Ask Claude About Your Funnel: A Practical Guide to Analytics over MCP

You are in your editor, or in a chat with Claude, and a question comes up. Did the new pricing page change the signup rate? Which campaign sent the visitors who left at checkout?

The usual way to answer is to stop what you are doing, open your analytics tool, build the report, and bring the number back. An analytics MCP removes those steps. You ask the question where you already are, and the AI tool gets the answer from your data.

This guide explains how that works, how to set it up with GrainQL, and how to ask questions that give answers you can trust.

What MCP is, in two paragraphs#

MCP means Model Context Protocol. It is an open standard that lets an AI tool use external services. A service publishes an MCP server, which is a list of functions with a description of each one. The AI tool reads the list and calls a function when your question needs it.

An analytics MCP server publishes functions such as "list the tracked events", "count this event by day", and "compare two periods". When you ask "how did signups do this week?", the AI tool chooses the functions, calls them, and writes the answer from the result. The numbers come from your analytics, not from the model's memory.

What the GrainQL MCP can do#

The GrainQL MCP gives an AI tool these abilities:

• See what you track. List your events and the properties on each one.
• Query. Count events, group them by country, device, campaign, or page, and return time series.
• Compare periods. This week against last week, with the changes that stand out already calculated.
• Get a digest. A summary of what changed recently across your site.
• Compare segments. Converters against non-converters, mobile against desktop, one campaign against another.
• Cluster sessions. Group sessions that behave the same way, for example the ones that leave at one checkout step.
• Correlate. Find what else moved when one metric moved.

Since GrainQL 6.5, one connection can cover every site on your account. If you run several shops or manage sites for clients, you can ask about all of them in one question.

The connection can read and analyze. It cannot change your tracking, your settings, or your data.

Connect it#

The setup takes about two minutes.

In Claude. Open the connector settings, add a custom connector, and enter the server address:

https://www.grainql.com/api/mcp

In Claude Code. Run this command in your terminal:

claude mcp add --transport http grainql https://www.grainql.com/api/mcp

In Cursor and other tools. Add a remote MCP server with the same address in the tool's MCP settings.

In each case a browser window opens and asks you to sign in to GrainQL. You then choose how far the connection reaches: your whole account, or one workspace. After that the connection is ready.

Every connection appears in your GrainQL settings with the time it was last used. If you revoke one, it stops working immediately.

Questions that work well#

The quality of the answer depends on the question. These five patterns give reliable results.

Start with what exists#

What events does my site track, and which ones look like conversion steps?

Ask this first in a new conversation. The AI tool learns your event names, and you avoid answers built on an event that does not exist.

Give a period and a comparison#

Compare signups in the last 7 days with the 7 days before. Break it down by traffic source.

A question with dates gives an answer you can check. "How are signups?" gives an opinion.

Ask where, then ask why#

Which step of the checkout funnel lost the most visitors this week compared with last week?

Then follow up:

What is different about the sessions that left at that step?

The first question finds the place. The second uses segment comparison and session clustering to find what those visitors have in common, such as a device, a browser, or a campaign.

Connect it to your code#

This is where an MCP inside an editor is most useful. You can ask about the change you are working on:

We released the new address form on Tuesday. Did the completion rate of the shipping step change after that, on mobile and on desktop?

The AI tool has your code open and your data connected. It can relate a number to the release that caused it.

Ask across sites#

Across all my workspaces, which site had the largest drop in conversion rate this week?

With an account-wide connection, this is one question. Before, it was one report for each site.

Where the limits are#

An AI tool with your data is still an AI tool. Keep these four points in mind.

It can choose the wrong event. If you have signup_started and signup_completed, a general question about "signups" can use either one. Name the event when it matters.

Ask it to show the query. A good habit is to end with "tell me which events and date range you used." You can check an answer only when you can see how it was made.

Small numbers are still small. The AI will report a 50% drop from 4 conversions to 2 with full confidence. Look at the counts, not only the percentages.

It answers what you ask. It will not tell you about a problem you did not ask about. For that, GrainQL has Kai inside the product, which watches your data and reports changes before you ask.

When to use MCP and when to use the dashboard#

Use the MCP when the question comes up during other work: in your editor, in a planning document, in a chat where you are writing a report. It is fast and it keeps you in one place.

Use the GrainQL dashboard when you need to see behavior. Session replays, heatmaps, and funnel views show you what a number cannot. A typical flow is to find the problem through the MCP and then open the sessions in GrainQL to watch it happen.

Both use the same data, so the numbers agree.