Most MMM teams don’t have enough weeks to estimate channel and subchannel effects cleanly. Nested “second-stage” regressions promise subchannel ROI anyway—by allocating a fixed channel story across its parts. The result is often precise-looking splits that are driven by spend share, not incremental impact.

Targeting engines can look “accurate” while mostly identifying people who would have bought anyway. Fit metrics and lift curves don’t answer the real question: what did the model change? The only decision-grade proof is incrementality—measured with a true holdout lift test.

CLV rarely fails because teams can’t build a model. It fails when one score quietly does two jobs—ranking customers for targeting and forecasting dollars for planning. The score earns trust under the easier standard (lift curves and deciles), then gets treated as planning-grade truth. When incentives and mix shift, rankings can still look right while dollars move.

Multiplicative MMM is often sold as “more realistic” because “everything interacts.” But structure is a commitment: it determines what kinds of counterfactuals the model is allowed to generate. When interactions are assumed universally, models can confuse overlap and frequency for synergy, inflate ROI through cross-channel scaling, and produce conclusions that aren’t decision-safe.

Adstock isn’t a technical setting. It’s a business assumption about when advertising impact is allowed to show up. In MMM, that timing story can change whether a channel clears your ROI hurdle rate—even if performance doesn’t change. The risk is simple: the model can fit history using carryover assumptions that don’t match your purchase cycle or channel intent, and the budget decision happens before you can validate the story.

Perfect fit in MMM can be cheap to obtain—and expensive to trust. When models use outcome-adjacent inputs and flexible response functions, they can look verified while remaining unsafe for decisions.