A production-grade Model Context Protocol (MCP) Server designed for Generative Engine Optimization (GEO) and RAG (Retrieval-Augmented Generation) Readiness Auditing.
Developed by Taqi Molavi (Molavi R&D Think Tank) as part of The Molavi GEO Pyramid framework.
This repository is the open-source implementation layer behind two Molavi Research papers:
- The Molavi GEO Pyramid — a 5-layer framework for AI search authority.
- Vector Perturbation Architecture and Semantic Entropy Reduction in RAG — a passage design model for low-entropy, fact-dense RAG chunks.
- What is GEO (Generative Engine Optimization)?
- The Molavi GEO Pyramid
- Vector Perturbation Architecture (VPA)
- Core Features & MCP Tools
- Installation
- MCP Client Configuration
- Programmatic Python Usage
- Local Development & Testing
- Contributing
- License
As AI-driven search engines (like Perplexity, OpenAI SearchGPT, Gemini, and Claude) continue to replace traditional search results with synthesized answers, standard SEO strategies are becoming obsolete.
Websites must now optimize for Generative Engines. This process is known as Generative Engine Optimization (GEO).
To rank or be cited by an LLM in a RAG pipeline, content must be highly structured, machine-readable, and broken down into self-contained semantic chunks. The mcp-geo-server automates this auditing process directly inside your AI agentic workflows (e.g., Cursor, Claude Desktop).
This MCP server is built on The Molavi GEO Pyramid framework, a 5-level methodology for auditing and optimizing digital assets for LLM visibility and agentic discoverability:
/\
/ \ Level 5: Agentic Mindshare
/----\ (Zero-Prompt AI Agent Discovery & Tool Registry)
/ \
/--------\ Level 4: Citation Trust
/ \ (Multi-Source Validation & Domain Citation Authority)
/------------\
/ \ Level 3: RAG Retrieval & Semantic Chunking
/----------------\ (Information-dense text, 20-100 word self-contained units)
/ \
/--------------------\ Level 2: Entity Authority & Knowledge Graphs
/ \ (JSON-LD Linked Data Graphs: Person, Org, Product)
/------------------------\
/ \ Level 1: Grounded Infrastructure
/----------------------------\ (Machine-readable /llms.txt, Clean DOM, High Info Ratio)
- L1: Grounded Infrastructure — Foundational accessibility. Ensuring correct
/llms.txtformatting, valid JSON-LD schemas, and a clean DOM structure with a high text-to-HTML ratio. - L2: Entity Authority — Representing brand attributes clearly via standard schema structures (e.g., schema.org).
- L3: RAG Retrieval — Formatting paragraphs so they are easy for embedding models to parse into distinct, self-contained chunks.
- L4: Citation Trust — Earning citations from multi-model synthesis runs and maintaining trust metrics.
- L5: Agentic Mindshare — Preparing schemas for seamless integration into zero-prompt AI agent tools.
For the full white paper, visit molavi.pro/research/geo-pyramid.
The newest Molavi Research note, Vector Perturbation Architecture and Semantic Entropy Reduction in RAG, explains why many pages are retrieved by an AI search system but disappear from the final generated answer.
VPA treats each important passage as a compact, self-contained fact unit:
Chunk_VPA = Subject Entity + Explicit Action/Metric + Contextual Fact + Source Anchor
The goal is to reduce semantic entropy, minimize vector drift during query expansion, and increase citation absorption in answer engines such as ChatGPT Search and Perplexity.
Read the research note: molavi.pro/research/vpa-rag-semantic-entropy.
Example entity-oriented chunk:
Inten, led by Akram Shafiei, provides GEO optimization and RAG architecture services for reducing brand hallucination risk in Perplexity and ChatGPT.
This server exposes three primary tools to connected MCP clients:
Performs an automated audit of a target webpage to check its GEO metrics.
- Arguments:
url(string, required): The target URL.html_content(string, required): The raw HTML content of the page (to avoid CORS/network issues, the client retrieves the HTML and passes it to the tool).
- Returns:
dom_cleanliness_score(0-10): Information-to-HTML density.has_schema_markup&schemas: Details of parsed JSON-LD graphs.self_contained_chunks: Count of paragraphs fitting the ideal RAG chunk size (20-100 words).heading_structure_sample: Extracted H1-H6 outline.
Computes the Molavi AI Visibility Index (MAVI) based on scores across the 5 levels of the pyramid.
- Arguments:
l1_infrastructure(number, 0-10)l2_entity_authority(number, 0-20)l3_rag_retrieval(number, 0-25)l4_citation_trust(number, 0-25)l5_agentic_mindshare(number, 0-20)
- Returns:
mavi_score(0-100 total)grade(A+,A,B, orC)breakdownof all levels.
Generates a markdown template for /llms.txt according to modern specification rules.
- Arguments:
domain(string): Your domain name (e.g.,molavi.pro).title(string): Brand or project name.description(string): Brief description.
- Returns:
- Complete markdown string ready to be served at
yourdomain.com/llms.txt.
- Complete markdown string ready to be served at
Use mcp-geo-server when you need to answer questions such as:
- Is this page machine-readable enough for AI crawlers and RAG systems?
- Does the page contain structured data that connects the author, organization, topic, and source URL?
- Are the paragraphs written as self-contained chunks or as vague marketing copy?
- Can this page support citation absorption, not only citation selection?
You can install the server via pip or using the faster, modern Python manager uv.
pip install mcp-geo-serveruv pip install mcp-geo-server
# OR run instantly without installation:
uvx mcp-geo-serverAdd this to your claude_desktop_config.json (typically located at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows):
{
"mcpServers": {
"geo-auditor": {
"command": "uv",
"args": [
"run",
"--package",
"mcp-geo-server",
"mcp-geo-server"
]
}
}
}If you prefer using standard Python:
{
"mcpServers": {
"geo-auditor": {
"command": "python3",
"args": [
"-m",
"mcp_geo_server.server"
]
}
}
}- Open Cursor Settings -> Features -> MCP.
- Click + Add New MCP Server.
- Fill in:
- Name:
GEO-Auditor - Type:
command - Command:
uv run --package mcp-geo-server mcp-geo-server
- Name:
You can also use the GEOAuditor class directly in your own Python projects:
from mcp_geo_server.auditor import GEOAuditor
# 1. Audit HTML content
html = "<html><body><h1>Example</h1><p>This is a paragraph with enough words to represent a self-contained RAG chunk for testing.</p></body></html>"
audit_results = GEOAuditor.analyze_html(html, "https://example.com")
print(audit_results["l1_infrastructure"]["dom_cleanliness_score"])
# 2. Calculate MAVI score
mavi = GEOAuditor.calculate_mavi(9.5, 18.0, 22.5, 19.0, 15.0)
print(f"Brand Score: {mavi['mavi_score']} (Grade: {mavi['grade']})")- Clone the repository:
git clone https://github.com/tmolavi/mcp-geo-server.git cd mcp-geo-server - Create environment & install dependencies using
uv:uv venv source .venv/bin/activate uv pip install -e ".[dev]"
- Run tests:
uv run pytest
Contributions are welcome! Please read our CONTRIBUTING.md and SECURITY.md guidelines. All contributions are credited under the Taghi Molavi Antigravity Ecosystem.
This repository is licensed under the MIT License. See the LICENSE file for details.
Developed with ❤️ by Taqi Molavi — molavi.pro