From Output to Insight: A Step-by-Step Guide to Making AI Ask Better Marketing Questions

Start with the verdict: the genuinely useful thing AI does for marketing is not producing more material, it is producing better questions — and a better question only pays off when a team has a routine for testing it. Nielsen's 2026 Annual Marketing Report describes nearly 90% of marketers as using AI inside their workflow, so adoption itself stopped being an advantage a long time ago. What still separates teams is whether they can travel from a pattern in the data to a decision that survived a real test. The five-step routine below is written for general readers as well as practitioners: no model to build, no code to ship, just a sequence of moves that turns customer records into judgement.

Two premises run through every step. First, an association that a machine surfaces is a lead, not a conclusion. Second, the buyer side of the conversation has already moved — Forrester's State of Business Buying, 2026 research describes generative AI searches as a starting point for business buyers — which means questions now arrive from outside your own dashboards, whether or not you planned for them.

Step 0: Understand the benchmark case, and what it really taught

The example that keeps resurfacing in this debate is the U.S. retailer Target. Its aim was to recognise customers who were probably expecting a baby before competitors did, because pregnancy is a window in which long-standing shopping routines can be reshaped. The company's statistician Andrew Pole worked through purchase activity and identified about 25 products which, considered together, could signal a pregnancy and roughly estimate a due date.

Most readers met the story through Charles Duhigg in a 2012 contribution to The New York Times Magazine titled “How Companies Learn Your Secrets.” The lesson is easy to misread. It was never that fortune-telling is possible; it is that combination-level analysis exposes relationships that no single purchase reveals. At the time the work required specialists, very large data holdings, statistical modelling and serious technical infrastructure. Today a slice of that capability sits inside tools that a marketer can open without writing a line of code.

Step 1: Replace one reporting question with a comparison question

Most marketing AI conversation still orbits production: drafting emails, writing subject lines, generating visuals, spinning up campaign variants, speeding up operations. None of that is harmful, and some of it is necessary. It is simply execution of plans that already exist. The piece I published in MarTech, “AI made marketing faster than it made marketing better,” argued exactly this: faster marketing is not automatically better marketing, and quicker execution has little worth when the underlying choices have not improved.

Here is the practical switch. Reporting questions describe what happened — best campaign, top segment, conversion count, average order value. Investigative questions describe what differs between groups. Take an ecommerce marketer staring at buyers who purchased once and vanished. The reflex is to slice that audience by recency, product line or order value, then push a second-purchase programme. Sensible, but the most useful behavioural contrasts often do not sit where standard cuts place them.

Work through three hypotheses that ordinary segmentation tends to miss:

• Buyers who spread a first order across two categories may come back more often than buyers who buy several items from one category.
• Discount-led first purchases may lead to different repeat paths depending on product category.
• The gap between a customer's first browse and first purchase may relate to later value.

Move the question in stages: stop at “How many customers made a second purchase?” only long enough to replace it with “What's different about those who returned compared with those who didn't?” and then immediately press further with “What other differences should we investigate?” That last question is the one most teams skip, and it is where the widening happens — because marketers, like everyone else, tend to study what they already believe matters, through segments and metrics they already trust.

Step 2: Let the tools search wider, but keep four guardrails on

Using ChatGPT or Claude to look over a dataset is not the same as building and validating a predictive model. These tools can misread field names, arithmetic wrong, ignore confounding variables and describe an unreal relationship with total confidence. Four guardrails are non-negotiable before any output reaches a decision:

• Field check: confirm the tool read the columns you think it read — date, currency, product identifier, nulls.
• Population check: ask which customers are in the slice and which are excluded by filters, returns, or missing history.
• Rival explanation: require at least two competing reasons for the pattern before naming a favourite.
• Data-handling check: customer records uploaded for analysis carry privacy and security obligations; a survey of marketer sentiment around AI in customer data has privacy near the top of concerns.

What genuinely changed is not that every marketer became a data scientist. It is that marketers can now join analytical work directly — posing questions, requesting comparisons, challenging outputs, drawing charts — while collaborating with analysts instead of waiting in a queue. For intricate predictive or causal work, specialist expertise still matters. Data scientists are not displaced; the analysis simply becomes more accessible and more collaborative.

Step 3: Read a pattern like an analyst, not like a headline

An instructive illustration comes from independent work on the University of California, Irvine Online Retail dataset. The Irvine extract contains more than 540,000 transaction records from a UK retailer covering December 2010 through December 2011. In the published analysis, repeat buyers accounted for 65.6% of customers while generating 92.9% of revenue — a number that looks like a mandate to pour money into retention.

Resist the reflex, for a simple reason: repeat buyers place more orders than one-time buyers by definition, so a larger revenue share is arithmetic, not evidence. The figures describe how revenue splits between groups; they do not explain why people return, and they say nothing about whether a specific retention tactic caused anything. Better questions come first: do returning customers also spend more per order? Which first-purchase attributes correlate with coming back? Which of those attributes can marketing plausibly influence at all?

