Photo by Carlos Muza on Unsplash
Marketing-mix modelling was supposed to die a decade ago, replaced by deterministic last-click attribution. Instead, 56% of US ad buyers said they would increase MMM focus in 2025 (IAB, December 2024 via eMarketer) and 49% of marketers worldwide already use MMM (Supermetrics, September 2024 via eMarketer). The cause is straightforward: deterministic signal collapsed under iOS, third-party cookies, and walled gardens, and the only causal-inference method that still works at scale is regression on aggregate spend.
Incrementality has moved with it. 71% of advertisers ranked incrementality as their top retail-media KPI in 2025 (DataSlayer, 2025), but only 52% are actively running tests, leaving a wide aspiration-versus-execution gap. Northbeam's 2024 guide to incrementality argues this is the only measurement layer that establishes causation rather than correlation, which is exactly what attribution cannot do once signal dropout is structural (Northbeam, 2024).
Geo-holdout testing is the cheap way in. Lifesight's 2025 methodology runs matched-market pairs over 4 to 8 weeks with 6+ months of historical baseline and a target 80% statistical power (Lifesight, 2025). The cost is opportunity, not software. The expensive way in is full MMM, which most teams budget against media spend rather than as a fixed tooling cost; 36.2% of marketers plan to invest more in incrementality methodology over the next year (Lifesight, 2025). Synthetic-control modelling, which builds a composite holdout from multiple regions, materially improves reliability over single-region tests.
The right sequence for a team scaling past $1m monthly spend: run a geo-holdout on the largest channel that attribution gives credit to (usually paid social or branded search), expect 4 to 8 weeks before the result is statistically clean, then commission an MMM that uses the holdout as a calibration anchor. The trade-off the research keeps surfacing is rigour versus speed: incrementality tests are slow but causal; MMM iterates faster with less certainty. Budget for both.