The Identity Graph You Use for Measurement Was Built to Optimize for Different Outcomes Than the One You Used for Targeting

Most campaign reviews frame the question simply: did we reach the audience we planned for, and did that audience respond? The comparison feels straightforward. You targeted a defined population, you received a measurement report, and the two should speak to the same people. The problem is that the graph used to build your targeting audience and the graph used to attribute your outcomes were almost certainly built by organizations optimizing for different things, and those different objectives produce structurally different populations even when the underlying customer data starts from the same CRM file.

What Graph Construction Objectives Actually Mean in Practice

An identity graph built primarily to support targeting activation is optimized to maximize addressable reach. Its commercial incentive is to find as many device and email and cookie-level identifiers as possible for a given person, because a wider graph means more biddable inventory. Probabilistic linkages that increase addressable scale are acceptable costs when the outcome being measured is reach volume.

A graph built primarily to support measurement and attribution has a different objective. It needs to be conservative about false linkages, because a spurious device-to-person connection that inflates reach in a targeting context becomes a spurious impression-to-conversion connection in a measurement context. Measurement graphs tend to favor precision over recall, accepting that some real people will not be resolved if the alternative is linking two different people together and misassigning credit.

These are not the same graph, even if they share a brand name or a parent company. They may use the same seed data, but the scoring thresholds, the linkage rules, and the confidence cutoffs are tuned differently because the cost of an error is different in each context.

How the Divergence Surfaces in a Real Campaign

Consider a hypothetical campaign scenario. A retail brand onboards a CRM file of recent purchasers. The targeting graph resolves that file to a population of, say, 800,000 addressable identifiers. The campaign runs, impressions are delivered, and at the end of the flight the measurement vendor runs an outcome analysis.

The measurement vendor uses its own graph, tuned for attribution precision. It resolves the same CRM file to 550,000 identifiers. It then matches delivered impressions to those identifiers and finds that 60 percent of the delivered impressions touched someone in the resolved measurement population.

The buyer sees that number and often interprets it as an accuracy problem: 40 percent of impressions did not reach the intended audience. But that framing misses the structural cause. Some share of those unmatched impressions likely did reach real people from the original CRM file. The measurement graph simply could not confirm it confidently enough to include those people in its resolved population. The gap is partly a delivery problem and partly a graph construction problem, and the two cannot be separated without knowing both graphs' linkage rules.

Why This Makes Audience Quality Evaluation Harder

When a media buyer tries to evaluate whether a campaign reached a high-quality, relevant audience, they are usually comparing targeting inputs to measurement outputs. If those two things resolve identity differently, the quality evaluation is always measuring something partially outside of its own control.

This creates a practical challenge for incrementality testing as well. If your holdout group is identified through a measurement graph and your treatment group was targeted through a different graph, the two groups are not necessarily comparable populations to begin with. Lift estimates built on that comparison carry structural uncertainty that cannot be resolved through statistical power alone.

Practical Steps Worth Taking Before a Campaign Launches

A few approaches can reduce the impact of graph divergence without requiring buyers to build their own resolution infrastructure.

First, ask both your activation partner and your measurement vendor to describe how their graphs weight precision versus recall, and in which direction they err when a linkage is uncertain. This is a question most vendors can answer, and the answers will tell you whether the two graphs are likely to produce similar resolved populations or divergent ones.

Second, where possible, run a pre-campaign graph alignment check. Some measurement vendors will accept your resolved targeting universe as an input and tell you what share of those identifiers they can confirm before the campaign runs. This gives you a baseline expectation rather than a post-campaign surprise.

Third, ask your measurement vendor how they treat impressions that touch identifiers outside their resolved population. Do they exclude those impressions entirely from outcome analysis? Do they attempt to probabilistically assign them? The answer changes how you should interpret reported reach and frequency numbers, and it changes how you should weight attribution outputs.

Fourth, if you are using a data clean room to connect impression logs to outcome data, confirm which identity graph the clean room uses for the join. Clean rooms do not automatically harmonize the graph used for targeting with the graph used for measurement. The join layer is a third graph decision, and it may differ from both of the others.

What to Communicate to Stakeholders

When campaign reports show gaps between planned reach and measured reach, the instinct is to assign blame: the DSP underdelivered, the targeting was too narrow, the publisher's inventory was off-target. Some of those explanations may be correct. But before that conversation happens, it is worth understanding how much of the gap is structural, meaning it would exist in any campaign simply because the two graphs involved were built to serve different objectives.

Framing the gap that way does not make it disappear, but it changes the remediation conversation. If the gap is primarily structural, the fix is graph alignment at the vendor level, not targeting adjustment. If the gap is primarily a delivery problem, targeting adjustments are appropriate. Without separating the two, campaign optimization decisions are responding to a mixture of signals that neither targeting nor measurement changes will fully address.

The useful habit is to treat graph construction objectives as a campaign input, something to ask about before launch, not a post-campaign variable to discover during the performance review.