ChatGPT for marketing: what it is good at, and where it stops

• First drafts and variations. Twenty headline angles, five email subject lines, a landing page outline.
• Reading an export you give it. Paste a search terms report and ask which terms spend without converting. It is fast and usually right.
• Research. What competitors claim, what customers ask, how a market talks about a problem.
• Explaining. Turning an audit into a note a client will actually read.

• Numbers it was not given. Ask about your CPA without data and you get an industry average dressed as advice.
• What a metric means in your account. It will read “conversions” as sales unless you tell it which actions are counted. Often they are page views or short calls.
• Recent days. It sees a drop in the last three days and suggests cutting bids. Conversions keep arriving for weeks after a click, so recent days always look worse.
• Adding up platforms. Meta, Google Ads and Analytics each claim the same sale. Summed, you have sold things twice.

Use a Project for each client or brand, and put these in the project’s files or instructions:

• What the business sells, to whom, and what a customer is worth.
• What each campaign is for and the one number that judges it.
• Which conversion actions count, and which are noise.
• Decisions already made, so it stops reopening them.
• Your brand voice, from real customer reviews, not adjectives.

This one instruction improves almost every analysis:

The “could not see” part matters most. It stops a partial look from reading as a full one.

ChatGPT can connect to outside tools through MCP servers in developer mode. OpenAI describes that as full read and write access, meant for people who know how to set it up safely. Google publishes an official Google Ads server that reads, and Meta publishes one that can also make changes.

Connecting is the easy half. The hard half is what stops a confident wrong answer becoming a live change. How to connect safely.

Even connected, a chat does not remember last month, does not check whether its advice worked, and does not look at your accounts while nobody is asking. The problems that cost us most were found that way: tracking that could not name the campaign behind 107 paid visits, and a test that had one impression in a week while everyone thought it was running.