ChatGPT for marketing: the copy-paste ceiling, and how to get past it

Most marketing teams using ChatGPT hit the same wall within a month, and it is not a wall about reasoning quality.

The model is entirely capable of looking at a week of campaign data and telling you something useful about it. What it cannot do is see the data. So you export a CSV, or screenshot a table, or paste a block of numbers into the chat — and because that takes four minutes, you only do it when a question is important enough to be worth four minutes. Which means the small questions, the ones that would have caught something early, never get asked.

That is the ceiling. Everything below is about what removing it changes, and what it does not.

What pasting actually costs

Three things, and the first is the least interesting.

The friction. Four minutes per question is enough to stop most questions. Anyone who has watched their own behaviour here knows the effect is not linear: you do not ask fewer questions, you ask a different, smaller set of them.

The provenance. A pasted block of numbers has had its origins stripped off. Which account, which date range, which attribution setting, whether “conversions” meant the platform’s default or the one your team overrode. The model cannot ask, so it assumes, and the answer arrives with the same confidence either way. A week later you cannot check the working, because the working was a clipboard.

The boundary. This is the one that matters most. You can paste from one place at a time, and the marketing questions worth asking usually span two: what the campaign cost against what the customers it produced actually paid. To answer that by pasting you have to export from both platforms, align the dates yourself, decide what counts as a match, and then ask — at which point you have done the hard part and the model is formatting.

What connecting accounts changes

An MCP server is a way of giving a client like ChatGPT authorised access to real accounts. You authorise the accounts once with the server’s operator; the server exposes them to the model as named tools; the model calls a tool when a question needs it. (What an MCP connector is, in full.)

Concretely, the change is this:

	Pasting	Connected accounts
Time per question	Minutes of export and cleanup	None
Date range	Whatever you exported	Whatever you asked for
Provenance	Lost at the clipboard	The account the tool read
Questions spanning two accounts	You do the joining	The server reaches both
Repeating the question next week	The same four minutes	Ask again

The third row is the quiet one. Once the account is the source, a follow-up question — and what did that look like the week before? — costs nothing, which is when this stops being a novelty and starts changing what gets noticed.

What it does not change

Three things, stated plainly because the category tends not to.

Your definitions are still yours. Two platforms reporting on the same campaign will disagree, and they are both right by their own rules: different attribution windows, different lookback periods, different views of what a conversion is and when revenue is recognised. A model that reconciles them without telling you has not solved the problem, it has hidden it. Ask what a figure is measuring when the answer matters.

Dashboards are still useful. A number that four people check every Monday belongs on a dashboard. Connected accounts are for the questions nobody built a view for, which is most of them but not all of them.

Generating copy is a different job. Most “ChatGPT for marketing” advice is about writing — ad variants, subject lines, briefs. Connecting an ad account does nothing for that and this post is not about it.

And the client is still a third party. When a tool returns, the result goes to ChatGPT, under your agreement with OpenAI. No MCP server can make a promise about what happens to it there. Worth reading once, rather than assuming.

Where to start

Start with the question you keep not asking because the export is annoying. If it is about one account, the vendor’s own MCP server is probably the shortest path, and we keep a list of which vendors publish one.

If it crosses two accounts — spend against revenue, traffic against orders, installs against subscriptions — that is the case for a server that carries several. Ours does, across ads, analytics, ecommerce, subscriptions and CRM, and every connector page says whether it works today. Setting it up in ChatGPT is a settings step, not an integration project.