Problem
Marketing budgets move on attribution reports, and most attribution is a rule: last touch, first touch, or a data-driven model fitted to observed journeys. None of these can see what would have happened without the ad. On real data there is no way to check them, because the truth is unknown.
Method
A simulated population of 50,000 users with a known causal ground truth: each channel's real effect on conversion is set by design, so every attribution model can be marked against the answer. Last-touch, first-touch and a data-driven model are run on the observed journeys. Then a randomised experiment (holding out exposure by channel, with confidence intervals) is run on the same population to show what a causal method recovers.
Evaluation
The dashboard models credited retargeting as the best-performing channel. The experiment showed it was second-worst: retargeting reaches people who were already going to convert, so it collects credit it did not earn. The other channels are ranked the same way, naive versus causal, in a Tableau Public dashboard (plum = causal, light rose = naive).
The decision it drives
Where to move budget, and how much to trust the report that says where. For an agency, the practical output is a case for running holdout tests before reallocating spend.