The Ecomm Analyst

ThoughtMetric’s MCP server, which the company markets as AI Connectors, lets you query your attribution data from Claude or ChatGPT in plain language instead of opening a dashboard. It exposes a small set of tools that let the assistant look up which metrics and dimensions exist in your account and then run aggregated queries against them. ThoughtMetric sponsors this blog, so read what follows with that in mind. I have spent enough time with the thing to have opinions about what it is good for and what it is not.

MCP stands for Model Context Protocol, an open standard for connecting AI assistants to external data. You authorize a connection once and the assistant can then pull live numbers from your account rather than asking you to paste a CSV.

What is actually exposed

Five tools, and the shape of them tells you a lot about how to use it. One lists the projects, meaning the stores or brands, your login can reach. One lists every available metric. One lists every available dimension. One returns a compatibility map showing which metrics can be grouped by which dimensions. The last one runs the actual query.

That fourth tool is more important than it sounds. Not every metric can be sliced by every dimension. Ad metrics like spend cannot be grouped by discount code, because spend does not have one. Rather than letting the assistant guess and return a plausible wrong answer, the compatibility map is checked first.

The metrics available

The list is broader than I expected. Traffic covers visits, pageviews, and session duration. Conversion covers orders, add to carts, leads, and conversion rate. Revenue is granular in a way that matters, with gross sales, net sales, discounts, refunds, shipping, and tax broken out separately rather than collapsed into one figure.

The advertising side has spend, clicks, impressions, CPC, CTR, CPM, ROAS, and MER. The customer side is where it gets useful, because almost everything is available split by new versus returning. There are separate metrics for new customer orders, new customer revenue, new customer ROAS, returning customer ROAS, cost per new customer order, and blended CAC. If you have ever exported two reports to work out whether a campaign was buying new customers or harvesting existing ones, that split being native saves real time.

Dimensions include time, channel, the full UTM set, account, campaign, ad set, and ad name, plus product name, SKU, product category, variant, discount code, and geography down to city. Filters support equals, does not equal, greater than, less than, contains, and does not contain.

What it is not

It is a query interface over aggregated data. It is not access to your raw order table. You cannot ask it to list every customer who bought twice in ninety days, because that is a row-level question and the API returns aggregates grouped by dimensions. If your analysis needs individual records, you still need an export.

It is also read-only. Nothing you ask changes your account, adjusts a budget, or writes back to an ad platform. Given the current enthusiasm for agents that take actions, I consider that a feature rather than a limitation.

The more subtle constraint is that natural language hides the schema from you. Ask for something the data model cannot express and you are relying on the assistant to tell you that clearly rather than to quietly answer a slightly different question. In my experience it is reasonably good about this, but it is the failure mode to watch for, and it is a good reason to sanity-check anything surprising against the dashboard before you act on it.

Where it earns its place

Quick checks, mostly. Someone asks in a meeting what last week’s ROAS was by channel and you have an answer before you would have finished loading the dashboard. Ad hoc slices you would not bother building a report for, like ROAS by product category for one campaign over a specific fortnight. Comparisons across date ranges that are tedious to set up manually.

ThoughtMetric’s own positioning is that this is a faster interface to the same data rather than a replacement for dashboards or deep analysis, and I think that is accurate. It does not do the analysis for you. It removes the friction between having a question and seeing a number.

The connector comes with the subscription rather than as an add-on, which follows from the pricing model. ThoughtMetric charges on monthly pageviews starting at $99 per month for 50,000 pageviews and includes every feature at every tier, with a two-week free trial and no credit card required.

Whether it is worth setting up

If you are already a ThoughtMetric customer, it takes a few minutes and there is no reason not to. If you are evaluating attribution tools, I would not choose one on the strength of its MCP server. The quality of the underlying attribution matters far more than the interface you use to query it, and a fast natural-language front end on numbers you do not trust is not an improvement.

The general case for connecting your marketing tools to an assistant, and where I think it stops being useful, is something I wrote about in the case for plugging your marketing tools into AI.