Attribution

Attribution assigns conversion credit to touchpoints. It is worth being precise about what that means: every attribution model is an opinion about causation expressed as an accounting rule, not a measurement of it. The model does not discover which ad caused the sale. It applies a policy for dividing credit among the touches it happens to be able to see.

The policies differ in predictable ways. Last-click gives everything to the final interaction, which rewards whichever channel sits closest to the purchase and systematically starves the channels that created the demand. First-click does the reverse. Linear and time-decay split credit by position or recency, which feels fairer without being more true. Data-driven models fit weights to observed patterns, which is a genuine improvement, but only across touches inside the platform’s own field of view.

Two structural problems sit underneath all of them. First, each platform grades its own homework: it can only see conversions it touched, and it credits itself for those, which is why per-channel reports routinely sum to more revenue than the business recorded. That is not any single platform lying. It is several parties independently claiming the same sale. Second, no model can see what would have happened anyway, and that gap is widest exactly where reported performance looks best — brand search, retargeting, and any campaign harvesting people who had already decided.

The number that follows from this is the reconciliation. If Google claims $180,000, Meta claims $120,000, and the business recorded $240,000 in total revenue, the claims exceed reality by 25% and no model reallocates that away. The overlap is real journeys touched by both, and the only way to learn how much of either channel was necessary is to remove one and observe.

So the workable posture has three parts. Pick one model as the operating lens and use it consistently for within-platform decisions, because a consistent lens with known bias beats switching between models to find the flattering one. Keep MER, computed from total revenue and total spend, as the referee for whether the whole system is profitable — it cannot double-count, because it never asks which channel deserves credit. And resolve high-stakes disagreements with incrementality tests rather than model debates, because a geo holdout answers the causal question that no attribution model is built to answer.

One practical discipline: whenever an attribution figure appears in a decision, note which model produced it and over what window. Most attribution arguments turn out to be comparisons between two different models, or the same model over different windows, which is a reconciliation problem rather than a disagreement about the business.