How to connect an AI agent to your GEO analytics (MCP and CLI)

An AI agent is only as good as the data it can reach. This is a step-by-step guide to giving Claude, Cursor, ChatGPT, or any agent live, read-only access to your AI-visibility data — so it can report on your GEO performance instead of guessing from stale training data.

Why connect an agent at all

Ask an agent “how is my brand doing in AI search?” with no connection and it will answer confidently and wrongly. To be useful, it needs your real numbers: your visibility score across engines, where competitors beat you, which sources get cited, and what changed. Connecting it once turns a chatbot into an agentic SEO analyst that pulls its own facts.

The blocker with most tools is access — the data lives in a dashboard and the API is gated to an enterprise tier. Visiblee includes API, MCP, and CLI access on every plan (and the free trial), so all three methods below work from day one. Pick whichever fits your agent.

Three ways to connect

• MCP server — the cleanest path for Claude, Cursor, and other MCP clients. The agent gets read-only tools it calls directly.
• CLI — best for coding agents and automation that can run a command and parse JSON.
• Skill — layered on top of either, it teaches the agent what the metrics mean and how to present them.

Option 1: the MCP server (recommended)

The Model Context Protocol is an open standard that lets an AI client call external tools. Connecting Visiblee takes one command:

1. Install the CLI: npm install -g @visiblee/cli
2. Run visiblee mcp install and pick your client (Claude, Cursor, and others). This registers the Visiblee MCP server in that client’s config.
3. Authenticate in the browser when prompted — account-based consent, no API key to copy and paste.
4. Restart the client. The Visiblee tools now appear to the agent.

The agent gets read-only tools such as list_projects, get_visibility, get_competitors, list_sources, list_suggestions, and list_reddit_leads. You then just ask — “summarise my AI visibility and the three biggest gaps” — and it calls the tools and answers.

Option 2: the CLI

If your agent can run shell commands (Claude Code, Cursor, an automation step), the CLI is the simplest path. Everything supports --json so the agent gets structured output to parse:

1. Install: npm install -g @visiblee/cli
2. Log in: visiblee login — opens the browser, no key to paste. (For CI or headless agents, set a VISIBLEE_API_KEY instead.)
3. Query: visiblee visibility --json, visiblee competitors --json, visiblee sources --json, visiblee leads --json, or visiblee suggestions --json.

Point your agent at the CLI once and it can compose these commands itself — pull visibility, cross-reference competitors, and draft an action list in a single run. Run visiblee --help to see every command.

Option 3: the analyst skill

Raw data is necessary but not sufficient — a model needs to know that average rank is better when lower, that mention rate is a fraction, and that an absent high-intent prompt is an opportunity, not a rounding error. Visiblee ships an analyst skill that encodes exactly this: what each metric means, how to interpret changes, and how to present the result as a clean report rather than a wall of numbers. Layer it on top of the MCP or CLI connection and the agent’s answers go from accurate to genuinely useful.

Prompts to try

Once connected, these are good first asks:

• “Summarise my AI visibility this period — score, mention rate, rank, and what changed.”
• “Which high-intent prompts am I absent from where a competitor is named?”
• “Build a share-of-voice table of me versus my top competitors.”
• “Which domains do the engines cite for my category, and is my own site among them?”
• “Give me a prioritised action list to improve my visibility next month.”

FAQ

Which agents can connect? Any MCP client (Claude, Cursor, and more) via the MCP server; any agent that can run a command via the CLI; and anything that can call an HTTP API directly. The API is included on every plan.

Is it safe? The MCP and CLI tools are read-only, and MCP authenticates through your account in the browser rather than a pasted key. An agent can read and analyse your data, not modify it.

Do I need to code? No. The connection is one install command; after that you ask questions in plain English. Coding agents can go further by scripting the CLI, but that’s optional.

What can the agent actually see? Your visibility score, mention rate, and rank across engines; competitor share of voice; cited sources; warm and Reddit leads; and recommended actions — the same data as the dashboard, explained here.