We Built an MCP Server for AI Search Visibility

Your customers stopped typing keywords. They open Claude or ChatGPT and ask a question in a sentence — and the answer they get names two or three sources, none of which they chose. That is the new front door to your site, and it means the questions you ask about your own site have changed shape too: not "where do I rank for X" but "which of my page-one keywords do AI engines ignore, and why?" Those questions belong in a conversation, not in a dashboard you have to remember to open. So we taught TriRank to speak the language assistants already use: the Model Context Protocol.

What we shipped

The TriRank MCP server is a remote MCP endpoint that exposes five read-only tools to any MCP client. You connect it once; after that you just ask.

You never name a tool. You ask a question; the assistant picks.

Here is what that actually looks like, verbatim from the production server. Ask about a competitor's authority:

Or ask about a site you own:

Those are real numbers from a live account, not a mock-up — including the uncomfortable ones. Note what the second answer refuses to do: it reports what each engine swept, not just what it cited. Perplexity checked 5 queries and cited 3; Google's grounding checked 9 and cited 2. An engine that never ran shows as unswept, never as a zero. That distinction is the whole reason the numbers are worth trusting, and it is why the citation-gaps tool exists at all — trirank_list_citation_gaps returns the keywords where you already rank on Google's first page but an engine explicitly does not cite you, with a per-engine verdict of cited, not cited, unknown, or not swept for every row. It is the gap between ranking and being cited, delivered into your chat window instead of a report you have to go find.

How to connect it

Three steps, about two minutes.

1. Create an API key. In TriRank, go to Settings → API keys and generate one. Keep it somewhere safe; it identifies your account and scopes every answer to the sites you own.

2. Add the server to your client. Point your MCP client at https://mcp.trirankai.com/mcp. On Claude.ai, add it as a custom connector and approve access — that path uses OAuth, so no key gets pasted into a chat window. In clients that support custom headers, such as Claude Code and Cursor, send your key as an x-api-key header instead. Both channels resolve to the same account and serve the same tools. The setup page has the exact snippets for each client.

3. Just ask. "How did my rankings move this week?" "Which page-one keywords have zero AI citations?" "What's the DR of my closest competitor?" No syntax to learn.

One quirk worth knowing: MCP clients cache the tool list when they first connect. If you connected before the citation-gaps tool shipped and don't see it, remove the server and add it again — the fresh handshake picks up the new tools.

What it will and won't tell you

The value of an answer in a chat window is exactly the value of the data behind it, so here is where the data comes from and where it stops.

Rankings and traffic come from your own Google Search Console property. That means they inherit Search Console's lag — typically 2-3 days — and every response says which date the data actually runs through, rather than implying it's live. It also means a site with no Search Console connection has no ranking data, and the tools say so instead of returning an empty list that looks like a zero. This matters most for citation gaps: the gap is a join between rankings and citations, so without Search Console the answer is "unknown," not "no gaps."

AI visibility comes from weekly sweeps across five engines — Google (Gemini grounding), ChatGPT (web search), Perplexity, Claude (web search) and Google AI Overview — run against the queries you've asked TriRank to monitor. The first four are API probes standing in for the consumer products, which is why each one carries that qualifier in its label. Google AI Overview is the exception: it is read straight off the live results page, so it is the consumer surface itself and needs no qualifier. Because AI answers are non-deterministic, a single check proves little; the discipline that makes it meaningful is repeated sweeps of the same prompts over time, the same reason citation tracking across Perplexity and ChatGPT has to be a schedule rather than a spot check. Any query no engine has swept yet is reported as unswept — never quietly folded in as "not cited."

Domain Rating comes live from Ahrefs on each call, for any public domain, connected to TriRank or not.

Everything is read-only, and reading is free — 500 calls per day per key, resetting at 00:00 UTC. Nothing in the MCP server publishes, edits, or deletes, no credits are spent when your assistant queries it, and every site-scoped tool answers only for sites connected to that key's own account. Ask it about a domain you don't own and it will tell you, politely, which sites you can ask about instead.

That's the whole trade: the numbers your dashboard already renders, reachable by asking. If you want to see your own citation gap first, start with a free audit — then connect the server and ask it what changed next week.

FAQ

What is an MCP server? MCP (Model Context Protocol) is an open standard that lets AI assistants call external tools. An MCP server exposes a set of tools; your assistant connects once and can then fetch live data on your behalf. You never call the tools by name — you ask a question and the assistant picks the right one.

Which AI clients can connect to the TriRank MCP server? Any MCP client. Claude.ai and the Claude desktop apps connect over OAuth in two clicks. Clients that support custom headers, such as Claude Code and Cursor, can send an x-api-key header with a key from TriRank settings instead.

Does the MCP server cost anything? Reading your data through the MCP server is free, with a limit of 500 calls per day per API key, resetting at 00:00 UTC. It reads data TriRank has already collected, so no credits are spent when your assistant queries it.

Where do the numbers come from? Rankings and traffic come from your connected Google Search Console property, which typically lags 2-3 days like Search Console itself. AI visibility and citation gaps come from TriRank's weekly sweeps across Google (Gemini grounding), ChatGPT, Perplexity, Claude and Google AI Overview. Domain Rating comes live from Ahrefs.

Can the MCP server change anything in my account? No. All five tools are read-only — they query rankings, reports, AI visibility, citation gaps and Domain Rating. Nothing publishes, edits or deletes, and every site-scoped tool only answers for sites connected to the API key's own account.