Should credit add up to 100% — or to the cost of removal?

Should credit add up to 100% — or to the cost of removal?

The question: there are two philosophically different ways to assign credit, and they answer different questions. One splits a fixed pie so all channels sum to 100% of conversions. The other asks, per channel, 'what breaks if this disappears?' These give different — and non-additive — answers. Which question are you actually asking?

The two frames:
— Fractional/allocative models (linear, U-shaped, Shapley) distribute total observed conversions across touches. By construction they sum to the total. Good for dividing a budget you've already committed.
— Removal-effect / incremental models (Markov removal effect, and true holdout incrementality) ask how much you'd lose by cutting each channel. These famously do not sum to 100% — they can sum to far more, because channels overlap and back each other up. Good for deciding what to cut.

Why this matters: practitioners conflate them constantly. They run a Shapley model (allocative) and then make removal decisions ('this channel is only 8%, let's cut it') — but an allocative 8% says nothing about what happens on removal, because other channels may not absorb the gap. The questions 'who gets credit?' and 'what's the cost of removal?' have genuinely different answers.

The nuance: true incrementality is the only frame that directly answers removal — and only an experiment measures it cleanly.

What to actually do: match the model to the decision. Dividing committed budget → allocative. Deciding what to cut or add → removal-effect or holdout. Never use one to make the other's decision.

Bottom line for practitioners: 'credit' and 'incremental value' are different currencies. Allocative models sum to 100% and answer division; removal models answer subtraction. Spend the right one on the right question.