How to use ChatGPT for marketing with real data

Most guides about ChatGPT for digital marketing are a list of prompts for writing copy. That works, and it is the smaller part of what the tool solves. The bigger part — analyzing what is happening in your campaigns — is what almost everyone leaves out, because it requires one extra piece.

This article covers both, and is direct about where each one stops.

Three things it does well with zero setup

Producing text variations. Headlines inside character limits, descriptions, scripts, client replies. Repetitive work with simple criteria, done in minutes.

Summarizing what you hand it. Paste a report, ask for five lines in client language. It is the task that saves the most agency time and the one that shows up least in prompt lists.

Explaining concepts and translating jargon. Turning "frequency rose to 5.2" into a sentence the client understands without feeling stupid.

In all three, the data needed is inside what you pasted. That is why they work with no integration.

Where it fails, and why

Ask "my CPL is US$ 38, is that good?" and the answer will correctly explain that it depends on ticket, margin and sales cycle, and cite generic ranges. It is correct. It is useless for deciding.

The reason is simple: it has not seen your account. It does not know what the number was last week, which campaigns pulled the average, or that one of them started three days ago and is still learning.

Without data, AI fills the gap with what is statistically plausible. Plausible sounds convincing and is not yours. That mechanism is behind essentially every frustration with AI in paid media.

The calculation that actually defines whether a CPL is acceptable is in contribution margin.

What changes with account access

There is a path for the assistant to query your campaigns live: an MCP connector. Authorized once, it lets ChatGPT read the account during the conversation.

The difference shows up on the first question:

Notice the pattern: the right column never asks for an opinion. It asks for a slice of your data. The opinion comes later, once you are looking at the right table.

The protocol is in what is MCP and the setup for Meta in Meta Ads MCP server.

A week of use, in practical order

Monday. "Compare spend, results and cost per result for the last 7 days against the previous 7, by campaign. List only what moved more than 20%."

Tuesday, if something rose. "Show frequency, CPM and CTR for that campaign in the same window." This separates more expensive media from a less interesting ad — two causes with opposite fixes.

Thursday, creative. "Which ads have been running over 30 days and lost CTR in the last two weeks?"

Friday, client. "Summarize in five lines what changed in this account this month, in plain language."

Four questions, always the same, always with the period stated. The repetition is what builds the history that makes comparison possible. The full list is in Meta Ads prompts.

Three mistakes that ruin the whole thing

Asking for a recommendation before asking for a number. "What should I improve here?" returns the same generic list every time.

Not stating the period. Without a date range, the assistant picks one, and it changes between answers.

Accepting the first answer. The conversation is the format's advantage. Ask where the number came from, request the campaign breakdown, challenge the conclusion.

Beyond paid media: where it also pays off

Digital marketing is not only paid media, and it is worth listing where the assistant delivers with no integration at all.

Content planning. Turning a list of client questions into article titles, grouped by intent. The input is the list of questions your sales team already hears daily.

Objection handling. Paste the three most common objections and ask for three answers each, in different tones. Works for sales scripts, captions and ads.

Creative briefs. Describe the audience, the offer and the format, and ask for five different angles. The value is not the final copy, it is breaking the blank page.

Reviewing a proposal. Ask it to flag what is vague, what has no number, and what a skeptical client would question. One of the most underrated applications.

Translating a report. Take the technical columns and get the email text that goes to the client, with no jargon.

What they share: you hand over the context along with the question. That is the rule for productive use without a connector. When the question needs data that only exists inside the ad account, there is no way around it: either you paste, or you connect.

Custom instructions worth setting once

A small setup step that improves every answer and takes two minutes. In the custom instructions, tell it three things about how you work:

Your currency and market. Otherwise it defaults to US dollars and US benchmarks, and you spend the rest of the session correcting units.

Your default period. "Unless I say otherwise, compare the last 7 days against the previous 7." This alone removes the most common source of inconsistent answers.

Your tolerance for hedging. Something like: "give me the number first, then the caveat, and skip the disclaimer that results vary." Assistants hedge by default, and in a daily routine the hedging is noise you already know.

You can add a fourth: your business model in one line. "I run paid media for B2B services with a 30-day sales cycle and a US$ 4,000 average contract" changes how every answer about CPL and scaling gets framed, without you restating it each time.

These are not prompt tricks. They are the context an assistant would have on day two if it were a person you hired, and giving it explicitly is the cheapest quality improvement available.

What stays yours

Worth being honest about the limit, because that is where expectations break.

The assistant reads the ad account. It does not know which leads became revenue — that lives in the CRM. It does not know the most expensive campaign is the only one bringing annual contracts. It does not know the client asked you to pause that product line this month.

Business context stays human, and it is what turns the table into a decision. The split between what the platform measures and what sales confirms is in MQL vs SQL.

For the assistant comparison, see Claude vs ChatGPT for marketing; for the wider view of AI in paid media, AI for media buyers. The reporting format is on the paid media reporting page.