Before You Deduplicate: Universe, Coverage, and Person Identity
A cross-screen reach number is a claim about people, but most of the raw evidence arrives as device, household, or platform records. This chapter follows that claim through its two hidden boundaries: the population a measurement is meant to represent and the evidence used to decide that records from different screens belong to the same person. The practical judgment is simple but demanding: before trusting deduplication, check the universe, the coverage gaps, the unit being counted, and the proof behind each identity link.
The denominator comes before the identity graph
Reach is often presented as a percentage, which makes the denominator look like a piece of arithmetic. It is more consequential than that. A cross-media reach percentage describes unique people with qualifying exposures relative to a measured population, universe, or target. If the population changes, the meaning of the percentage changes even when the observed exposure records remain exactly the same.
The MRC Cross-Media Audience Measurement Standards require the cross-media universe to be appropriate and disclosed. At minimum, it is the deduplicated total of persons in the media universe for each medium, or more generally total persons. A component report can use a narrower population while it is measuring that component: an internet population for digital-only activity, a mobile population for mobile-only activity, or a television population for return-path or set-top-box data. But that component population cannot simply become the denominator for a cross-media total. It must be projected to the broader population with access or ability to access taken into account.
This rule prevents a subtle comparison error. Suppose a digital platform reports reach among people it can observe online, while a television dataset represents households connected to a particular provider. The two reports may each be internally consistent. Their percentages are not automatically comparable because their universes differ. A combined number that uses the digital denominator for one component and the television denominator for another has no single population interpretation, even if the report presents one clean percentage.
Consider a hypothetical campaign in which a platform reaches 20% of the people in its observable digital population. That may be a meaningful platform result. If the observable population excludes people without the relevant access, devices, or measurement permission, the same 20% cannot be read as 20% of the total cross-media population without a supported projection. The issue is not that the platform must observe every person directly. It is that the bridge from its observed population to the intended universe must be defined, tested, and disclosed. Otherwise, the report may compare a rate inside a window with a rate across the whole population while calling both reach.
The universe also defines what “missing” means. If the intended population includes people who use an unsupported device or a service outside a data partner’s footprint, their absence from the raw records is not evidence that they were unexposed. It is a coverage limitation. MRC guidance requires measurement organizations to describe material exclusions and quantify them where they are known and measurable. Estimates should be projected to the intended universe, not quietly to an adjusted universe that happens to match the available data.
This is why “census-like” does not mean complete. A large transactional dataset may capture enormous numbers of exposure or tuning events, but still omit particular MVPDs, over-the-air households, Smart TV types, websites that do not carry the relevant measurement asset, unsupported browsers, or environments where measurement is restricted for privacy or technical reasons. The scale of the stream reduces some sampling problems; it does not remove systematic gaps. A large, biased window can still produce a biased view of the audience.
Coverage has to be monitored as the environment changes. OTT devices can enter and leave households, new operating systems and device models appear, and different sources may feed linear, streaming, and application activity into the same reporting product. A measurement service therefore needs to identify device type, platform, and operating system, understand which environments are excluded, and update universe and coverage assessments as the mix changes. Without that maintenance, an identity method may be accurate for the records it receives while the reported audience drifts away from the population it claims to represent.
The difference between sample-based and large-scale collection is useful here. A panel or sample projects observed activity to a population and must demonstrate representation, weighting quality, standard errors, and the effects of bias. A large-scale collection tracks many accesses directly but must disclose what it cannot see and how its adjustments work. Neither approach gets a free pass from coverage analysis. The evidence needed differs; the obligation to define the boundary does not.
An analyst can therefore ask two separate questions about a reach percentage. First, what people are in the intended universe? Second, what fraction of that universe could the measurement actually observe or project? The first is a definition. The second is a coverage claim. Blending them makes the denominator appear more stable than it is.
From records to people
Once the universe is clear, the measurement system still has to decide what its records represent. A device is not automatically a person. A browser is not automatically a device. A household is not automatically one viewer. The same person may use a television, phone, tablet, and laptop; a household may contain several people sharing one connected television; and one device may be used by different people over the reporting period. The word “unique” hides these distinctions unless the unit is stated.
The MRC standard prefers unique-user attribution for combined and deduplicated cross-media video where it is possible, while recognizing that large passive datasets may begin at the device level. If a service reports at the person level, its underlying identification and attribution process must be at least partly traceable to information obtained directly from people. The standard rejects a purely algorithmic leap from a browser, device, or other non-human element to a person. Device-level observations can be useful inputs, but they are not self-authenticating person records.
