Pair Marketing Mix Models with Experiments as Tracking Erodes

Last-click and basic data-driven attribution share a bias towards channels close to conversion. They miss incremental lift and can steer budgets towards channels that appear in converting paths without generating new demand.

The user-level data these models depend on is also getting thinner. Consent refusals, Safari's and Firefox's tracking prevention, Apple's App Tracking Transparency (introduced with iOS 14.5) and walled-garden platforms that do not share user-level data all break individual journeys. The GA4 export remains valuable behavioural data, but it cannot reconstruct every cross-device, cross-platform path.

Aggregate methods avoid the problem. A marketing mix model (MMM) relates daily or weekly outcomes to spend per channel while controlling for seasonality, pricing and other drivers, so it needs no user-level identifiers. Bayesian MMMs add carry-over (adstock) and saturation effects and allow prior knowledge to rule out implausible results; Google's open-source LightweightMMM is one implementation. Geo experiments work at the same level, comparing regions rather than stitching individual journeys, which makes both approaches compatible with privacy-first measurement.

An MMM on its own is still observational. Spend often rises when demand rises, which inflates estimated effects. Structured incrementality tests, such as geo experiments, holdout groups and uplift modelling, supply causal estimates that can calibrate the model's priors or check its outputs. The two answer different questions: an experiment measures one channel precisely at one point in time, while the MMM covers the whole mix continuously.

In BigQuery, dbt models turn GA4 export events into the aggregates an MMM needs: conversions and revenue per day and region, joined with spend from each ad platform and with control variables such as promotions and price changes. For incrementality tests, privacy-first collection such as Robin Signals provides consistent conversion signals without depending on user-identifiable data.

The sustainable approach combines top-down modelling with bottom-up experiments rather than trying to track every fragmented touchpoint. It takes more effort than reading an attribution report, but it is the method that can withstand scrutiny of marketing ROI.