What a finished MMM hands you

Reading the decomposition

The waterfall below is the canonical form: total revenue on the left, base peeled off, then each channel's incremental slice.¹ Read it with three checks. Does it sum to actual revenue? Do the intervals (in a Bayesian model) overlap so much that channel rankings are noise? And does the base share pass the smell test against what happens when you have paused marketing before? A decomposition that fails the third check is not wrong by definition, but it owes you an explanation.

Response curves into reallocation

Each channel's curve shows expected revenue across spend levels, bending through its saturation point. The reallocation logic is mechanical: read the marginal return at current spend for every channel, move budget from the flattest slopes to the steepest until slopes equalize.² The caveat that belongs in every readout: curves are best-estimated near spend levels the data has visited. A recommendation to triple a channel is an extrapolation wearing an optimizer's badge; step toward it, test, refresh.

What goes in the board deck

One slide, three elements. The decomposition for last period, with base share stated plainly. Channel incremental ROI at a printed contribution margin (see ROI vs ROAS for why the margin must be printed). And the single reallocation the model most confidently recommends, sized in dollars with its expected revenue range. Leave the other forty charts in the appendix; a board that wants the Hill parameters will ask.

The caveat sentence worth standardizing: "Estimates are model-based with stated uncertainty; the model is calibrated against N experiments and validated out of sample; allocation recommendations are reliable within observed spend ranges." If any clause in that sentence is not yet true of your model, that is your roadmap, in order: see validation and the vendor questions.