The trap is treating a relationship as an explanation. Numbers can be perfectly correct while the lesson drawn from them is wrong.

Step 4: Borrow behavioural science before you borrow budget

Customer behaviour is not a heap of transactions and clicks. It is shaped by habit, motive, perceived risk, cognitive effort, expectation and context — and the numbers usually locate behaviour without naming the driver. Consider shoppers who browse constantly yet rarely purchase. Analysis may tie them to a product category, a price band, a device or an acquisition source. That tells you who and where, not why.

Run the candidate explanations out loud: is the decision genuinely hard to make? Are they comparing options? Do they read buying as risky? Are they researching here and purchasing elsewhere? Or are they simply not in-market? Each answer points to a different move — clearer product detail and reassurance for the uncertain buyer, comparison help for the one weighing choices, useful content rather than a further discount for the one who is not ready, and honest acceptance that some people have already decided to buy somewhere else.

Step 5: Convert the finding into a test that can lose

AI can point at relationships worth exploring. Behavioural science can propose explanations. Only testing shows whether changing something actually shifts the behaviour you care about. Skip the last two and you have manufactured fancier assumptions. Go back to the repeat-purchase case: if review suggests that first-time buyers who explore products across multiple categories return more often, one option is to switch on broader cross-category recommendations immediately. The more cautious option asks what is really going on — were those buyers already more engaged, do they simply have wider needs, does browsing multiple categories build trust, or is it just something committed buyers do anyway?

That distinction decides the test. If wider product discovery might encourage returns, test whether helping single-category shoppers find relevant complementary products lifts second-purchase rates. Score it on incremental buying impact, not recommendation clicks. Design the test honestly: fix the sample size before you start reading results, and change one variable at a time, because peeking early and stacking variables are among the most frequently documented causes of false positives in experimentation platforms such as Adobe Target.

How the same data set looks at each stage

Stage	Typical question	Stronger question	Output
Reporting	Which campaign won?	—	Ranked results
Association hunting	What correlates with returning?	What groups differ in ways I never cut for?	Shortlist of leads
Explanation	Why would that be true?	Which rival explanation fits best?	Named hypothesis
Verification	Did we send more emails?	Did incremental purchase rise versus a holdout?	Decision or dead end

Five misconceptions that quietly burn analysis budgets

• “A clean percentage is a finding.” It is a description of the slice you pulled; the same number can support opposite conclusions.
• “Every pattern needs a campaign.” Some patterns deserve a question, some deserve nothing, and a few deserve a deeper data check first.
• “Confident wording means correct maths.” Generative tools deliver both at the same confidence level, verified or not.
• “Better segments prove themselves.” Predefined lifecycle segments may describe no distinct behaviour at all; test that before automating journeys.
• “The tool replaces the statistician.” It replaces queue time, not validation.

FAQ

I do not have an analyst. Where do I start?

With one comparison question on one behaviour you already measure. Take a group that converted and a group that did not, ask what differs, and demand that any answer name the fields used. You are building a habit, not a model.

How do I know an AI-produced number is trustworthy?

Recompute the headline figure independently — a spreadsheet total or a single query run by a colleague is enough. If a tool cannot reproduce the number when asked to show its steps, treat the whole output as a draft.

Is this still relevant if my data is small?

Yes, with lower ambitions. Small samples rarely support causal claims, but they are plenty for generating hypotheses worth watching in the next quarter. The failure mode is acting on small-sample associations as if they were established laws.

How long before a pattern becomes a decision?

A useful rule of thumb: association to hypothesis in days, hypothesis to test result in one full purchase cycle. Any timeline shorter than your customers' buying rhythm is guessing.

Your first 30 days: action checklist

1. Pick one metric your team reports on every month and rewrite it as a comparison question.
2. List three to five segments you have never cut, and ask the tool to contrast them on the metric.
3. Run the four guardrails before sharing any output: field check, population check, rival explanation, data-handling check.
4. Write down two behavioural explanations for the strongest pattern, and mark which one marketing can influence.
5. Design one test with a holdout group, a pre-set sample size and a single changed variable.
6. Define the winning measure as incremental purchase or revenue, never clicks or opens alone.
7. Close the loop: log the questions that produced nothing, so the next round starts from a shorter list.

There is a joke in the current excitement: we spend enormous effort asking machines for answers while the scarce skill is writing better questions. Ask what separates lapsed customers from active ones before requesting another re-engagement sequence. Study how groups respond to incentives over time instead of shipping more discount headlines. Check whether predefined segments reflect distinct behaviour before adding journeys. Look for an alternative reading before accepting the one your report implies. Spotting correlations will matter less and less as an edge; the durable advantage belongs to teams that can tell a signal from noise, explain behaviour instead of describing it, and recognise when an attractive finding still does not justify action.