How to let Claude read your website analytics

An MCP server lets Claude pull real figures from your analytics instead of guessing from a screenshot. How the connection works, how to add one, what to ask first and where the answers usually slip.

7 min read

Summarize with

In this article
  1. 1What an MCP server is
  2. 2Local or remote: check this first
  3. 3Does your analytics tool have one?
  4. 4Adding it to Claude
  5. 5First, ask something you can check
  6. 6Questions that work, and ones that won't
  7. 7Where the answer can go wrong
  8. 8What the AI tool gets to see
  9. 9A short checklist

Most analytics questions are small. Which pages did well last week, where Tuesday's visitors came from, whether mobile is growing. If you already spend the day in Claude, asking there can be quicker than opening a dashboard and setting a date range.

An MCP server is the piece that lets Claude read your real numbers instead of guessing. Here's how the connection works, how to add one, what to ask first, and where the answers tend to go wrong. It applies to any analytics tool that offers an MCP server.

What an MCP server is

MCP stands for Model Context Protocol. Its documentation describes it as an open-source standard for connecting AI applications to outside systems. As of November 2026, Claude and ChatGPT support it, and so do code editors such as Visual Studio Code and Cursor.

There are three parts:

  • The server. Your analytics tool publishes an MCP server that offers a set of tools, such as "top pages for a date range".
  • The client. An AI app like Claude connects to that server and learns which tools it has.
  • The call. When a question needs data, the model calls one of those tools, gets figures back and writes its answer from them.

For analytics, the benefit is that the AI reads numbers from the source, not from a screenshot or a CSV that was out of date the moment you exported it. The limit is just as simple: the AI can only see what the tools return. If the server has no tool for entry pages, Claude can't see entry pages, even if your dashboard shows them.

Local or remote: check this first

MCP servers come in two kinds, and the protocol's own docs call them local and remote. Which one your provider offers decides how you set it up.

Local serverRemote server
Where it runsOn your computer, started by a commandOn the provider's side, reached over the internet
How you add it to ClaudeAn entry in Claude Desktop's config file, then a restartA custom connector: a name and a URL
How it signs inA key or credentials you set up on that computerOften a sign-in page from the provider, where you approve access; some use a fixed key
What to watchNeeds the right runtime installed, and only works on that computerHas to be reachable from the public internet

That last row matters if you host your analytics yourself. Claude connects to remote servers from Anthropic's cloud, not from your computer, so a server on a private network or behind a VPN won't connect unless you allow Anthropic's IP addresses.

Does your analytics tool have one?

Search your analytics provider's docs for "MCP". You'll find one of three things.

  • An official server, built and maintained by the vendor. Google, for example, documents an MCP server for Google Analytics that you run on your own computer. Setting it up takes Python 3.10 or later, a Google Cloud project with the Analytics Admin and Data APIs turned on, and credentials (your own Google account or a service account) with at least the Viewer role on the property. As of November 2026, Google's docs say it handles read requests only.
  • A community server, built by someone else and often run with an API key from your analytics account. These can work well, but the key goes into code from someone other than your provider. Check who maintains it, read what it does with the key, and use the most limited key your provider offers.
  • Nothing. An API or a CSV export still works, but then you're back to pasting snapshots into the chat.

Prefer an official server when there is one. Your credentials then go to code from the company that already holds your data, not to a third party.

Adding it to Claude

For a remote server, as of November 2026, Anthropic's help page on custom connectors gives these steps:

  1. Get the server URL from your analytics provider's docs.
  2. In Claude, go to Customize, then Connectors, choose Add and then Add custom connector.
  3. Give it a name, paste the URL and continue.
  4. Choose how it signs in, following your provider's docs. If the provider uses OAuth, that means a sign-in page where you approve access.
  5. Turn the connector on in a conversation, from the + button at the bottom left of the message box, and ask a test question (more below).

Anthropic lists custom connectors on its Free, Pro, Max, Ultra, Team and Enterprise plans, with Free limited to one custom connector. On Team and Enterprise, an owner adds the connector in the organisation settings first, and members then connect to it.

For a local server, Claude Desktop reads a file called claude_desktop_config.json. The MCP guide for local servers has you open it from Claude Desktop's settings, under Developer, with Edit Config. Add the entry your provider gives you, then quit and restart Claude Desktop. Menu names change, so follow the current help pages rather than an old screenshot.

