Marketing attribution return on investment: a causal method

Start with the buyer choice that a return figure might influence, then follow each gate in order. A failed test sends the calculation to an operational result, not to a guessed ROI. Leave numerical fields empty until the evidence, economic inputs and named approvals are ready.

If no decision depends on the return

If no decision depends on the return, set ROI aside. Name the choice, the alternative under consideration, the review date and the conditions that would stop the work. Without an action to inform, the ratio has no settled purpose and cannot tell the buyer what to do.

Treat channel-model credit as an operational allocation, not a causal numerator. Use dashboard reporting and the planning template to capture the proposed choice and stop conditions before labelling any output ROI. Keep that report separate from incremental contribution, so attribution credit is not mistaken for return.

• Rule of thumb: No buyer choice to inform, no ROI calculation.

If a choice is live and its boundaries are clear

If a choice is live, freeze its boundaries before measurement begins. Write down the eligible population, geography, outcome event and outcome window, then record the alternative being assessed. If any boundary changes after launch, treat the original comparison as potentially invalid rather than carrying its result into a revised decision.

Set contribution per outcome after refunds, and assign an analysis owner and statistics reviewer. Name the review date as well. These choices separate the commercial result from a platform conversion count; team roles can help assign responsibility for checking and releasing the evidence.

• Rule of thumb: Set the population, outcome and window before measurement.

If a comparison design can be fixed in advance

If a comparison design can be fixed in advance, select it before the campaign starts.

A randomised holdout, matched geo test or synthetic control may be appropriate; incrementality testing sets out the kinds of claims these approaches can and cannot sustain. Keep the method tied to the defined population and outcome window.

Write down likely contamination and attrition before results are visible, then arrange an evidence review against the chosen design. Do not pick a control after inspecting which answer is most favourable. A documented comparison design lets reviewers judge whether the estimate follows from the planned test rather than post-result selection.

• Rule of thumb: Decide how to compare results before seeing them.

If the planned comparison fails

If assignment fails or the comparison group no longer represents the eligible population, leave the incremental outcome blank.

A modelled allocation cannot repair a weak counterfactual, and precise arithmetic cannot make evidence from a different population causal. Stop the ROI calculation at this gate rather than filling the gap with a proxy.

An outcome definition change, identity coverage below the approved floor or expired privacy approval also calls for a pause. Use an operational result while the control issue is unresolved. Review data handling where needed, but do not treat that review as a substitute for a valid comparison or as approval to reuse the old estimate.

• Rule of thumb: A comparison that fails cannot support a causal numerator.

If the causal estimate passes review

If the comparison survives review, calculate incremental contribution using the accepted incremental outcome and the buyer-approved contribution per outcome.

Keep those inputs distinct: one is the estimated lift, while the other is the contribution after refunds. Record reviewer decisions alongside each input, and keep the ROI field empty until both are approved.

The table records formulas and placeholders, not approved results. Keep dated sign-off with the calculation so each input can be checked later. If an input lacks approval, the ROI field stays blank; launch review can check whether evidence and controls are ready for release.

• Rule of thumb: One unapproved numerator input means no reportable ROI.

A worked example: the randomised holdout

The holdout is the simplest design that supports a causal numerator. All figures below are invented to show the arithmetic, not results from any campaign.

Then add the uncertainty. The standard error of the difference between the two rates is about 0.15 percentage points, so a 95% interval for the lift runs from roughly 0.11 to 0.69 points. That gives 97 to 623 incremental purchases, £2,910 to £18,690 of contribution and an ROI between about -50% and +210%. The central 80% should be reported with that range, because the low case loses money.

If the holdout is too small for a narrow range, enlarge it or lengthen the window before launch. Do not widen the claim afterwards.

If the cost base is still incomplete

If the numerator is approved, assemble whole-life economic cost before dividing. A supplier invoice alone is not the denominator. Include supplier charges, buyer labour, data engineering, compliance review, exit, support, incidents, renewal and migration, then reconcile cash and economic cost views.

For every cost row, state currency, VAT basis, unit, period and owner. Ask the buyer's accountant to confirm recoverable VAT treatment rather than inferring it from invoice presentation. Keep this ROI cost base distinct from any decision about statutory software-asset recognition, and do not substitute a partial supplier charge for total economic cost.

• Rule of thumb: Divide by whole-life economic cost, not the invoice alone.

If assumptions or controls change

If the numerator and cost base have passed review, show low, central and high cases and identify the input that drives the range. Put uncertainty beside the estimate rather than concealing it in a single value; HM Treasury's optimism bias guidance shows one public-sector way to test cost and benefit assumptions. Keep the delivery review process attached to the assumptions.

Record the assumptions for each case and note how each one moves the result. Stop if assignment breaks, outcome definitions change, identity coverage falls below its approved floor or privacy approval expires. A changed control condition requires reassessment, not an unchanged carry-forward.

Keep source extracts, model version, formula version and reviewer decisions together in the review record. Recalculate when a documented input changes, retain the earlier figure and have the named owner sign the change log. This makes the evidence trail available for later checking without replacing the original assumptions with a revised result.

• Rule of thumb: Show uncertainty and pause when an approved control fails.

If every decision gate has cleared

If the counterfactual, numerator, denominator and reviews have all cleared, publish ROI with its low, central and high cases. Keep the population, outcome definition and cost basis consistent across the figure and its supporting record. That alignment ties the ratio to the choice it was designed to inform.

Retain signed approvals and the audit trail with the result. If the comparison is not credible, the cost base is incomplete or an input has no approval, leave ROI blank and report the operational result instead. Change the record only when a documented input moves, preserving the earlier figure for later review.

• Rule of thumb: Release ROI only when every input shares the approved scope and period.