AI marketing analytics: turn questions into useful decisions

Ask a question with boundaries

“How is marketing doing?” leaves too many choices implicit: which market, which period, which outcome, which sources, and which comparison? A more useful prompt says what decision is under review and what evidence should be considered—for example, compare the last four complete weeks with the preceding four using matched orders and connected ad spend.

If you do not know which metric is appropriate, ask the assistant to clarify before calculating. The system should not silently select a flattering window, switch from net to gross revenue, or treat a missing source as zero. Scope is part of the analysis, not prompt decoration.

Ask the assistant to show its basis

For every material claim, a reader should be able to find the data period, source accounts, metric definition, calculation, and model assumptions. Ask for the records or totals that support the explanation and for any important sources that were unavailable. If the interface cannot trace a claim back to evidence, keep it out of the decision brief.

A transparent calculation is often simple: attributed revenue divided by eligible spend, using stated inputs and a stated model. But the phrase “return” can conceal different numerator and denominator definitions. Label the measure precisely and avoid treating platform-reported revenue as interchangeable with finance-confirmed revenue.

Separate observation from interpretation

An observed fact might be that recorded spend rose in a given period. A calculated result might be the change in matched revenue under one attribution rule. An interpretation might be that a campaign drove the change. The first two can be checked against data and calculation; the third needs additional evidence and may require a causal test.

A good AI answer labels those layers. It can propose explanations—such as a budget shift, a tracking change, or a different sales mix—but should present them as hypotheses until verified. A confident narrative should never convert correlation or attributed credit into causal proof.

Use conversation to investigate, not to skip review

An assistant is useful for summarizing a period, finding an unusual movement, explaining a model, or suggesting the next check. The user still needs to confirm the account, date range, and coverage before acting. If the answer proposes changing a campaign or budget, require the appropriate human approval and inspect the supporting evidence first.

Treat a mismatch as a useful question. If an AI summary conflicts with a finance report, pause and reconcile definitions, time zones, refunds, and deduplication. Do not prompt the system until it produces the preferred answer. The discrepancy itself may reveal a data-quality problem.

A concise review checklist

Before sharing an AI-assisted conclusion, a marketer should be able to explain how it was produced and what it leaves out.

• What exact question and period did the assistant use?
• Which connected accounts and records support the answer?
• What calculation, attribution model, and conversion window were applied?
• What data is missing, late, estimated, or unmatched?
• Is the final statement descriptive, attributed, or causal—and is that distinction clear?

A conversation that gets to a useful answer

Start with a specific request: “Compare paid marketing for the last four complete weeks with the preceding four. Use net order revenue and connected ad spend. Show source coverage before explaining any changes.” This gives the assistant an outcome, a period, a comparison, and a reporting priority. If a source is missing, the answer should explain that before suggesting a channel ranking.

Then ask: “Which channels changed most under the same attribution model?” A useful response keeps the model and conversion set stable. It separates the spend change from the credit change and tells you whether the result is driven by a different number of orders, a different order value, or a shift in journey mix. None of those explanations alone proves why the change occurred.

Follow with: “Show examples of journeys behind the largest change, and tell me what to check next.” The assistant should point back to available records, name any sampling limits, and propose concrete checks such as campaign tags, landing-page changes, returning-customer share, or delayed imports. End by asking for a short brief with findings, open questions, and owners.

Use prompts that name the reporting job

A good prompt does not need to be long. It needs to name the task clearly enough that the system can preserve its scope. Save prompts that your team repeats, but review their metric definitions when the business changes. A prompt written for paid orders should not be reused unmodified for qualified leads.

If the assistant lacks the requested source or detail, the useful response is a direct explanation of what is available and what additional data is needed. Do not encourage it to fill the space with a likely-sounding narrative. A narrower answer grounded in actual totals is more valuable than a broad answer that cannot be checked.

Check arithmetic before accepting the explanation

Language models can produce a persuasive explanation even when a calculation is incorrect or its inputs are inconsistent. A reporting assistant should retrieve and calculate with appropriate tools, then explain the result. Users should still be able to see the numerator, denominator, comparison period, and rounding convention for any important ratio.

Suppose fictional attributed revenue rises from £20,000 to £24,000 while spend rises from £5,000 to £8,000. Revenue increased 20%, spend increased 60%, and attributed ROAS fell from 4.0 to 3.0. A summary saying “revenue growth means efficiency improved” would miss the denominator. Ask the assistant to separate volume from efficiency and show the calculation.

Also check whether totals can be added. Channel rows from one deduplicated model may sum to a model total. Platform conversion claims from separate ad networks may overlap. A fluent answer that sums overlapping claims is still wrong. The meaning of the data matters as much as the arithmetic.

Keep context intact across follow-up questions

A marketer naturally asks “why?” or “what about last month?” after an answer. The assistant should keep the workspace, accounts, metric definition, and model in context while making any changed period explicit. If a question switches from acquisition to retention, it may need a different customer cohort and outcome. Clarifying that switch is part of useful conversation.

Before exporting an answer, ask it to restate the final scope in one sentence. This catches accidental drift after several follow-ups. If the question moved from all customers to new customers only, the brief should say so. The answer's confidence should reflect the available records, not how long the conversation has lasted.

Turn recommendations into reviewable actions

A suggestion to increase a budget should include the supporting observation, the assumptions, and what would change the recommendation. For example, a channel may have rising credited revenue but a falling margin because discounts increased. An assistant that sees only revenue should ask for the missing cost context or limit its recommendation accordingly.

Keep external actions distinct from analysis. Drafting a recommendation, generating a report, and changing an ad account are different jobs. Before a consequential campaign action, review the account, budget, period, and supporting evidence. After an action, measure against the original expectation so the team can learn whether the recommendation was useful.