Claude, Coupler.io and Markifact: our AI marketing stack

Our AI marketing stack is a chain. Coupler.io first pulls in the data from more than 400 sources, Markifact opens up Google Ads, Meta Ads, GA4 and Tag Manager, among others, and via MCP everything reaches Claude. Claude turns it into an HTML page with the analysis and slides, which we publish on Cloudflare or share in the meantime as a Claude artifact.

This is the setup we use ourselves as a Google Analytics agency. The answers are only as good as the measurement underneath. Below you will read what each part does, what you can ask, and where to be careful.

Four links, each with its own role

Coupler.io and Markifact connect to Claude via MCP. The Model Context Protocol is an open standard for setting up secure two-way connections between data sources and AI tools ^[1]. So Claude does not talk directly to your ad accounts, but to an intermediate layer whose permissions you decide.

What you can ask

Via Coupler.io, Claude can search your datasets, view their columns, query them and combine them with SQL ^[7]. The computing happens on Coupler.io's side, so large datasets don't have to fit in Claude's memory ^[3]. A few examples:

• "Which channels and campaigns delivered the most conversions last month?" That is Coupler.io's own example ^[8].
• "Put ad spend per channel next to the deals won from the CRM."
• "Which campaigns cost more per enquiry than last quarter?"

If a source is missing, Claude can set up a new data flow itself: connect a source and choose a destination ^[7].

Via Markifact, Claude pulls reports from Google Ads, Meta Ads and GA4, and creates or manages campaigns ^[5]. It can also work in Tag Manager, up to and including publishing a container version ^[6]. Claude first describes such a write action, and asks for confirmation ^[5].

What the page with analysis and slides looks like, which metrics belong in it and how to protect it is covered in HTML dashboards with Claude.

Where it stops

The stack is fast, but not all-knowing. Five limits you need to know:

• Claude only sees what you share. Via Coupler.io, only the datasets from flows that have Claude as their destination ^[7].
• Data is as fresh as the last refresh. That ranges from every 15 minutes to monthly, depending on your plan ^[9].
• Confirmation is an instruction, not a lock. Markifact asks the AI tool to confirm, but whether you really have to approve depends on that tool's settings ^[5], ^[10]. Markifact itself advises against setting the write tool to "Always allow" ^[5].
• Actions cost credits, usually one per operation; searching and looking up is free ^[5].
• Bad measurement gives bad answers. If your tracking is wrong, Claude analyses wrong figures. So start by getting your tracking in order.

Also always have Claude show how it calculated, and check one figure against the platform itself. Why GA4 and Google Ads differ in that respect is explained by Cédric in why GA4 and Google Ads show different numbers.

Privacy and access rights

An AI that can write in your ad accounts needs the same care as a new employee. You add a connector in Claude with the URL of the MCP server; on Team and Enterprise, an owner first does that for the whole organisation ^[11]. With Markifact, you then grant access in a login screen from Markifact itself ^[5].

Anthropic recommends only connecting servers from organisations you trust, and limiting the permissions requested. Choose "Allow always" only for tools you dare to let work unsupervised ^[11].

What the tools themselves promise:

• Coupler.io has SOC 2 Type II and says it is GDPR compliant. It only asks your sources for the permissions needed to pull data, and never modifies records there ^[12].
• Markifact lets you restrict each MCP server to certain operations and certain accounts. A request can narrow those permissions, never expand them ^[5]. Stored data is kept in the EU ^[13].
• Anthropic only uses chats on Free, Pro and Max for training if you allow it ^[14]. On business products, such as Team, Enterprise and the API, not by default ^[15].
• Claude artifacts are visible only to you at the start ^[16]. So an interim analysis is only shared when you do so yourself.

Three rules of thumb: client data only in a business account, a separate MCP server per client with only that client's accounts, and no names or email addresses in the chat. Who may have which access to an ad account, and who owns the account, is explained by Benjamin in transferring a Google Ads account.

Want to know first whether your measurement is right before you let AI loose on it? The free tracking audit lays it all out.