You have probably done it. Exported a campaign report to CSV, dropped it into ChatGPT and asked what on earth is happening with your CPAs. The answer came back fast, confident and beautifully written. It was also frozen in time, blind to nine of your ten platforms, and built on whatever slice of data survived the export.
That workflow has a name now: copy-paste analytics. And it finally has an expiry date, thanks to a quiet little standard called MCP. This post explains what MCP is in plain language, what building one actually involves, and how we turned the idea into Upbuild, a connector that lets marketers ask their entire ad stack anything from inside Claude or ChatGPT.
AI is brilliant. It is also blind.
Large language models know an astonishing amount about marketing in general and precisely nothing about your accounts. They have never seen your Google Ads search terms, your GA4 conversions or your Meta frequency curves. The gap between a smart assistant and a useful one is not intelligence. It is access.
Until recently, the only way to close that gap was manual: screenshots, exports, pasted tables. Every one of those is stale the moment it leaves the platform, capped by upload limits, and one fat-fingered filter away from being wrong. If your analyst spends their morning ferrying numbers between dashboards and chat windows, they are not analyzing. They are working as a human API.
What is MCP, in plain words?
MCP stands for Model Context Protocol. It is an open standard, introduced by Anthropic in late 2024 and since adopted across the industry by OpenAI, Microsoft and Google. Think of it as a universal socket: one standard way for an AI assistant to plug into an external system and use it.
The mechanics fit in one paragraph. An MCP server sits in front of a data source and publishes a menu of tools: small, named actions like “run a GA4 report” or “list the tags in this GTM container”. The assistant reads the menu, decides which tool answers your question, calls it, and works with the live result. No export in the middle, no copy of your data sitting anywhere it should not.
For marketers, the shift is simple to describe. Without MCP, your assistant is a well-read intern. With MCP, it is a well-read intern with a login.
What building an MCP actually takes
You do not need to build one to benefit from MCP, but knowing what goes into it makes you a sharper buyer. Conceptually, there are five steps:
- Pick the API you want to expose. Every major ad and analytics platform already has one; the MCP server is a translator, not a replacement.
- Wrap it as a short menu of tools. A handful of well-named, well-described tools beats an exhaustive one-to-one mapping of every endpoint. The AI has to choose between them, so clarity wins over coverage.
- Solve authentication properly. Real users need OAuth sign-in, token refresh and scopes, per user, not one shared key for everybody.
- Decide what the AI is allowed to do. For marketing data the answer should almost always be read-only. An assistant that can look at your campaigns is an analyst; one that can edit them is a liability.
- Host it and hand out a URL. The finished server is just an address your assistant connects to.
Here is the practitioner truth the tutorials skip: the tool code is the easy 80 percent. The hard 20 percent is auth, scoping and making sure one user can never see another user’s accounts. If you fancy a weekend project to feel the shape of it, our step-by-step guide to building a conversational GA4 assistant with the Gemini CLI is exactly that, in miniature.
Why we built Upbuild
As a martech consultancy, we live inside our clients’ ad accounts. And for years we watched the same scene on repeat: someone senior asks a perfectly reasonable question, “which channel actually drove revenue last month?”, and the answer requires opening four platforms, exporting three reports and rebuilding a spreadsheet that existed last quarter under a different name.
So we built the connector once, properly. Upbuild plugs ten platforms into Claude, ChatGPT, Copilot Studio and Gemini Enterprise: Google Ads, Google Analytics, Google Tag Manager, Campaign Manager 360, Display & Video 360, Meta, TikTok, Snapchat, Adobe Analytics and Adobe Launch. Fifty-five tools, every single one read-only, answering with live data from your own accounts.
The design choices are the lessons from the section above, applied without compromise. Read-only by design, so the AI can see your data but never touch a campaign. Per-user sign-in, so you only ever see the accounts your own logins already have access to; there is no shared service account pretending to be everyone. And zero installation: you paste one connector URL into your assistant, click Sign in with Google, and you are asking questions about two minutes later.
What people actually ask it
The magic is not in any single feature. It is in the questions that suddenly cost nothing to ask:
- “Compare ROAS across Google, Meta and TikTok for the last 30 days, and tell me where the budget should move.”
- “Which Google Ads search terms spent money with zero conversions last week?”
- “Audit this GTM container before the migration: list every tag, trigger and variable, and flag anything that looks abandoned.”
- “It is launch day. Show me GA4 realtime by source every time I ask, and shout if the purchase event stops firing.”
- “Pull the Floodlight lookback windows on this advertiser and check they match what we documented.”
None of these need SQL, a dashboard build or an export. They need an assistant that can reach the data, and a marketer curious enough to ask the next question. In our experience, the second follow-up question is where the real insight lives, and follow-ups are exactly what dashboards are worst at.
Build your own, or plug one in?
Honest answer: it depends on what you are connecting. If the data lives in proprietary internal systems, a custom MCP built around your warehouse or your CRM is a genuinely great investment, and the five steps above are your roadmap. If the data lives in the standard ad and analytics platforms every marketer uses, that server has already been built, hardened and hosted for you.
Either way, the direction of travel is clear. The era of pasting CSVs into a chatbot is ending, and the marketers who get comfortable asking their data direct questions will simply move faster than the ones still exporting.
Curious what your own accounts would say if you could just ask? Upbuild has a 14-day free trial, no card required, live in about two minutes. And if what you need is a custom MCP built around your own data stack, that is literally a service we run.