Other MCP clients, such as ChatGPT, Visual Studio Code and Cursor, have their own setup steps. The MCP introduction linked above points to each one's docs.

First, ask something you can check

Start with a question whose answer you can already see in your dashboard.

Then open the dashboard for the same day and compare. If the numbers differ, check:

  • Time zones. "Yesterday" for the AI tool may not be yesterday in the time zone your dashboard uses.
  • Date wording. "Last week" can mean the last seven days or the previous calendar week. Give explicit dates.
  • Metric names. Pageviews, visits and unique visitors are three different numbers. Name the one you want.

It also helps to look at the list of tools the connector offers, because that list is everything it can see. For Claude Desktop, the MCP guide for local servers finds it through the button at the bottom left of the message box: Connectors, then Manage connectors, then the server's name.

Questions that work, and ones that won't

Questions with one clear metric and one clear date range work best:

  • Which ten pages had the most pageviews between 1 and 31 October?
  • What were the top entry pages this month, compared with last month?
  • Which traffic sources sent the most visitors last week?
  • Did any AI assistants send visitors this month, and which ones? (Background: how AI referrals show up.)
  • How did the split between mobile and desktop visits change over three months?
  • Which UTM campaigns brought visits in March? (Clean tags make this much easier: UTM naming conventions.)

Three kinds of question won't get a real answer:

  • Data your tool doesn't collect. If you never set up funnels or custom events, there's nothing to read.
  • Data the server doesn't expose. A server can offer fewer reports than the dashboard.
  • Why people did something. Analytics records what happened, not the reason.

A careful AI tool will say the data isn't there. If it produces a number anyway, ask which tool call that number came from. If it can't say, treat it as a guess.

Where the answer can go wrong

The server returns figures. The AI still has to read them, and that step can slip.

  • Unique visitors don't add up. Summing daily unique visitors counts a returning person once for every day they came. In cookie-free tools that count from a key that changes every day, the same person on two days is two visitors by design.
  • Noise read as a trend. A jump from 12 visits to 20 is a 67% rise and still only eight visits. Ask for the raw numbers next to any percentage.
  • Similar metrics mixed up. Visits and unique visitors, top pages and entry pages. Ask the tool to say which metric it used.
  • Advice that isn't in the data. If the AI suggests what to do next, that comes from the model, not from your analytics. Weigh it accordingly.

If you're comparing the answer with a second analytics tool, the gap may not be the AI's fault at all. No two tools count alike, and the reasons are mostly predictable.

Ask for the figures and the date range first, and the summary second.

What the AI tool gets to see

  • Your AI tool's policies apply. Whatever the server returns becomes part of the conversation. Check how your AI tool handles that data, especially on a work account.
  • Read-only is enough. An analytics connector has no reason to change settings. If a server offers write tools, turn them off in Claude's Search and tools menu. On Team and Enterprise, an owner can set them to Blocked for everyone.
  • Shared keys mean shared access. On Claude's Team and Enterprise plans, a connector set up with a fixed key gives everyone who uses it the same access, so Anthropic advises a key limited to what members should see.
  • Watch for hidden instructions. Anthropic warns that a malicious MCP server can include hidden instructions that try to make Claude do things you didn't ask for. Connect only servers from providers you trust. Remember too that analytics data holds text that strangers write, such as the UTM tags on a link anyone can make.

A short checklist

  • Find out whether your provider's server is official, community-built or missing.
  • Check whether it's local or remote, and set it up the matching way.
  • Prefer read-only access and the most limited key available.
  • Test with a question you can check in the dashboard.
  • Ask with explicit dates and named metrics.
  • Ask the tool to show the figures it used before it summarises.
  • Don't expect answers about data your tool doesn't collect.

Set up this way, an AI tool won't replace your dashboard, but it's a quick way to ask the small questions without leaving the conversation you're already in.


Lomen has an MCP server that lets AI tools such as Claude query a site's figures, and its own chat, Lomen AI, which answers from baselines, weekday patterns and outlier days that Lomen computes first, with Google's Gemini model writing the reply. Lomen doesn't track custom events or funnels, so neither route can answer questions about those. It's free during the public beta, which includes three AI insights a month.

Numbers you can ask questions

Lomen is one cookie-free script: sources, campaigns and AI referrals on one dashboard, and an AI that answers from your own figures. Free during the beta, no credit card.