The Contract That Holds the Measurement Stack Together
The cross-channel advertising measurement system now has a clear shape: a randomized experiment can define a causal question, MMM can estimate an aggregate channel contribution, and IMA can refine that contribution into daily campaign allocations. This checkpoint connects those layers around one practical judgment: before trusting a handoff, verify that the outcome, exposure, time window, hierarchy, and evidence status still mean what the receiving user thinks they mean.
The point is not to merge different methods into one score. It is to give each method a place in a chain of claims. A channel total can remain an anchor while a campaign decomposition adds detail. A randomized lift can remain a causal reference while a reporting system limits which segments can be judged. Shared semantics are the contract that says what has been preserved, what has been inferred, and what can no longer be known from the available output.
The same claim needs an identity as it changes scale
Consider one advertising activity moving through the system. At the causal layer, the question is whether treatment changed an outcome compared with a control condition. At the strategic modeling layer, the question is how a channel’s media activity contributed to an outcome over a weekly interval. At the tactical layer, IMA asks how that channel-level incremental contribution can be represented among campaigns on a daily scale. These are related questions, but their units and evidence are not interchangeable.
The first contract field is the outcome. “Sales,” “conversions,” “visits,” and “spend” are different quantities. So are total sales and incremental sales. If a weekly channel model estimates incremental sales but a campaign dashboard reports total conversions, an apparent agreement in direction does not mean the models are checking one another. A downstream user must know whether a value is the outcome itself, a model’s estimate of contribution, or an effect relative to an unexposed or control condition.
The second field is exposure. The IMA framework can use impressions, clicks, or spend as media series. Those inputs do not carry the same interpretation. An impression count measures delivered opportunity, spend measures investment, and clicks measure an observed response to exposure. A coefficient attached to one cannot be moved to another without changing the model’s question. When the series is adstock-transformed, the exposure also includes a rule for how earlier media persists into later periods.
The third field is time. MMM’s weekly contribution and IMA’s daily allocation are connected through temporal disaggregation. The daily target is produced using adstock parameters converted from the weekly model and tuned where necessary for daily dynamics. It is therefore not a direct daily observation of incremental sales. It is a daily representation of a weekly modeled quantity under a carryover assumption. The contract must record both the calendar resolution and the memory of prior exposure.
The fourth field is hierarchy. A channel contains campaigns, but a reporting segment may cut across channels or campaigns. A geographic or customer-history partition may contain only part of a campaign’s audience. The unit randomized in a lift experiment may also differ from the unit used to record media.