Frequency capping across channels, and why it usually is not happening

Each platform can see its own delivery and nothing else. When it caps at four, it means four exposures within its own inventory, to the identifiers it recognises. It has no visibility of the same household being reached by a different seller an hour earlier.

The consequence compounds with the number of channels. Two channels overlap somewhat; seven overlap substantially, particularly when they are all targeting the same narrow audience — which is precisely what a well-briefed multi-channel plan does. The better your targeting, the worse the overlap.

None of this shows in a per-channel report. Each one reports its own frequency accurately, and the aggregate figure — the one a marketing director actually cares about — is not computed by anybody.

To deduplicate across sellers you need every channel writing exposures into one record, keyed on an identifier they can all supply. The difficulty is that they cannot all supply the same one. A streaming platform can offer a household; a billboard can offer a postcode sector and nothing finer; a podcast can offer a delivery log.

The workable design is multi-key with graceful degradation: an opaque household key where available, a hashed advertising identifier where one exists, a consent-based identifier where the chain supports it, and a postcode sector as the coarsest key that always works. Precision varies by channel and the record says which key it used.

That degradation is not a compromise, it is the requirement. Advertising identifiers are a declining asset and a spine that depended on them alone would lose accuracy on a schedule set by Apple and Google. Multi-key is what keeps the record useful as that happens.

A single record of who saw what across seven channels is exactly the kind of dataset that deserves scrutiny, and the design decision that matters is where identity resolution happens.

The defensible arrangement is that the exposure record itself never performs an identity join — it stores whichever key it was given and never resolves it back to a person or across key spaces. Resolution happens in one separate module, which is then the only place a reviewer has to read to understand what is joined to what, and the only place a retention rule has to be enforced.

In AdBuyMCP that module holds hashed advertising identifiers for at most ninety days, after which the hash is persistently stripped from the event and the record re-graded to a coarser key rather than deleted. Being exact about the current state: the spine is designed, schema-backed and unit-tested, and is fed only by the deterministic sandbox today, because a live report proves aggregate delivery rather than household exposure and the system refuses to mint identities from one.

A launch is the worst case and the best argument. Launch campaigns deliberately buy the same audience across several channels in a compressed window, which maximises overlap at exactly the moment somebody is going to report a combined reach figure to a board.

The second case is a narrow B2B audience. If your addressable market is fifty thousand people in three cities, seven channels targeting them will overlap heavily by construction, and the difference between the deduplicated and naive reach figures can be large enough to change a decision.

The case where it matters least is a broad national consumer campaign with genuinely different audiences per channel — which is rarer than media plans imply.