for competitor analysis.

FAQ

Are these ChatGPTs actual internal fanout queries?

No — this is a high-quality simulation. Real fanout queries can only be captured via the DevTools trick on chatgpt.com (manually). We use Perplexity Sonar Pro + Gemini in parallel and score queries both models agree on as high confidence. The tool is explicit that it's simulation, not interception. The Ahrefs study shows that simulated fanout via the same principle provides sufficient signal to improve titles and H2s.

What is a good cosine similarity score?

The strategy doc sets ≥0.65 as the target for "well covered". 0.40-0.65 is "partially covered" (consider adding an H2). Below 0.40 is a real gap — your content doesn't cover the semantic space of that query. Average similarity ≥0.65 across top 8 fanout queries is a concrete KPI from the Ahrefs study.

How do the AI H2 suggestions work?

When you provide a reference URL or title, we pick the 5 worst-similarity gap and partial queries and send them to Gemini Lite with an editorial system prompt: "write H2s that answer this question, don't repeat the question". You get 2-3 candidates per query with rationale (the angle), max 75 chars (Google snippet limit). Copy each candidate with one click, or use smart H2 export where gap queries are automatically replaced by AI candidates. Without a reference we skip this step (no similarity = no gap detection).

Which embeddings are used?

OpenAI text-embedding-3-small (1536-dim) via our existing generate-embedding Edge Function. This is the same model we use for memory-vault and knowledge-graph, so scoring is consistent. The strategy doc recommends Vertex text-embedding-005 long-term for full GDPR posture — that's on the roadmap.

How many credits does it cost?

30 credits per analysis: 1× Perplexity Sonar Pro + 1× Gemini + 1× Gemini Lite for insight + 1× Gemini Lite for H2 candidates (parallelized with insight to save ~500ms-1s) + up to 16 embeddings. The price is flat whether or not a reference is provided — H2 generation is layered on top of the existing flow at no extra credit cost.

Try Fanout Query Extractor free