Myth: get attribution right first, then layer on MMM and experiments later

Myth: get attribution right first, then layer on MMM and experiments later

The question: the sensible-sounding roadmap is to perfect your multi-touch attribution, then add media mix modeling and incrementality tests as advanced refinements. Is that the right sequence?

Why the order is backwards: this treats attribution as the foundation and experiments as the polish, when the causal hierarchy runs the other way. Randomized experiments sit at the top of the evidence pyramid — they measure causation directly. MMM sits in the middle — aggregate, confounder-aware, calibratable to experiments. Attribution sits at the bottom — observational credit allocation with no counterfactual. Building your measurement program on the weakest causal layer and bolting the strongest on as an afterthought inverts the epistemics. You end up using rigorous experiments merely to 'tweak' a model whose core claims they should be governing.

The nuance: the modern best-practice framework (often called unified or triangulated measurement) explicitly runs the other direction. Experiments establish causal anchors. MMM is calibrated to hit those anchors and extended to channels and history experiments can't cover. Attribution is then used only for fast, in-flight tactical decisions within the bounds the other two methods validated — never as the source of truth about channel value.

Bottom line for practitioners: don't sequence experiments last. Stand up even a lightweight always-on geo-holdout and a basic MMM early, and let them constrain what your attribution dashboard is allowed to conclude. Attribution is the fast tactical layer governed by causal methods above it — not the foundation everything else refines.