The Measurement Problem Behind Cross-Channel Growth
The subject of this series is the modern advertising measurement system: the chain that connects channel-level planning, campaign-level optimization, randomized incrementality, and privacy-constrained reporting. This opening chapter focuses on the problem that makes the chain difficult to trust: the same marketing activity can be represented by several credible-looking numbers that do not mean the same thing. The practical goal is to learn how to recognize those differences before comparing models or acting on a dashboard.
That problem is not merely technical. A marketing team may need one measurement system to answer two very different questions. Which channels deserve more budget over the next planning period? Which campaigns should be changed this week? And underneath both questions sits a stronger one: did the advertising cause outcomes that would not otherwise have occurred? Privacy restrictions add another layer, because the experiment or customer journey may be sound while the released report no longer preserves enough information for a fine-grained conclusion.
Strategic and tactical measurement ask different questions
Marketing Mix Modeling, usually called MMM, is designed for the strategic question. It works with aggregated data over longer time horizons and estimates how channels contribute to outcomes such as sales. Because it does not require a complete record of every person’s path across websites and devices, it can support channel-level planning across digital and offline media. Its strength is breadth and robustness at the aggregate level.
That same design becomes a weakness when the question moves down to individual campaigns. A model that tries to estimate a separate effect for a very large number of campaigns has more parameters, stronger correlations between campaigns, and fewer observations per parameter. Campaigns often run at the same time or belong to the same family, so their media series move together. Standard regression can then produce unstable coefficients, inflated variance, sign reversals, or overfitting. A weekly channel result cannot simply be relabeled as a daily campaign result.
Multi-Touch Attribution, or MTA, was built for a different purpose. It uses user-level journeys to assign credit to digital touchpoints and can return fast feedback for campaign optimization. But the very identity continuity it needs has become less dependable as third-party cookies disappear and cross-site tracking is restricted. When part of a journey is missing, attribution can become biased or operationally fragile. This does not mean every MTA implementation is useless. It means its conclusions depend on an observation system that may no longer see the journey it claims to measure.
Integrated Marketing Attribution, or IMA, addresses the gap between these two levels in a specific way. It starts from channel-level incremental contributions estimated by MMM, then uses aggregated data and channel-specific Bayesian models to produce campaign-level effects. Its purpose is not to pretend that weekly channel evidence and campaign-day evidence are identical. Its purpose is to add tactical detail while keeping the detailed allocation tied to an aggregate anchor.
Randomized incrementality supplies a related but distinct reference point. In a clean randomized experiment, treatment and control groups allow an analyst to estimate the difference between outcomes under advertising and outcomes without it, subject to the experiment’s assumptions. That causal question is stronger than assigning descriptive credit. Yet a randomized design does not guarantee that every downstream report can support every decision. If outcomes cannot be matched, identities cannot be linked, conversions fall outside a reporting window, small cells disappear, or noise overwhelms a segment, the released evidence may be too weak even though the original assignment was randomized.
The important conclusion is that MMM, IMA, MTA, and randomized experiments are not four competing answers to one already-defined metric. They operate at different resolutions and rely on different evidence. A trustworthy system must say how the outputs relate before it places them on the same page.
The diagram is a map of relationships, not a claim that one method automatically validates the next. The upper path shows how an aggregate channel anchor can be refined into campaign-level modeled effects. The lower path shows that the evidence used for a causal decision may pass through a reporting system that changes what is observable. The final three states are deliberately different: a result can clear a threshold, fall below it, or remain undecidable with the available report.
Why shared meaning matters more than a shared dashboard
“Shared data semantics” sounds abstract until the same campaign is followed across the measurement chain. Start with the outcome. Is the quantity total sales, conversions, visits, spend, or incremental sales? A campaign’s observed conversions are not automatically its incremental conversions. Next define the exposure: impressions, clicks, spend, or an adstock-transformed version of one of these. The choice matters because an exposure can affect sales after a delay rather than only on the day it occurs.
