Sign inBook a setup callStart tracking
Product
Product overviewHow it worksLive demoFree tools
Product
DashboardsClient reportsAutomationsProfilesAssistant
By what you run
Online storeServices and leadsCreatorsAgencies
Platforms
Meta adsGoogle AdsTikTokInstagramEmailSMS
Resources
DocsGuidesGlossaryWalkthroughsEvents APICompareSecurity
PricingSign in
Start trackingBook a setup call

Metrics explained

Ask your analytics in plain words: MCP and AI assistants

What the Model Context Protocol is, how an AI assistant reads your analytics through it, what it answers well, and the safety rules for keys.

By the Sporetag teamUpdated 5 min read

On this page

To ask Claude, ChatGPT or Cursor about your analytics, connect the assistant to an MCP server: a service, run by your analytics tool or on your computer, that hands the assistant your numbers through a small set of tools such as "get stats for these dates". You then ask "how did last week compare with the week before?" and the assistant answers from the real data.

This guide covers what the Model Context Protocol is, remote versus local servers, the questions assistants answer well and the five ways they go wrong with numbers, and the safety rules for keys and data.

What MCP is

The Model Context Protocol is an open standard for connecting AI applications to other systems (MCP). A tool such as an analytics product runs an MCP server. An assistant such as Claude, ChatGPT or Cursor is the client.

The server publishes a list of tools, each a named action with a description and the inputs it takes, such as get_stats with a site and a date range. When you ask a question, the model picks a tool, fills in the inputs, and the client calls the server. The server answers with data, and the model writes the reply from it (MCP).

The model never sees your database. It sees only what each tool returns.

Remote or local

  • A remote server is a web address the product runs. You paste the address into your assistant, with a key. Nothing to install. In Claude, for example, it is added as a custom connector (Claude).
  • A local server runs on your computer and calls the product's API with your credentials. Google's experimental Google Analytics MCP server is one: it runs locally with Google Cloud credentials and reads GA4 reports (GitHub).

Remote servers suit most people. Local servers suit developers who want to change what the server does.

What it does well

  • Comparisons over a period: "Visitors and conversions last week versus the week before."
  • Top lists: "Top five sources by sales this month."
  • Crossing two reports: "Which campaigns had spend but no conversions?"
  • Drafting words from numbers: "Write three lines for my client about October." You still check the numbers.
  • Follow-ups: "And on phones only?", without rebuilding the report.

Where it goes wrong

An assistant is a fluent writer reading a table. Know its weak spots:

  1. Date ranges. "Last week" can mean the last 7 days or Monday to Sunday. "This month" in which time zone? Good servers document how they read dates, and good questions name them: "Monday October 13 to Sunday October 19."
  2. Arithmetic. Models can add and divide wrongly, especially over long tables. Ask the server for the totals and rates rather than letting the model compute them from rows.
  3. Definitions. A "user" in one tool is a "visitor" in another. The model may use a word loosely; the number still means what the tool counts. See sessions, users and pageviews.
  4. Attribution. "Which ad made the sale?" depends on the model of credit. Ask which one the answer used. See attribution models explained.
  5. Confident gaps. If a tool has no data for a question, a model may answer anyway from general knowledge. Ask it to say which tool call each number came from.

Safety rules

Use a read-only key, scoped down

Connect with a key that can only read, and only the sites the assistant needs. A key for one client's site cannot leak another client's numbers. The MCP specification also asks applications to keep a human able to deny tool calls, and to confirm sensitive operations (MCP); read-only tools take most of that risk away.

Treat the address like a password

Some assistants accept only an address, so the key goes inside it. Anyone with that address can read what the key reads. Do not paste it into shared documents, and revoke the key when you stop using the connector.

Visitor-supplied text is data, not instructions

Analytics is full of text written by strangers: referrer names, page paths, campaign names, search terms. Someone can put "ignore previous instructions" in a referrer and wait for an assistant to read it. That is prompt injection, the first risk on OWASP's list for AI applications (OWASP). A careful server labels such values as untrusted data, and read-only tools mean that even a fooled model cannot change anything.

Personal data stays out

Aggregate numbers are enough for almost every question. A server that never returns visitor emails, names or phone numbers is one you can connect without a new privacy review of the assistant.

Where Sporetag fits

From Sporetag's MCP code and help: a remote server at https://sporetag.com/mcp on paid plans, over Streamable HTTP, taking a read key made in Settings, API for all sites or one. Seven tools, all marked read-only: sites, stats with the span before, breakdowns, goals, attribution, ad spend with ROAS, and short links. Dates are whole local days in the site's time zone. The server tells the model to treat source, page and campaign names as data, never sends visitor emails, names or phone numbers, and shares the read API's limit of 120 requests a minute per key.

Common questions

What is an MCP server?

A service that publishes tools an AI assistant can call, using the open Model Context Protocol. An analytics MCP server lets the assistant read your numbers to answer questions.

Can I connect Google Analytics to Claude or ChatGPT?

Google publishes an experimental local MCP server for GA4 that you run with your own Google Cloud credentials. Other analytics tools offer remote servers you connect with an address and a key.

Is it safe to give an AI assistant access to my analytics?

It is when the key is read-only and limited to the sites needed, the server returns no personal data, and you revoke the key when you are done. Treat any address that contains a key like a password.

Can the assistant get numbers wrong?

Yes. It can misread a date range or compute a rate wrongly. Name exact dates, ask for totals from the tool rather than sums of rows, and check numbers before sending them to anyone.

Sources

Visitors, sessions and pageviews: how each tool counts themWhat a user, a session, a visit and a pageview are in GA4, Shopify and privacy-first tools, why the counts differ, and which one to report.Metrics explained · 4 min readAttribution models explained, with one buyer's path worked throughFirst touch, last click, linear, position-based and data-driven attribution, worked through on one buyer path. What each model is good for, and its limits.Attribution · 5 min readClient reporting for agencies: a template clients actually readWhat to put in a monthly client report, how often to send it, which numbers to lead with for stores and lead-gen clients, and how to explain gaps.Agency reporting · 5 min read

Spotted something wrong or out of date? Email hello@sporetag.com and we will fix it.