Marketers of Katan: Digital Measurement Can Be Solved With a New Toolkit
Identity measurement is probabilistic on a sample nobody randomized. Place and time are representative by construction: canonical geography, canonical time, one unit for both sides of the transaction.
Nate Woodman ยท July 24, 2026
Anyone who has played Settlers of Catan knows the feeling: the board is a grid of hexagons, the dice decide which hexes pay out, and you win by understanding the board better than the people across the table. You do not win by knowing your opponents' names.
Digital measurement is in the middle of an identity crisis, and I mean that literally. The identifiers the industry built its measurement on are structurally decaying: cookies are gone, and every platform release restricts mobile identifiers a little further. The response has been an identity repair kit: clean rooms to match what remains, authenticated traffic to re-collect identity at the login, ID graphs and UID frameworks to stitch the fragments back into something portable.
I have built with every tool in that kit, and I want to be precise about what it is: a partial answer, and a probabilistic one. The pitch for identity-based measurement has always been determinism: exact matches of known people, no modeling required. But look at what the matches cover. A clean room sees the overlap of two parties' identified users. Authenticated traffic sees the slice that logged in. An ID graph sees whoever it managed to stitch. Each tool sees a sample, none of the samples is random, and the population you sell to includes everyone else. The moment you project a matched-set result onto the full market, you are running a probabilistic model on a sample nobody randomized, carrying selection bias the system cannot see from inside. The matching is deterministic. The measurement never was.
There is a harder version of that point, and Chris Williams has been pushing me to it in a long-running measurement thread: attribution is probabilistic inference all the way down. Outcomes move with price and weather and with every other channel at the same time, so no user-level match, however exact, isolates what caused a sale. "Deterministic attribution" described the join, never the inference. Multi-touch attribution is better understood as multi-touch analysis, one input into a model, and the model is where you estimate causality.
The selection bias also runs in a known direction, which turns a methodological complaint into an operating problem. Digital touchpoints are the ones carrying identifiers, so the media that can be measured is the media that gets credited, and digital collects credit earned by channels no identifier can see. Attribution overstates digital, and a buying agent optimizing against that overstatement moves budget the wrong way at machine speed. A human planner who distrusts a number discounts it. An agent takes the number at face value and optimizes.
Here is the reframe I would offer the industry: stop trying to reassemble "who" and start structuring "where" and "when." Place and time have a property no identity system can offer. They are representative by construction. Every impression, every store visit, every conversion happens somewhere, at some hour, whether or not anyone logged in. A cell-and-hour frame is a census, not a panel. The measurement you run on top of that frame is openly probabilistic, and I will take openly probabilistic on the full population over nominally deterministic on a biased slice.
I am not the only one making that trade. Chris reduces the same reframe to a slide that reads geography over identity, and he is arguing that case in front of the joint industry committees that measure television and out-of-home.
The other thing that happened while identity was unraveling is that AI agents arrived, and agents change what measurement can use. An agent can reason across more dimensions than any planner or spreadsheet ever could. It can hold place structure, temporal pattern, weather, events, commerce flows, and creative context in one decision. But only if we feed the agent two things: the data, and the structure. Data without structure is noise to an agent, the same way a Catan board with no hexes is just a pile of resource cards.
The board: canonical geography
The structure for "where" already exists. Hexagonal cell grids (H3 is the reference) turn geography into a canonical, composable unit: every place on earth gets a cell, cells nest across resolutions, and two parties naming the same cell mean exactly the same ground. The standards work is already underway. The Ad Context Protocol community is standardizing a geo_cells unit right now, in public, on GitHub, and I am in that thread arguing for two requirements that make the standard trustworthy rather than merely convenient.
First, a declared resolution is a joint claim. When a seller says it supports a given cell grain, that one number carries two statements at once: I can deliver at this grain, and I have assessed that delivering at this grain is safe in the geographies I serve. A 460-meter hex is genuinely low risk. A one-meter hex in a sparse rural area re-identifies almost like a raw coordinate. Precision and privacy pull in opposite directions on the same knob, which is exactly why the declaration has to carry both, and why a fine-grained claim is a privacy posture, not a spec sheet line.
Second, conversions must carry provenance. The moment a request in cells gets converted, to a radius, to a zip crosswalk, to an admin boundary, the effective grain changes, and both halves of the declared claim can silently break. The fix is boring and essential: report requested versus executed geography on every conversion. Declarations you cannot verify are marketing.
The objection I hear from measurement veterans is operational, and it is the right question: platforms do not speak hexagon. Spotify and Hulu take postal codes. YouTube and Meta take a point and a radius. The Trade Desk takes polygons on some inventory types. This objection is exactly why the grid works as a standard rather than as another proprietary scheme. H3 converts into each of those shapes and back out, the conversion tooling is open source and supported in tools like BigQuery, and we have published datasets that normalize census tracts, postal codes, and counties onto the grid with population-weighted mapping. The grid is a substrate under the encodings platforms already use, not a replacement for them. Adoption can start today, before any platform goes native.
