Attributed Revenue Is Not the Same as Causal Lift

Attributed revenue is not the same as causal lift.

Attribution answers a crediting question: Which touchpoint or channel receives credit under the selected model? Incrementality asks a counterfactual question: What happened because of the intervention that would not have happened otherwise?

Both can be useful. Confusing them leads to decisions the evidence cannot support.

Begin with the decision

The measurement method should match the decision being made.

If the decision is whether to send a campaign again, a randomized holdout may provide a useful comparison. If the decision is how to divide credit across several channels, an attribution model may help describe the observed path. If randomization is not available, a cohort, matched comparison, time-series design, or other quasi-experimental method may be the strongest feasible option.

The method does not need to be perfect. Its limits need to be visible.

Build the measurement contract

Before launch, state:

1. the primary decision;
2. the baseline or comparison;
3. the assignment method;
4. the primary result and time window;
5. guardrail measures;
6. exclusions and known sources of bias;
7. the review date;
8. the rule for keeping, changing, or stopping the intervention.

This prevents the team from choosing the most flattering result after the data arrives.

Use holdouts carefully

A holdout is useful only when its design supports the comparison.

Check whether assignment was random, groups remained comparable, customers crossed between treatments, the observation window was long enough, and other channels changed during the test. Document material limitations.

“The holdout was lower” is not a complete finding. The decision needs the size of the difference, uncertainty, operational cost, guardrails, and evidence that the comparison remained valid.

Separate direct and influenced results

CRM and lifecycle programs often affect customers beyond the last click. That does not justify assigning every subsequent purchase to the program.

Use separate views:

• Direct response describes behavior tied to the defined interaction rule.
• Attributed or influenced revenue follows a disclosed crediting model.
• Incremental lift estimates the difference caused by the intervention relative to a valid comparison.

These numbers may differ because they answer different questions.

Close with a decision

A useful measurement review should end with one of four outcomes:

• keep the mechanism;
• change the mechanism and retest;
• expand it under stated conditions;
• stop it.

Record the owner and next review date. If the evidence is inconclusive, say which uncertainty matters and what test could reduce it.

Measurement earns trust when it constrains the claim as carefully as it supports the action.