What MCP actually changes for a marketing team (and what it doesn't)
Today an assistant mostly reads your stack, and that alone changes how fast you can answer a question. Call it the read/write line.
MCP lets an AI assistant use your marketing tools directly, so you stop exporting and pasting between them. The part the hype skips: most marketing connectors today can only read your data, not act on it. So the real win right now is faster answers across your whole stack, not an AI running your campaigns. Here's what MCP changes for a marketing team today, what's still hype, and when it isn't worth wiring at all.
- 01What MCP changesThis piece
- 02Which connectors to wire firstRead part 2
- 03Wiring write access safelyRead part 3
You've seen the acronym in a dozen LinkedIn posts. Most of them imply the same thing: connect MCP and an AI will run your marketing. That's not what's happening.
The honest version is smaller and more useful. Today an assistant mostly reads your stack, and that alone changes how fast you can answer a question. Call it the read/write line. Knowing which side of it you're on saves you from buying a fantasy this quarter.
What is MCP, in a marketer's terms?
MCP, the Model Context Protocol, is a shared standard that lets an AI assistant use your tools directly, the way one USB-C port runs many devices. Instead of exporting a report and pasting it into ChatGPT, the assistant queries HubSpot, GA4, or Ahrefs itself. Anthropic built it in 2024. Anthropic, OpenAI, and Google now back it.
This stopped being one vendor's idea. On 9 December 2025 Anthropic donated MCP to the Agentic AI Foundation, a new fund under the Linux Foundation co-founded with Block and OpenAI, with support from Google, Microsoft, Amazon Web Services, Cloudflare, and Bloomberg. Those eight companies sit behind the standard now. That matters for a buyer: you're not betting on a single supplier's roadmap.
The adoption is real, not marketing. There are more than 10,000 active public MCP servers and tens of millions of monthly downloads of the developer kits. You don't need to know how any of it works under the hood. You need to know it's stable enough to build on.
How is this different from Zapier or an API you already use?
The difference is who decides. A Zapier automation or an API call follows a path you wired in advance. With MCP, you hand the assistant a set of tools and it works out which to use, and in what order, in the moment. Zapier still earns its place as the catch-all for the thousands of apps you'll never wire directly.
Think of it as the gap between directions and a map. An automation is turn-by-turn directions: it does exactly what you told it, every time, and breaks the moment the route changes. MCP hands the assistant the map and the destination, and lets it pick the route. Ask a question it's never seen, and it still works out which tools to touch.
That flexibility has a limit worth naming. Leah Miranda of Zapier describes MCP as acting through your tools while you're chatting, rather than running in the background the way an automation does. So the two aren't rivals. Zapier's own MCP offering reaches across its roughly 8,000-app catalogue, which keeps it the right tool for the set-and-forget flows that should never need a human in the loop.
MCP is for the questions you didn't pre-plan.
What can an AI actually do across your stack today?
Today the reliable win is reading. You ask one question and the assistant pulls from several tools at once: which posts lost traffic this quarter, what they rank for, and how many leads they drove. That used to be three browser tabs and a spreadsheet. Now it's one question and one answer.
That single read is the part people underrate. The value isn't a clever automation. It's the collapse of the gather step, the 20 minutes you spend pulling the same numbers from three places before you can even start thinking.
The point isn't the specific question. It's the shape of it. Any "show me X across these tools" question that used to mean a manual gather now collapses into one ask.
Why can't it just run your campaigns for you yet?
Because most marketing connectors can only read, not act. Google's official Google Ads server is read-only in its current release. It can query the account, but it can't change a bid or pause a campaign. As one practitioner put it, the server can query Google Ads but not manage it, and those are different things. Acting on your tools needs other servers, and other risks.
This is the read/write line, and it's where most explainers quietly mislead you. The official servers from the platforms a marketer actually lives in are read-only today. Google Ads, GA4, and Ahrefs all ship official MCP servers that report and diagnose but make no changes. Google calls its Ads server an initial, read-only release, so this is where things stand now rather than a permanent ceiling. Either way, the assistant you connect this quarter reads.
Writing is possible, but it's a different decision. To let an AI change a bid, edit a record or trigger a workflow, you reach for third-party servers built for write actions, and those carry their own safety questions: what can it change, who approved the change, and what happens when it gets one wrong. That's a real choice with real exposure, not a checkbox. We'll cover how to make it safely in a later piece.
So the honest answer to "can AI run my campaigns" is: not through the official connectors you'd reach for first, and not without a deliberate decision to hand it write access. Reading is the win that's sitting there today.
Do you need this if you already use ChatGPT or Claude?
Yes, and it works in both. MCP started at Anthropic, but it's now a cross-vendor standard. ChatGPT supports it through what it renamed apps in December 2025, and Google's Gemini has backed it since May 2025. Wiring your tools this way doesn't lock you to one assistant. Switch, and the connections come with you.
The specifics differ by tool. ChatGPT renamed its connectors to apps on 17 December 2025; they read your data by default, and write actions sit behind Developer Mode, with full write capability gated to Business plans and up. Gemini has supported MCP since May 2025, and Google shipped fully-managed remote servers for its own services in December 2025. The practical upshot for a marketing lead: if your team switches from ChatGPT to Claude next year, the tools you connected this year keep working. You wired the stack once, not per assistant.
When is MCP not worth it for a marketing team?
When your workflow is small and fixed, when you're a one-tool team, or when your data is a mess. MCP gives an AI access to your data. It doesn't clean it, reconcile your campaign naming, or build attribution that was never there. Connect a bad dataset and you've just made it faster to get confident, wrong answers.
That last failure is the expensive one. As the analytics platform Improvado points out, MCP gives an AI access to your data, but it doesn't clean it, transform it, or govern it. An assistant reading across messy data doesn't flag the mess. It answers smoothly, in fluent prose, with numbers that look right and aren't. Garbage in, garbage out, now conversational. The cleaner your data, the more MCP is worth; the messier it is, the more it'll mislead you at speed.
The other two cases are simpler. If your real workflow is one fixed sequence you run every Monday, that's an automation, and Zapier or a scheduled job does it more cheaply and more reliably. And if you live in a single tool, there's nothing to read across; the gain is marginal. MCP pays off when you have several tools worth querying together and questions you can't predict in advance. That's also why the smart first move is to wire fewer connectors well, not all of them at once, which is the subject of the next piece.
Where to start
The move this quarter isn't "let AI run marketing." It's "stop exporting and pasting." Wire the two or three tools you check most, ask one cross-tool question you'd normally build a spreadsheet for, and see how much of the gather step disappears. That's the win that's real today.
Book a call