E-commerce attribution: seven checks before you move budget

1. Choose the outcome you mean by revenue

A store's gross order value, net sales after discounts and refunds, payment captures, and advertising-platform conversion value are different measures. Pick the one appropriate to the decision and label it. For profitability decisions, revenue alone is not contribution margin; product cost, fulfilment, discounts, payment fees, and returns may matter.

Write down how cancelled orders, partial refunds, tax, shipping, gift cards, and repeat purchases are handled. If a customer places two orders, decide whether the reporting unit is the order or the person. Apply the same convention across periods.

2. Reconcile platform-reported conversions

Ad platforms can report conversions using their own click and view windows, modeling, identity, and event integration. Their totals may overlap because multiple platforms can each claim credit for the same order. Treat them as each platform's reporting view, not as independent revenue that can safely be summed.

Compare platform events with store orders using a fixed period and clear conversion definition. Investigate duplicate purchase events, missing order identifiers, browser/server event deduplication, and timezone boundaries. Keep unexplained differences visible until reconciled.

3. Inspect the whole recorded journey

Last-touch reports are easy to understand but can give all modelled credit to the final recorded visit. That can favor branded search, direct visits, or retargeting that appears close to checkout. It can understate earlier interactions that introduced or educated a shopper.

Use a few attribution models to understand how the answer changes when credit is allocated differently. Treat the comparison as a sensitivity check, not proof that one model reveals the true cause. Sample individual journeys to see which sources were recorded and what is absent.

4. Check returns and repeat-customer mix

A campaign can acquire high order volume while attracting low-margin or high-return purchases. Where the data supports it, compare new and returning customers, return rates, average order value, and contribution margin. State the cohort window and avoid implying that a customer was truly new if the identity history is incomplete.

Post-purchase adjustments can arrive days or weeks after conversion. Decide when a period is mature enough for review and whether recent outcomes remain provisional. A recent campaign's apparent efficiency may change after cancellations or refunds are recorded.

5. Validate spend and source coverage

Confirm that all relevant ad accounts, markets, currencies, and campaign types are included. A missing account or failed import can make a blended comparison misleading. Check the last successful sync, date range, spend currency, and whether fees or credits are included.

Unknown spend is not zero spend. Mark a period incomplete when source coverage is partial, and repair the connection before making a precise blended return claim. Keep channel-level and total coverage visible to anyone using the report.

6. Keep attribution and incrementality separate

Attribution assigns credit under a selected rule to recorded interactions. It can help standardize reporting, prioritize investigation, and understand how a journey is represented. It does not estimate by itself how many purchases would disappear if the campaign stopped.

When the decision is whether spend caused additional sales, use a randomized holdout, geographic test, or another design suited to the campaign and available scale. The test's uncertainty and limitations matter too, but it answers a different question from attribution.

7. Turn discrepancies into a repeatable review

Document the source of truth for orders, definitions, windows, and timing. Reconcile on a recurring cadence, sample customer journeys, and log meaningful tracking or campaign changes. Compare periods only after checking that the measurement setup remained comparable.

A useful budget review can state what is known, what is modelled, what is incomplete, and what should be tested next. Avoid universal claims such as a fixed percentage of revenue being lost to one attribution mistake unless a credible study directly supports the exact population and method.

Build a reconciliation worksheet

Create one worksheet for a completed reporting period. Start with the store's authoritative outcome and separate gross order value, discounts, refunds, cancellations, and the final revenue definition. Then list spend from each connected ad account. Add the time zone, currency, date coverage, and source update time beside each total.

Next record platform conversion claims separately from journey-based credit. The purpose is to understand differences, not force every source to agree. For each discrepancy, note its likely definition or timing cause and what you checked. If the cause is unresolved, leave it unresolved in the worksheet instead of applying a balancing adjustment.

Only calculate a blended ratio when the revenue and spend scopes are compatible. A store-wide revenue total divided by one selected campaign's spend is not a campaign ROAS. If you intentionally compare overall revenue with all paid-media spend, label it as a blended revenue-to-ad-spend measure and say which costs it excludes.

Separate acquisition from returning-customer demand

A high-return campaign can be dominated by customers who already know the brand. That can still be commercially useful, but it answers a different question from acquiring new customers. Segment first-time and returning purchases using a consistent customer definition and the history available in your commerce system.

Be careful with an incomplete history. If you only imported one month's orders, an existing customer may appear to be new because their earlier orders were not included. State the historical coverage behind the classification. Also consider whether the business defines new at the store, brand, or group level.

Compare channel credit, acquisition cost, order value, and later repeat purchasing by cohort when you have adequate data. Avoid crediting a channel with future lifetime revenue that has not occurred. Observed repeat revenue and forecast lifetime value should be shown as different measures.

Check margin before scaling a strong revenue result

Imagine a fictional order with £100 of net revenue, £45 of product cost, £10 of fulfilment and payment costs, and £25 of allocated advertising spend. Under this simplified contribution calculation, £20 remains. A promotion that increases revenue while sharply increasing product discounts or returns may reduce that contribution even when the dashboard's revenue trend looks healthy.

Your business may use a different cost convention. Agree it with finance and keep the definition visible. Attribution allocates revenue credit; it does not automatically allocate every cost or tell you which costs vary with the next order. When deciding how much more to spend, understand the economics of additional sales rather than relying only on an average attributed return.

Investigate branded search and retargeting carefully

Branded search and retargeting often appear near checkout. Under a last-touch view they may receive substantial credit for demand that began elsewhere. This does not make them worthless; it means the journey and decision context matter. Compare their role across models and inspect how often they appear after a customer has already engaged.

Separate brand and non-brand search where the source data supports it. Distinguish prospecting from retargeting campaign groups and new from returning customers. Avoid treating every campaign within a platform as one acquisition engine. If the question is whether a campaign adds sales that would not otherwise happen, use an appropriate experiment rather than treating a model comparison as the answer.

Create a weekly operating rhythm

Use a completed period and review source health first. Reconcile the store outcome, inspect spend movements, compare primary-model channel credit, and review the most material exceptions. Limit the meeting to the questions that could change an action. Keep detailed checks available for anyone who needs to follow the result.

Write down each proposed action with its reason and what success would look like. An action might be repairing a purchase-event duplicate, reviewing a landing page, or testing a small budget change. At the next review, compare the observed result with that expectation while accounting for offers, stock, and other business changes. This turns attribution from a disputed scorecard into a repeatable working process.