A great deal of energy goes into perfecting attribution, on the unstated assumption that certainty is available if the model is good enough. It is not available, it has never been available, and waiting for it is more expensive than deciding without it.
I have built attribution models, and the useful thing I learned from doing it is how much of the value came from the arguing rather than the output. Building the model forces a team to say out loud what it believes about how customers decide. That conversation is worth more than the number it produces.
Where precision misleads
The risk with a well built model is not that it is wrong. It is that it is precisely wrong, and precision reads as authority. A dashboard with two decimal places invites decisions that the underlying data cannot support, and the confidence is manufactured entirely by the formatting.
The risk is not a model that is wrong. It is a model that is precisely wrong, because precision reads as authority.
What works better
- Contribution thinking. What would we lose if this channel went to zero, rather than what did it get credit for.
- Holdout tests. Turn something off in one market, deliberately, and watch. Uncomfortable and extremely informative.
- Directional agreement across several imperfect views, rather than one authoritative view.
- A stated tolerance. Decide in advance how wrong a number can be and still support the decision.
Holdouts are the ones teams resist, because switching off spend feels like deliberately losing money. It is, briefly. It is also the only method that answers the question directly, and the cost of not knowing compounds every quarter you postpone it.
The standard to hold
The goal is not the correct number. The goal is a decision you would make again, given the same information. That is a lower bar than accuracy and a far more useful one, because it is achievable this quarter.
It also changes the tone of the conversation. Teams arguing about whose number is right tend to stop arguing when the question becomes what they would each do differently if their own number were true.
One practical habit
Write the decision down with the number that drove it, then revisit it a quarter later. Not to score anyone, but to calibrate. Teams that do this develop a feel for which of their numbers deserve trust, and that feel outperforms any model.