Your AI visibility produces data every week: citation rates on the questions your market asks, share of voice against the names you are compared with, the tone of what gets said, the sources each answer stood on. Reading that data is a job. Someone has to compare engines, pull the answers behind a number that moved, cross a weak rate with the sources that engine actually listens to, and turn the whole reading into this week's actions.
That someone can be Claude. Connect it to Epovest once and the assistant you already talk to becomes a GEO analyst: an analyst for generative engine optimization that works on your own measurement rather than on screenshots of it. You ask in plain language; Claude reads the series, quotes the engines' own answers as evidence, and poses the follow-up work in the same conversation.
Claude plays two roles in this story: it is one of the engines your trackers measure, and it is the analyst you talk to. This guide is about the second role: the connection, the questions of a first session, and the weekly loop where an AI visibility analyst earns its place.
What an AI visibility analyst reads
An analyst is only as good as what it can read. Connected, Claude reads:
- The series. Tracking measures how often each engine cites your names on fixed question panels: citation rate per keyword and per engine, survey after survey, share of voice across the names of your market, and tone counts where the sentiment analyst runs.
- The answers themselves. Every survey archives the engines' actual text with the sources each answer cited: the evidence behind every number.
- The map. Atlas ranks the sources of your answers by AI Authority: what each source weighs for each engine, measured on your own questions over the last thirty days.
- The registries. Your surfaces, the pages you control with their alignment state, and your corroborations, the third-party proof about you, with its watch.
- The working file. The quest file, where open work waits, and the Logbook, where actions get dated.
And it reads engine by engine, never as an average: each engine reads a different web, so an average erases exactly what you can act on.
Connect Claude to your data
The connection is a standard MCP server, and there are two ways in.
On claude.ai, add Epovest as a custom connector with the server URL https://mcp.epovest.com/mcp. The connector finds the sign-in on its own: you sign in to Epovest, approve the access once, and there is no key to paste. You review or revoke the connection any time in the app.
In Claude Desktop, Claude Code, or any client that takes a header, use the same URL with an API key created in the app, sent as Authorization: Bearer. Both paths reach the same tools and act for the same account.
The walkthrough per client is on Connect your AI. The analyst acts within the scopes you granted: it reads the measurement, it writes to the registries you point it at, and anything involving money is approved by you, in the app.
The first session: six questions
Open with the questions that produce a decision, not a chart.
"Where am I cited this week, engine by engine?" The trend per engine and per keyword, read from the series: which engine moved, on which question family, since which survey. It comes first because everything else hangs on it.
"Show me the answers behind that negative count." On a tracker that runs the sentiment analyst, Claude brings back the very answers judged negative, each with the sources it cited. You read what was actually said before deciding whether the fix is a page of yours or a mention to obtain from a third party.
"Which sources does my weakest engine listen to, and where am I absent?" Claude sorts your Atlas by that engine's measured authority and crosses it with your corroborations registry. The deliverable is a short list of named sources where a mention would carry, which is how share of voice is won: one source at a time.
"What changed since the last survey, and why?" The comparison names the movements, and the archived answers say why: a source that appeared, a competitor that entered, a question family whose answer tilted.
"What is left to do?" The quest file answers with the work the measurement derives on its own: pages that no longer carry your current wording, proof worth asking a third party to refresh, discovered domains waiting for a decision, plus everything you posed yourself.
"Log what we shipped." Claude dates your actions in the Logbook: the page rewritten, the listing obtained, the campaign started. Next week's reading then has causes standing in front of the curves.
The weekly loop
An analyst pays in cadence. The session is short and always runs in the same order: trend first, evidence second, sorting third, actions last.
Read the trend engine by engine. Drill into anything that moved, with the answers as evidence. Sort what accumulated: discovered domains to accept or dismiss, candidate mentions to record. Pose the week's quests, close what shipped, and log it. On a quiet week the whole loop is two questions long: did anything move, and is anything left. Through a launch, tighten the cadence of the measurement itself and keep the loop daily, because a young record moves fast.
Nothing in the loop depends on memory or goodwill: the quest file holds what is open, the Logbook holds what was done, and the series holds what it changed. The analyst reads all three, so week twelve starts exactly where week eleven ended.
What the conversation adds
A conversation follows up. "And on that engine only?", "Quote the answer.", "Which source carried it?": each drill-down is one sentence, answered from the same data in the same breath. Evidence arrives as the engines' own words. Actions get posed in the same gesture, in the registries, dated. The app remains the place where you see the curves whole and where you approve anything that spends; the conversation is where you interrogate them.
Under the hood
The analyst runs on the same machinery the app reads. The MCP server exposes the measurement to the assistant you connect: the series, the archived answers, the source map, the registries, the quest file, the Logbook. Every question in this guide resolves to those tools, on your account's data, with nothing passing through an export. The technical reference lives at epovest.com/docs/api.md, and the per-client walkthrough on Connect your AI. The first session is one connection away.