Data Methodology

We send structured prompts to 7 AI search providers: ChatGPT (OpenAI), Perplexity, Gemini (Google), AI Overview (Google), AI Mode (Google), Copilot (Microsoft), and Grok (xAI).

Prompts are designed to reflect real user queries across multiple industries, intent types (informational, comparative, transactional), and geographies. Each prompt is localized to the target country and language.

Every monitoring cycle processes thousands of prompts across all providers and countries. Responses are captured in full, including inline citations, source links, and recommendation context.

Each AI response is parsed to extract source citations. We identify:

• Source domains — which websites are cited (e.g., reddit.com, wikipedia.org)
• Citation position — where the source appears in the response (1st, 2nd, 3rd, etc.)
• Citation type — how the source is referenced (direct citation, recommendation, comparison, etc.)
• Source classification — domain type (institutional, media, blog, forum, social, educational)

Citations are normalized to root domains and deduplicated within each response. Position data is recorded to measure source prominence, not just presence.

All data is segmented by provider (which AI engine gave the response) and country (the geographic context of the prompt). This allows us to answer questions like:

• Which sources does ChatGPT cite most in France vs. the US?
• Does Perplexity favor different sources than Gemini for the same query?
• How does source diversity vary across AI providers?

The Data Hub currently covers 15+ countries with localized prompts in their respective languages.

Key metrics computed from the raw citation data:

• Mention Rate — percentage of responses that cite a given source
• Average Position — mean citation rank across all appearances (lower = more prominent)
• Position Score — normalized 0-100 score derived from average position for easier comparison
• Provider Coverage — how many of the 7 providers cite a source
• Source Diversity Index — how evenly citations are distributed across sources
• Trigger Rates — how often AI features (shopping, local results, images) are activated

For the GEO Index brand rankings, we additionally compute a composite GEO Score weighted across mention rate (35%), average position (25%), provider coverage (20%), sentiment (10%), and prompt type breadth (10%).

Data is refreshed on a weekly cadence. Each update cycle:

• Runs all prompts across all providers and countries
• Extracts and classifies new citations
• Recomputes all aggregate metrics (top sources, trends, trigger rates, etc.)
• Generates period-specific snapshots (last week, last month, last year, all time)

Historical data is retained to power trend analysis and period-over-period comparisons.

We believe in being upfront about what our data can and cannot tell you:

• Sample-based — our data represents a sample of AI responses, not the complete universe of all queries ever made
• AI responses change — the same prompt can produce different citations at different times. We capture snapshots, not continuous streams
• Prompt design matters — citation patterns depend on how prompts are worded. We use diverse, realistic prompts but cannot cover every possible query
• Provider APIs evolve — AI providers frequently update their models and citation behaviors. Historical comparisons should account for model changes
• No personal data — all prompts are generic research queries. We do not use personalized or logged-in sessions