This creates an interface between exposure collection and identity. The collection layer may know that a qualifying exposure occurred on a particular device at a particular time. The identity layer tries to connect that device to a person, household, or panel participant. The deduplication layer then tries to recognize that person when a separate platform reports another exposure. Each interface can introduce uncertainty. A technically precise exposure record does not become a precise person record merely because it enters an identity database.
Direct person evidence provides the strongest assurance, especially when people have confirmed their details or demographics and their media activity can be attributed through ongoing credentials or compliance mechanisms. The MRC standard acknowledges that direct person-level measurement at full scale can be difficult and costly, so it permits other approaches with empirical support and clear limitations. That permission is not the same as a conclusion that all approaches are equally accurate. The more an estimate depends on an overlay, inference, or modeled assignment, the more important its validation and error disclosure become.
Cross-device deduplication is the most visible version of the problem. To combine a television exposure with an OTT or mobile exposure, a measurement organization must establish that the records belong to the same person within the relevant reporting period. The data may come from different businesses, hardware types, operating systems, and collection rules. The standard says deduplication methods should be supported by research and actual consumer duplication information gathered as a learning or truth set. A model trained only on its own assumptions cannot, by itself, demonstrate that its links reflect real duplication behavior.
This requirement changes how an identity graph should be interpreted. It is not enough to ask whether a system can connect records. Ask how the connection was validated, whether the validation covers the media types and device combinations being measured, how rates differ by device or media, and how often the evidence is refreshed. Deduplication patterns can change as people add devices, switch services, share screens, or alter privacy settings. A stable-looking identity score can conceal a moving underlying population.
A truth set also has to match the claim being made. Evidence that links a panelist's phone to a laptop may say little about linking a shared television to that same person. Evidence collected in one market or for one device mix may not transfer unchanged to another. The point is not to demand a perfect census of identity relationships; it is to prevent a narrow validation sample from being presented as proof for every cross-screen combination. The closer the production estimate comes to the conditions represented in the truth set, the more relevant that evidence is. Where the conditions differ, the difference belongs in the limitation and uncertainty statement.
Some convenient signals are also weak foundations. The MRC standard specifically identifies IP address as insufficient in quality, granularity, or stability to serve as the basis for audience uniques by itself. An address can be shared, reassigned, or associated with multiple people and devices. It may contribute context in a broader method, but it cannot settle the person-level question alone. The same caution applies to any identifier whose apparent precision exceeds its actual stability.
Privacy makes the interface more constrained, not less important. MRC requires clear privacy notices, permission and opt-in practices where people voluntarily disclose data, protection of identifying information, and attention to rules such as the GDPR and local requirements. Privacy restrictions may remove or anonymize fields that would otherwise help link records. The standard recognizes that a universal person identifier would be useful but difficult in the current environment, and it encourages study of methods that preserve privacy while supporting accurate measurement.
That trade-off is often described incorrectly as a choice between perfect matching and no measurement. In practice, the system can preserve fewer direct identity details while retaining aggregate evidence about overlap. The subset-reach research paper starts from this kind of constraint: privacy and computation can make a complete reach Venn diagram infeasible, even when aggregate subset reaches are available. An absent person-level join is therefore not proof of zero duplication. It is a reason to represent the remaining overlap as uncertain and use evidence that does not require exposing every individual relationship.
The entities and interfaces can be kept distinct:
The diagram is deliberately not an identity solution. It shows why a unified audience depends on several different objects: the universe says whom the estimate is about; platform data record exposures; devices generate or carry those records; people are the intended unit of unique reporting; linkage evidence supports, but does not erase, the connection between them; and privacy rules constrain which evidence can be collected or retained.
A defensible deduplicated reach claim must keep those roles visible. It should say whether its unique unit is a person, household, or device; identify the intended universe and coverage exclusions; explain what direct or modeled evidence supports linkage; and disclose the limitations introduced by privacy, missing data, or changing device behavior. If the system cannot observe every overlap directly, it should not convert that absence into certainty. It should carry the uncertainty forward.
The central question is therefore no longer whether an identity graph exists. It is whether the claimed link matches the reporting unit, the universe, the media behavior, and the evidence available to validate it. That distinction will matter when several platform unions are observed but the full overlap structure is not.