Marketing Data Analysis AI

The short answer

AI is excellent at the mechanical part of marketing analysis: cleaning exports, joining tables, calculating ratios, drawing charts and describing what changed. It is unreliable at the part that matters most, which is judging why something changed and what to do. It will also, if allowed, state numbers it has not calculated. The method below keeps the first strength and guards against the rest.

Step 1: start with a decision, not a dataset

"Analyse our marketing data" produces a tour of charts. "Should we move budget from campaign A to campaign B next month?" produces an answer. Write the decision in one sentence and list the two or three numbers that would change your mind.

Step 2: choose the right method for the question

Marketing questions, methods and AI's role

Question	Suitable method	What AI can do	What it cannot do
What changed this week?	Period comparison and anomaly detection	Compute changes, flag outliers, write the summary	Know whether a change was caused by a tracking fault
Which campaign is most efficient?	Cost per result and return on spend by campaign	Calculate and rank from exports	Correct for attribution differences between platforms
Where do visitors drop off?	Funnel analysis	Build the funnel from event data and describe gaps	Explain why people leave without research or testing
What do customers talk about?	Text analysis of reviews, calls or surveys	Group themes, count them, pull quotes	Decide which theme matters commercially
Did this change cause the uplift?	A controlled test, or a before and after with a comparison group	Set up the calculation and check significance	Turn an uncontrolled comparison into proof of cause
How should budget split across channels?	Marketing mix modelling	Help prepare data and interpret outputs	Replace the data volume and expertise the model needs

Step 3: prepare the data yourself, or check the preparation

Export the data for the period and a comparison period. Before any analysis:

• • Check totals in the export match what the platform shows.
• • Note known gaps: tracking changes, consent banner changes, site outages, sale periods.
• • Remove or anonymise personal data unless the analysis needs it. The UK GDPR applies to personal data in AI tools; the ICO's guidance on AI and data protection covers minimisation and lawful processing.

Understand what your figures represent. For example, if you use Google's consent mode, some conversions in Google Ads and GA4 are modelled rather than observed; Google documents that modelled conversions appear in reports alongside observed ones. That is useful, and it is also a reason not to over read small changes.

Step 4: make the AI show its working

Use an assistant that can run code on uploaded files (several major assistants can) and instruct it:

• • "Load the file, print the column names and row count, and confirm the total of the spend column before doing anything else."
• • "Do every calculation in code and show the code."
• • "For each finding, give the exact figures it rests on."

This turns the AI's answer from an assertion into a calculation you can check. If the tool cannot run code, ask for spreadsheet formulas and apply them yourself.

Step 5: verify before you believe

Recalculate the one or two numbers your decision depends on by hand or in a spreadsheet. Check that the AI used the right date range and did not drop rows. Ask it directly: "What would make this conclusion wrong?" Good analysis names its own weaknesses.

Be most sceptical of causal language. "Revenue rose because of the new campaign" is a claim that needs a test or a comparison group. Without one, the honest wording is "revenue rose in the same period as the new campaign".

Step 6: write the decision, not the analysis

End with a short note: the decision, the two or three numbers behind it, the known weaknesses, and when you will check the result. That note is what gets used; the charts rarely are.

Worked example

Illustrative example with round, hypothetical figures.

A business runs two paid campaigns and wants to know whether to move budget from A to B.

• • Campaign A: spend £2,000, 40 enquiries, 8 sales at an average gross profit of £400.
• • Campaign B: spend £1,000, 25 enquiries, 3 sales at an average gross profit of £400.

The AI, asked for efficiency, reports that B is better because its cost per enquiry is lower. Checking by hand:

• • Cost per enquiry: A is £2,000 ÷ 40 = £50. B is £1,000 ÷ 25 = £40. The AI is right on this measure.
• • Cost per sale: A is £2,000 ÷ 8 = £250. B is £1,000 ÷ 3 ≈ £333.
• • Gross profit after ad spend: A is 8 × £400 − £2,000 = £1,200. B is 3 × £400 − £1,000 = £200.

On the measure that matters to the business, A is far stronger. The AI answered the question it was asked (efficiency per enquiry) rather than the decision (profit). And with only 3 sales, B's figures are too small to conclude much either way; a sensible decision is to keep A's budget and give B another month before judging. Assumptions: all sales are correctly attributed to one campaign and gross profit per sale is the same for both.

If you want to run your own figures, the ROAS calculator works through the same arithmetic for ad spend, revenue and margin.

Built in AI analysis tools

GA4 generated insights summarise notable changes in your analytics in plain language at the top of reports. They are a good early warning. They do not know about your sale last week or your broken form.

Marketing mix modelling estimates how each channel contributes to sales from aggregated data over time, without tracking individuals. Google's open source Meridian uses Bayesian methods for this. It needs a data scientist and a long history of spend and sales data, so it suits larger advertisers.

When to get help

If the numbers your decisions depend on do not reconcile across platforms, fix measurement before analysis; no AI can analyse its way out of broken tracking. AfM's managed service includes reporting on agreed outcome measures, with the founder reviewing what the data does and does not show.

FAQ

Can I upload my Google Ads or GA4 export to ChatGPT or Claude?

Technically yes. First remove personal data, check the provider's data settings for your plan, and ask it to show its calculations as code so you can verify them.

Why do the AI's numbers sometimes differ from the platform?

Usually a different date range, filtered rows, currency or time zone settings, or modelled conversions counted differently. Ask it to print row counts and totals first and compare them with the platform.

Is marketing mix modelling worth it for a small business?

Usually not yet. It needs a long history of spend and sales data across channels and specialist skills. Simple controlled tests, such as pausing a channel in one region, often answer the same questions more cheaply.

What is the most common mistake in AI assisted analysis?

Accepting a fluent explanation of why something happened. AI is good at describing what changed and poor at establishing cause. Ask what would make the conclusion wrong.