Last-click attribution is not a model. It is the absence of one: a rule that gives full credit to whatever happened most recently, which in practice means giving full credit to branded search and retargeting for demand that something else created.
The three views you need
No single method is trustworthy alone. Useful measurement triangulates between three, and treats disagreement as information rather than as a problem to be reconciled away.
- Platform-reported: fast, granular, and structurally biased toward the platform reporting it.
- Multi-touch modelling: better at relative channel contribution, dependent on identity resolution that keeps degrading.
- Incrementality testing: the only one that answers the actual question, and the slowest and most expensive to run.
Start with holdouts, not models
Teams reach for a multi-touch model first because it feels rigorous and requires no lost revenue. But a geo-split holdout on a single channel, run for four weeks, will tell you more about whether that channel is incremental than six months of model refinement. Run the cheap experiment before building the expensive apparatus.
Report ranges, not points
Every attribution number is an estimate with meaningful uncertainty. Reporting it as a single figure to two decimal places invites decisions the data cannot support. Present a range, state which method produced it, and be explicit about what would change your mind.
The goal is not a perfect model. It is a shared, defensible basis for deciding where the next pound goes.
- Attribution
- Measurement
- Media
