One Outcome, Several Measurement Scales
This chapter examines how the cross-channel advertising measurement system changes the meaning of an “outcome” as it moves between randomized experiments, weekly channel models, and daily campaign reports. The practical judgment is straightforward: before deciding whether two numbers agree, identify the causal question, the unit, the time window, and the evidence status each number represents.
That discipline matters because measurement scale is not just a display choice. Moving from a channel-week total to a campaign-day value changes the data, the assumptions, and the uncertainty. A shared semantic layer lets the system refine a quantity without silently replacing it with another one.
Incrementality is a counterfactual question, not a credit score
Incrementality asks what advertising caused beyond what would have happened without it. In a randomized experiment, some units receive treatment and comparable units do not. The observed difference in their outcomes estimates the difference between two potential results: the result for the same kind of unit with advertising and the result without it. Randomization is valuable because treatment assignment, rather than a marketer’s prior choice, creates the main comparison.
This is different from observing that people who saw an ad converted more often. Those people may already have been more likely to buy, or they may have differed in geography, timing, device, or prior intent. A conversion after an exposure is an observed sequence, not automatically a causal effect. A post-hoc attribution rule can assign credit to that exposure, but it does not by itself reconstruct the missing alternative outcome.
The distinction becomes clearer when the three measurement approaches are placed side by side. MMM estimates how aggregated media activity is associated with changes in an outcome over a longer horizon, while accounting for other modeled factors. Its practical output is a channel-level contribution useful for budget planning. MTA uses user-level journeys to distribute credit among digital touchpoints, often to provide rapid optimization feedback. IMA uses a different bridge: it starts from the channel-level incremental contribution estimated by MMM and models how that contribution can be allocated among campaigns within the channel.
Those outputs can inform one another without being identical. An MMM contribution is not the sum of all user-level touchpoint credits unless the definitions and accounting rules make it so. An IMA campaign contribution is not a newly observed causal outcome simply because it is estimated from campaign media. It is a granular model output anchored to a channel-level target. A randomized lift estimate is stronger evidence for a causal effect under its assumptions, but it may be available only for a particular experiment, population, outcome, and follow-up window.
This is why a common label such as “campaign performance” is too weak for cross-channel decisions. It might mean incremental sales, predicted conversions, assigned touchpoint credit, or a share of a modeled channel total. The first semantic question is not whether the number is high. It is: high relative to which counterfactual, for which unit, and over which period?
The denominator matters as much as the numerator. A conversion lift for all randomized units is not automatically a lift for a customer-history segment, a device group, or a campaign audience. The privacy-robust framework makes this explicit by defining incrementality for a segment as the difference between its treatment and control potential-outcome means. That definition ties the effect to a population and a reporting boundary. If the segment changes, or if the reporting system retains outcomes unevenly across treatment and control, the meaning of the difference changes with it. A shared semantic record must therefore preserve not only what outcome was counted, but whose outcome it was and which units were eligible to appear in the comparison.
Changing scale changes the object
The first field to stabilize is the outcome. The source material uses several possible outcomes, including sales, conversions, visits, spend, and advertiser value. These are not interchangeable. A conversion count may be the reported outcome, while incremental sales is the target of a downstream allocation. A useful measurement record states whether it describes a raw result, a modeled contribution, or an effect relative to a control condition.
The second field is exposure. Marketing activity may be represented by spend, impressions, or clicks. Each is a different signal of what happened. Even within one exposure measure, timing matters: media effects can persist after the original impression or click. IMA therefore uses adstock-transformed media series when modeling campaign contributions. The transformation retains current and earlier exposure with weights determined by a window length, a peak or delay, and a decay parameter. It is an attempt to represent carryover, not a claim that every channel has the same response curve.
This makes a calendar day an interval in the causal story rather than an isolated timestamp. Suppose a campaign runs on Monday and sales rise on Tuesday. A same-day report may record only Monday’s exposure and Monday’s conversions, but a carryover model may treat part of Tuesday’s response as related to Monday’s activity. Conversely, a long window may connect an outcome to exposures that are no longer visible in a narrow daily report. Without a declared window and carryover rule, two “daily” metrics can refer to different processes.
The third field is the unit and hierarchy. MMM works at an aggregate channel level in weekly intervals. IMA’s downstream layer works with daily campaign media within a channel and produces campaign-level incremental-sales contributions. A segment report may partition users, campaigns, geographies, devices, or time windows. These hierarchies create useful relationships, but they do not erase their differences. A campaign belongs to a channel according to a classification rule; a reporting segment may cut across several campaigns; and the unit randomized in an experiment may not be the same unit shown in a media table.
