The logic was clean. That spend was not converting. The last-click numbers were poor. On the deck, it was the easiest line to cut: move the money to the channels that were “actually working,” protect margin and look disciplined.
For about six weeks, the number agreed with me. Then it stopped.
What I had cut was not converting directly. It was helping the converting channels convert. Demand did not disappear the day I turned it off. It disappeared later, quietly, in channels I had assumed were self-sustaining. By the time the P&L showed the effect, the decision and the revenue consequence were far enough apart that no dashboard connected them.
The problem was not that the model was useless. It was that the model was a snapshot and the business had a lag. Attribution naturally rewards what sits closest to the sale. It is much less reliable at showing the upstream activity that made the sale possible.
The lesson was not “don’t cut.” Sometimes you should. It was to be much more careful when the model and the commercial reality disagree — because the missing variable may simply be time.
The deck shows you the quarter. The P&L makes you live in the one after.