AgentGeoScore: A Website GEO Scoring Tool Built for the AI Era

As large language models like ChatGPT, Claude, and Gemini become the primary entry points for users to access information, traditional SEO is evolving toward GEO (Generative Engine Optimization). AgentGeoScore is an open-source GEO scoring system designed to help websites evaluate their visibility and citation probability among AI agents, providing a complete scoring framework and optimization recommendations.

Traditional SEO focuses on the ranking of web pages in Search Engine Results Pages (SERP). However, in the AI era, when users obtain answers through conversations, the way websites are cited has undergone a fundamental change: AI directly generates answers and selectively cites sources. The optimization goal has shifted from "being indexed" to "being understood and cited". AgentGeoScore simulates the behavior of mainstream AI agents to evaluate the actual performance of websites.

It adopts a front-end and back-end separation architecture:

• Backend (FastAPI + Pydantic): Provides API services, including a scanner (5-dimensional detection), probes (integrated with AI services like Gemini/Mistral), a scoring engine (0-100 weighted score), and a repair recommendation module;
• Frontend (Vite + React + TypeScript + Tailwind): Provides visual reports (score cards, category details, repair lists, llms.txt preview).

The scoring system covers key links in AI discovery and understanding of websites:

1. Agent Access (25%): Detects AI crawler accessibility (robots.txt, CDN restrictions, HTTPS, etc.);
2. Discoverability (20%): Evaluates AI-friendly discovery mechanisms (llms.txt, sitemap.xml, metadata, etc.);
3. Structured Data (20%): Checks the completeness of JSON-LD/schema.org tags;
4. Content Clarity (15%): Analyzes content clarity (titles, meta descriptions, semantic tags, etc.);
5. Citation Probe (20%): Initiates queries to mainstream AI services and counts the domain citation ratio (practical test of visibility).

Supports multiple deployment methods: locally manage Python dependencies via uv and build the frontend via npm; the API is concise (POST URL to get reports). In production environments, the LLM probe supports graceful degradation (automatically skips if the key is missing), and AI services provide free quotas. The system automatically generates llms.txt recommendations (standard practice for GEO optimization).

Privacy considerations in design:

• Scan results are not persistently stored;
• No user authentication required, no cookies or tracking mechanisms;
• API keys are only used for backend calls and not exposed to the frontend/third parties.

AgentGeoScore is open-sourced under the MIT license, with a clear code structure and comprehensive test coverage (unit/frontend/end-to-end tests). As the AI agent ecosystem develops, GEO will become a mandatory course for website operations. The tool not only provides scoring but also serves as a methodology to help examine the discoverability and citeability of websites from an AI perspective, supporting AI-native content strategies.