Prompt Frequency Analyzer

Run a prompt N times against Gemini with Google Search grounding enabled. Collect and report the frequency of search queries Gemini generates across all runs.

Why Gemini 3 Flash? This is the model that powers Google Search AI Mode and AI Overviews — the most important answer engine for AEO. Running prompts through Gemini 3 Flash with grounding simulates what Google's AI actually does when users ask questions. 20 samples provides reliable frequency distribution for directional insights.

The Retrieval Framework: Influence happens at retrieval, not inside the model. You can't edit a model's training data — but you can enter the "candidate set" the model selects from when it searches the web. Gemini is search-first: it fires real Google Search queries before nearly every answer, making it more influenceable than GPT. This tool reveals the recurring retrieval set — the queries, sources, and themes Gemini consistently draws from. Understanding query frequency is the first step to entering that set.

Usage

GEMINI_API_KEY="$GEMINI_API_KEY" python3 scripts/analyze.py "your prompt here" [--runs 20] [--model gemini-3-flash-preview] [--concurrency 5] [--output text|json]

Run from the skill directory. Resolve scripts/analyze.py relative to this SKILL.md.

Options

• --runs N — Number of times to run the prompt (default: 20; 20 samples gives good directional signal)
• --model NAME — Gemini model to use (default: gemini-3-flash-preview — the model powering Google AI Overviews)
• --concurrency N — Max parallel API calls (default: 5; keep ≤5 to avoid rate limits)
• --output text|json — Output format (default: text)

Output

Reports for each unique search query:

• Frequency percentage (how many runs used that query)
• Raw count
• Intent classification — each query is classified as informational, commercial, navigational, or transactional
• Intent distribution summary — breakdown of query intents across all unique queries
• Top web sources referenced

Intent Classification

Every search query is automatically classified by intent:

• informational — knowledge-seeking queries ("what is X", "how does X work", "X explained")
• commercial — evaluation/comparison queries ("best X", "X vs Y", "top X for", "X review")
• navigational — brand/site-specific queries (contains domain names, "X login", "X website")
• transactional — purchase/action queries ("buy X", "X discount", "X free trial", "download X")

Example output:

Search Query Frequency:
  85% (17/20) [commercial] — best seo tools 2026
  60% (12/20) [informational] — how seo tools work
  40% (8/20) [navigational] — semrush.com features
  20% (4/20) [transactional] — seo tools free trial

Intent Distribution:
  45% informational, 30% commercial, 15% navigational, 10% transactional

Use intent data to understand what kind of content enters the recurring retrieval set. Each intent type maps to a content format the model searches for:

• informational → explanatory/educational content (guides, explainers)
• commercial → comparison/review content (vs pages, best-of lists)
• navigational → brand/product pages (homepages, feature pages)
• transactional → conversion pages (pricing, free trial, download)

If 60% of queries are commercial, the model is searching for comparison content — and that's the content type you need to create to enter the candidate set.

Further Reading

Notes

• Gemini API key must be in GEMINI_API_KEY env var (stored in macOS Keychain under google-api-key)
• Each run is independent — Gemini may use different search queries each time
• Retries failed requests up to 3 times with exponential backoff
• Use --output json for programmatic consumption