Then define the unit and hierarchy. A channel contains campaigns, a campaign may be observed across segments or geographies, and a randomized experiment may assign treatment to a different unit from the one used in a media report. “Campaign performance” is incomplete unless the reader knows which rows or people are counted and how campaigns are grouped. A channel-week total and a campaign-day contribution can be reconciled only if the relationship between their units is explicit.
Time has the same problem. IMA uses weekly MMM outputs but produces daily campaign-level attribution. It does not distribute a weekly contribution by a simple same-week proportion. It carries forward the adstock structure learned by MMM, converting and, where needed, tuning the parameters for daily media dynamics. The meaning of a daily contribution therefore includes the effect of earlier exposure. Remove that carryover definition and the number may still be computable, but it no longer describes the same temporal process.
Finally, attach the evidence status to the metric. A value may be directly observed, estimated by a model, allocated within an aggregate constraint, or supported by a randomized experiment. It may also be calculated from a privacy-degraded report whose clean value is only partially identified. These labels are not bureaucratic additions. They tell a decision-maker how far the number can travel.
Consider an illustrative planning conversation. One dashboard says that a search channel contributed strongly during a week. Another says that three campaigns within it had different daily contributions. A lift test reports a positive conversion difference. A privacy system reports only aggregated counts for some customer groups. Those statements can all be true, but they answer different questions. Before deciding which campaign to expand, the team must know whether “conversion” has the same window in each report, whether the campaign exposures include carryover, whether the lift is incremental or descriptive, and which segments were omitted or perturbed. Without those definitions, agreement in the headline number can conceal disagreement in the underlying object.
One system, three kinds of evidence
The measurement system can be understood as three linked layers. The first is the causal layer: a randomized experiment asks what changed because of treatment. The second is the modeling layer: MMM estimates channel-level contributions, and IMA uses those contributions to allocate detail among campaigns while preserving the channel total. The third is the reporting layer: privacy and identity constraints determine which parts of the experiment or media record reach the analyst.
The layers should be connected, but they should not be collapsed. IMA’s reported representative-channel fit, for example, shows that daily campaign predictions aggregated to weeks can track the weekly MMM contribution closely; the paper reports an R-squared of 0.98 for that channel. That is useful evidence of reconstruction consistency. It is not, by itself, an experiment proving that each campaign caused exactly its allocated amount. Likewise, a positive clean lift in a randomized dataset does not automatically become a certifiable segment decision after reporting noise and signal loss are included.
This distinction changes what “cross-channel measurement” should mean in practice. It should not mean forcing every channel, campaign, and segment into one universal score. It should mean maintaining a visible relationship among quantities: a channel-level anchor constrains a granular decomposition; a clean causal estimand is transformed into a report; and the report is tested against a business threshold. The relationship is more important than visual uniformity.
That is why the unresolved state matters from the beginning. If a report is compatible with both a clean effect above a meaningful threshold and a clean effect below it, the honest output is not a more confident point estimate. It is an unresolved claim. The privacy-robust framework formalizes this idea through bounds on the clean effects compatible with the released report. Later chapters will examine how those bounds are constructed and how cell size, reporting noise, and heterogeneous signal loss widen them.
The opening judgment is therefore simple but demanding: measurement integration starts before model integration. A shared definition of outcome, exposure, unit, time window, hierarchy, and evidence status is what allows one layer to be checked against another. It does not remove multicollinearity, privacy loss, or causal uncertainty. It makes those limits visible instead of allowing a familiar label—“attribution,” “lift,” or “channel contribution”—to hide them.
In the next chapter, the focus moves from the overall tension to the measurement objects themselves. The question is what must remain invariant when one advertising activity is seen as a weekly channel signal, a daily campaign series, and a randomized outcome. That vocabulary will make it possible to inspect IMA’s handoffs without confusing aggregate consistency with independent causal proof.