Attribution assigns credit. Only an experiment estimates cause.

Before-and-after is not a control

Switching a campaign off and watching the number move confounds the campaign with seasonality, promotions, pricing and every other channel running that week. Comparing this month's attributed ROAS to last month's does the same thing with more decimal places. Neither comparison has a control group, so neither separates the campaign from the calendar.

What a test is allowed to risk

The Campaign Constitution caps the spend any single test may put at risk, sets a minimum run length, and marks which campaigns are exempt from suppression entirely. Every suppression is written by the action service as a logged, reversible change, so a test can be stopped mid-window and the prior state restored. No design starts outside those limits without a named approver.

Budget moves that survive a CFO

When a holdout estimates that most customers credited to a channel would have arrived anyway, moving money out of it becomes a defensible decision rather than an argument. When the estimate runs the other way, spend can rise on a channel whose attributed ROAS reads unremarkably. Each result is filed with the design that produced it and the dates it covers, so a decision made in March can be re-read in September.