Attribution is the most contested topic in marketing measurement and most of the argument rests on a category error. An attribution model does not measure what caused a sale. It applies a rule for dividing credit among the touchpoints that happened to be observed.
A model is a policy, not a measurement
Last-click gives everything to the final touchpoint. First-click gives everything to the first. Linear spreads it evenly. Data-driven models distribute credit according to observed patterns in your own conversion paths.
None of these is measuring causation. Each is applying a rule to a set of observations. The rules disagree because they encode different assumptions, not because some are more accurate than others.
This matters practically. When someone says a channel “drove” a number of conversions, the honest reading is that a model assigned it that credit under a particular policy. Change the policy and the number changes without anything in the world having changed.
What the models systematically miss
Every model only sees touchpoints it can observe. It cannot see the conversation with a colleague, the podcast mention, the billboard on the commute, or the two years of brand familiarity that made the eventual search feel like a safe choice.
Cookie restrictions, cross-device journeys, and privacy tooling remove more of the observable path every year. The model does not report reduced confidence when its inputs degrade — it produces the same clean-looking output over a thinner picture.
This is why brand and upper-funnel activity chronically under-report. They influence outcomes through paths the model structurally cannot observe, and a model that cannot see something reports zero rather than unknown.
The incrementality question
The question worth answering is not how credit should be divided but what would have happened anyway. If you turned this channel off tomorrow, how many of the conversions it currently receives credit for would still occur?
For branded search the honest answer is often “most of them.” Someone typing your company name has already decided. The ad captures a click that the organic result would have received at no cost. Attribution reports this as excellent performance because the model cannot distinguish capture from creation.
Answering the incrementality question requires experiments — geographic holdouts, scheduled pauses, matched-market tests. These are more work than reading a dashboard and they are the only method that addresses causation directly.
How we use attribution anyway
We do not discard attribution. We use it for what it is good at: comparing like with like over time, spotting directional shifts, and allocating within a channel where the confounders are broadly similar.
We do not use it to settle cross-channel budget arguments, to value brand activity, or as evidence of causation. For those we run experiments, and we accept that experiments are slower and less satisfying than a number that is already on a dashboard.
The practical discipline is to say which one you are doing. A report that presents modelled credit in the same visual language as measured revenue invites the reader to confuse them — and readers reliably accept the invitation.
Have a version of this problem? Tell us about it — we are usually happy to give a straight answer even when there is no engagement in it.