The fourth field is the reporting boundary. A clean experimental outcome, an observed media series, and a privacy-degraded arm mean do not carry the same amount of information. The privacy-robust incrementality framework describes a released report that may include arm means, counts, suppression indicators, and noise metadata. If the report retains only part of the outcome signal, the number reaching the analyst is not simply a noisier version of the clean value. It may be compatible with a range of clean values.
The fifth field is evidence status. A reader should be able to label a number as observed, modeled, allocated, or causally identified, and those labels should not be treated as synonyms. Observed means the reporting system recorded the quantity under its own rules. Modeled means a statistical procedure inferred it from data and assumptions. Allocated means a constrained model distributed an aggregate quantity among lower-level components. Causally identified means that, under a stated design and assumptions, the evidence identifies a treatment effect rather than merely a pattern.
These categories can overlap in a single pipeline, but the overlap must be explicit. IMA’s campaign-level output is modeled and allocated within a channel-level contribution. The channel-level contribution is itself a model output. A randomized experiment can provide the causal anchor for a lift estimate, while privacy reporting can limit which segment-level claims remain identifiable. A number does not become causal by moving farther down a data table.
The relationships in the diagram are conditional, not automatic. A randomized experiment can anchor a causal question, but it does not determine the definitions used by every media report. A channel measurement can constrain a campaign decomposition, but the decomposition still depends on the model and its assumptions. A segment report can be compared with either layer only when outcome, exposure, window, hierarchy, and reporting rules line up.
A semantic check for cross-channel claims
A useful way to test a measurement handoff is to ask five linked questions. What is the outcome? If the source value is incremental sales, do not compare it with total sales as if they were the same target. What is the exposure? If one report uses spend and another uses adstocked impressions, their coefficients describe different relationships. What is the unit? A channel total cannot be read as a campaign result without a defined aggregation path. What is the time window? A daily value with carryover is not a same-day-only effect. What is the evidence status? A modeled allocation should not be presented as an independently measured causal effect.
A sixth question becomes essential under privacy constraints: what portion of the intended signal reached the report? A conversion outside the attribution window, an unmatched outcome, a broken identity link, or a suppressed cell can change the set of effects compatible with the released number. If two channels have different retention or reporting quality, the same apparent conversion rate may carry different uncertainty. Shared semantics therefore includes the conditions under which a value was observed, not only the value’s name and unit.
Imagine a retailer comparing two campaign summaries. Campaign A has a higher daily contribution in an IMA report; Campaign B appears stronger in a randomized lift test. That is not yet a contradiction. The IMA result may allocate a channel-level incremental-sales target using daily, adstocked media. The lift test may measure conversion incrementality for a different population and follow-up window. One result may be a model-constrained allocation and the other a causal estimate. To compare them, the retailer would need compatible outcome definitions, campaign membership, exposure windows, experimental units, and reporting conditions. If those cannot be aligned, the right conclusion is not that one model “won,” but that the comparison was not yet well-defined.
The distinction also protects against an easy but damaging shortcut: treating successful reconciliation as proof of every lower-level claim. If campaign allocations sum to a weekly channel total, the aggregation rule is internally consistent. It does not show that the model has uniquely discovered each campaign’s causal effect. The source paper’s IMA evaluation is framed at the channel level for this reason: daily predictions are aggregated to weekly totals and compared with the MMM-derived channel contribution. That check answers whether the downstream layer preserves its anchor, not whether all causal uncertainty has disappeared.
Conversely, failure to obtain a clean comparison does not prove that a method is wrong. A mismatch may indicate a different outcome, a different window, a missing carryover definition, or a privacy boundary. The evidence analyst’s task is to locate the semantic break before attributing the difference to model quality. This is the difference between evidence of disagreement and evidence that one method is invalid.
The most important invariant is therefore not a single number. It is the meaning of the question as it moves through the system. The outcome should remain identifiable as the same business quantity or be explicitly transformed. The exposure should retain its definition and carryover rule. The hierarchy should explain how channels, campaigns, and segments relate. The time window should travel with the estimate. The evidence status should remain visible. Under privacy constraints, retention, suppression, and noise metadata must travel with the report as well.
With those fields in place, a daily campaign value can be useful without pretending to be a raw observation, and a weekly channel value can remain a meaningful anchor without pretending to answer every tactical question. The next task is to inspect the mechanism that performs this handoff: how IMA uses temporal disaggregation, adstock, and channel-specific Bayesian models to add detail while preserving the aggregate reference.