A useful media mix does not rank channels only by observed ROAS.
A useful media mix does not rank channels only by observed ROAS. It distinguishes demand creation, intent capture, conversion, retention and interactions between roles. Measures do not share the same horizon or counterfactual: search can appear immediately in journeys while video gradually changes search probability and future response. Comparing them in one window mechanically shifts budget to journey end.
Decision and method.
Allocate media by incremental contribution to expected role and decision horizon, explicitly reserving a result share for measured interactions. State audience, intermediate behaviour, expected delay and outcome metric for each activity; an activity without explicit mechanism is spend to diagnose. Match individual experiment, geo test, MMM, survey or journey data to role and available granularity, keeping incompatible estimands separate. Project direct, assisted and persistent effects without double counting; apply margin, media cost and uncertainty range to each allocation scenario. Compare the next budget increment, not historical average return, and reserve learning budget where uncertainty prevents robust choice.
Worked example.
Two plans each spend €80,000. Capture plan puts €60,000 in search, €10,000 video, €10,000 CRM. Role plan puts €30,000, €40,000, €10,000. Margin per incremental sale is €55. Capture: 1,100 search + 150 video + 50 assisted + 230 CRM = 1,530 sales; net contribution €4,150. Role plan: 900 + 600 + 350 + 250 = 2,100 sales; net contribution €35,500. Retain role plan only if its 350 assisted sales are confirmed causally; otherwise they are sensitivity, not certain credit.
Checks and limits.
Every activity has role, audience, delay and expected proof; direct, persistent and interaction effects are not double-counted; scenarios use the same margin and economic horizon; revise allocation by marginal return and uncertainty. Roles change with targeting, creative and spend. Brand effects are difficult to turn immediately into contribution; interaction observed in one sequence is not automatically transportable to another.
Resources and sources.
Download the media-role model. Related: assign paid-owned-earned roles, measure media interactions, compare next increments. Sources: Naik and Raman (2003); Stephen and Galak (2012); Lewis and Rao (2015).