In AdExchanger last week I argued that the open web is shifting from an audience economy to a context economy. Audience systems explain who. Context systems explain what matters right now.
This is the engineering follow-up. If agents are going to act on context, they need more than faster access to the old segments. An agent can hold who, where, when, what, and why at the same time, which is more than any human planning interface could carry. Space and time are components of the decision, and today they are not machine-readable in any consistent way at the moment the decision is made.
The blocker is that no two systems agree on the unit.
The problem the unit solves
A CTV ad airs Wednesday at 8pm Eastern. A billboard shows the same creative at 8:15pm. A web conversion lands at 8:30pm. These three events only compose into one analysis if they resolve to the same place and time buckets. Today each channel buckets place and time its own way, so composition requires a custom integration for every pair of partners.
Identity is one axis of that composition, and it answers person-level questions such as frequency management and suppression. This unit is orthogonal to identity, and it complements identity rather than substituting for it. Where and when exist on every impression whether or not an identifier does, and every operator sees the same geographic units: every publisher, every advertiser, every measurement vendor, every channel including CTV, out-of-home, audio, and in-store outside the browser. The unit is consistent across the ecosystem by construction, and composition on it does not have to wait for identity bridging.
The two dimensions compose. An identity-keyed record gains a place and time dimension, and an impression with no identifier still resolves to a bucket. An agent or an embedding model can hold both on the same record and dissect them independently, which is what agentic decisioning requires: more components of the decision, each machine-readable on its own.
The definition
Why the unit must be hierarchical
Digital devices do not report location at one accuracy. A GPS fix and an IP lookup can put the same impression anywhere from 5 meters to a whole metro. A fixed-resolution grid forces a choice between discarding the coarse signals and pretending they are precise.
A hierarchical unit resolves each signal at the resolution its accuracy supports. A GPS-derived impression lands at R8. An IP-derived impression lands at R5. Parent and child rollup makes the two composable at the coarsest resolution the analysis requires, and the declared resolution travels with the data so no consumer mistakes one for the other.
This is also how the unit manages signal versus noise. Location noise is acceptable as long as it is equally applied across the buckets being compared. What breaks measurement is noise that differs between the things being measured. Uniform cells and declared resolutions keep the error structure even, and finding the resolution where signal survives noise is an empirical question. It requires testing and modeling at multiple resolutions of the same hierarchy. A hierarchical standard makes that testing routine.
What standardizing the unit enables
- Consistent bucketing across campaigns, operators, and measurement paradigms, at the resolution each signal honestly supports. Observational attribution, causal geo experiments, marketing mix models, and retail trade-area analysis all read the same units.
- Every channel resolves to the same buckets. CTV, digital, out-of-home, audio, and in-store impressions and conversions become composable without a per-partner integration.
- Buy-time and measurement-time share one key. If the cell is in or derived from the bid request, an agent can manage reach and frequency during the buy on the same unit that measures the outcome after it.
- Intelligence published on the unit is portable. A demand baseline or an experiment design keyed to (cell, week, hour) works in any workflow that uses the unit.
The unit belongs to no protocol
The definition above carries no protocol branding on purpose. The same bucket works as an attribution histogram index, an audience-signal payload, a bid-request field, and an experiment randomization unit. I proposed it in June 2026 on the W3C Private Advertising Technology Working Group public list, in the context of attribution measurement; the archived thread is public. Once the unit is standardized, the inputs and the experimental designs that consume it can remain open. The unit is a convention, and any system can adopt it without adopting anything else.
Where Ether Data stands
Ether Data publishes its place and time model on this unit today: a demand baseline for any American neighborhood at any hour of the week, built from public structural data at H3 R8. The model reads from the unit. The unit stands on its own.
Agents need to do more. Space and time are components of the decision, and they need to be inserted so AI workflows extend the intelligence layer instead of replicating the old one. Standardizing the unit is how they get inserted.
