Build a Weekly Ecommerce Performance Review With AI

A weekly store review should explain what changed, what might be driving it, and what you will investigate next. AI can prepare the first pass, but it needs consistent data and explicit metric definitions.

Begin with a small report you can verify manually. Avoid combining every available dashboard before you know whether their dates, currencies, and revenue definitions match.

Prepare the input

Export the relevant period and a comparable prior period. Include net revenue, orders, sessions, purchasing sessions, new customers, acquisition spending, product costs, fulfillment costs, refunds, and stock issues where available.

Remove personal customer information unless it is essential and your approved tool can handle it appropriately. Document the store timezone, currency, treatment of refunds, and which costs are actual versus estimated. Preserve zero values separately from missing values.

Add a short context note: promotions, stockouts, price changes, tracking updates, holidays, or unusual traffic. Without this context, the model may invent a business explanation for a data difference.

Use a structured prompt

Analyze these two comparable periods using the supplied metric definitions. First list missing, inconsistent, or duplicate data. Then calculate absolute and percentage changes, showing formulas. Separate observed facts from possible explanations. Do not infer causation or invent benchmarks. Return three questions to investigate and at most three proposed actions, each tied to evidence and a guardrail. Use “not calculable” when a denominator is zero or missing.

Ask for a machine-readable calculation table as well as prose when possible. Use deterministic spreadsheet or code calculations for the arithmetic and the model for interpretation.

Check the important numbers

Recalculate revenue, order counts, CVR, AOV, CAC, and contribution independently. Verify that refunds are not deducted twice and that total acquisition costs are not confused with ad spend.

Suppose revenue rises 20% while contribution rises only 2%. That is a reason to inspect discounting, product mix, fulfillment, and returns. It is not proof that any one of them caused the gap.

If a prior-period value is zero, percentage growth is undefined in the usual formula. Report the absolute change rather than an invented infinite or 100% improvement.

Convert observations into decisions

Use a short review format: one paragraph on business health, a table of material changes, three investigations, and an action log. Assign owners and dates outside the model if it does not have authoritative staffing information.

A useful action is “Inspect mobile checkout errors for paid-social visitors because checkout starts stayed stable while completed purchases fell.” “Improve marketing” is too vague to execute or evaluate.

Keep a consistent history

Save the input definitions and final reviewed report. Mark tracking or attribution changes so later comparisons do not treat them as customer behavior. Keep a record of whether prior hypotheses were supported.

Do not let the weekly cadence force action when the sample is too small or the system is stable. Sometimes the correct decision is to gather more data, complete an existing improvement, or investigate an operational exception.

Start with a report you could explain without AI. Then use AI to shorten preparation and sharpen questions while retaining responsibility for the calculations and decisions.