The dice: canonical time
Geography is getting its standards moment. Time has not even started, and time is half the board. The industry still runs on blunt instruments: dayparts and holiday flags, weather averaged across a whole city. Those features do not vary by place, which means they cannot interact with place, which means the models never see that a stadium hex and a residential hex live in different temporal worlds on game night.
Broadcast makes the failure concrete, and Chris calls broadcast the weak link in cross-media measurement. Campaign exposure data can arrive months after the flight, and when the delivery detail is thin the supplier averages that exposure across every week the campaign ran. A model fed that average cannot see a Tuesday. No amount of geographic precision repairs a flattened calendar.
The structure that works is canonical time states per cell: a small taxonomy of temporal conditions, defined at the same grain as the geography, so that "where" and "when" compose. We have built and validated that taxonomy. It is not exotic, just the same normalization discipline the industry finally accepted for geography, applied to the clock and the calendar.
Put the axes together and you get the unit I have been arguing for in the standards threads: an H3 cell at declared resolution, crossed with the ISO week and the hour in UTC. A week holds 168 hours, which is where the name of our public walkthrough comes from. The 168 Grid explainer, from my work with Ether Data, shows the full proposal in one place: the grid itself, the crosswalks to postal codes and DMAs, the activation paths per platform, and the privacy math that makes a declared resolution a compliance posture. If the essay is the argument, the walkthrough is the diagram.
One unit then serves both sides of the transaction. A CTV spot airing Wednesday at 8pm Eastern, a billboard impression at 8:15, and a web conversion at 8:30 compose across channels only if they resolve to the same cell-and-hour bucket. A buying agent can use the bucket at bid time, deciding how many impressions each cell needs to hit a reach goal and managing frequency during the buy instead of auditing frequency after. Measurement reads the same bucket back out, so the dimension you targeted is the dimension you evaluate.
Chris made the strongest commercial case for the unit: a cell has a known population, so incremental reach inside a buy becomes a bounded number instead of an assumption stacked on de-duplication guesses. One caveat, and time makes it obvious: the denominator is not constant. An airport cell holds a few hundred people at 03:00 and tens of thousands at 14:00, so the population model has to declare its time grain too. That caveat is the argument in miniature. Where and when only work together.
Why orthogonal matters
Here is the part I find genuinely beautiful, and it is an empirical result, not a design preference. When you normalize both axes properly, they turn out to be close to orthogonal. In our validation across roughly 900,000 observation rows, the place prior held in every time stratum we cut, business hours or midnight, summer or winter; the prior never collapsed. The time prior split cleanly into a universal component that applies everywhere and an event component that concentrates where places have gravity. And the two priors were additive to within roughly a percent: each prior contributes its explanation independently.
Orthogonal axes are what make measurement composable. You can decompose an outcome into place structure and moment without double counting, and you can transport a model across markets because the axes mean the same thing everywhere. You can hand an agent a decision space instead of a reconciliation project. Orthogonality is what the toolkit buys, and nothing in that toolkit requires knowing anyone's name.
Marketing mix models already run on geo structure. Google's Meridian works at geo level, and Google's guidance warns that different geo groupings produce different results. When the grouping changes the answer, standardizing the grouping is the fix.
Chris would add a third normalization, and he is right to. An impression means something different in every medium, so exposure needs normalizing too before a model can treat a billboard and a pre-roll as the same kind of evidence. That work composes with the cell-and-hour unit rather than competing with it.
What the industry should do
The list is short. Adopt cell-grid geography with declared resolution in the transaction protocols, and align on the geo_cells unit already in progress rather than inventing another. Treat declared resolution as a joint precision and privacy claim, and name the safe band. Make requested-versus-executed provenance mandatory wherever geography converts. Give time the same canonical treatment geography is getting, per-cell states, published taxonomy, a standards home. And validate orthogonality empirically before composing axes, because the whole point is measurement you can defend.
The unit also travels, and nothing else on offer does. Canada's census metros, the UK's broadcaster regions, and Australia's metro definitions are mutually incompatible, and none of them crosses a border. DMAs stop at the American shoreline. A hexagonal grid means the same ground in every market, which for a multinational advertiser may be the strongest adoption argument.
The cross-media prize makes the standard bigger than digital. Panels, out-of-home, television, the markets the joint industry committees measure: none of them will ever be ID-addressable, and all of them happen at a place and a time. Out-of-home is the clearest case. Nobody will ever attach a device identifier to a billboard audience, so the count has to come from a population model built on census-grade geography, and that model needs the cell-and-hour unit or its numbers will not join anything else. Those same committees now have to work out how to govern buying agents, and a standard unit gives them something to write rules about. A cell-and-state standard is a board every medium can play on. The argument is running in public, in the Ad Context Protocol GitHub threads and in the W3C attribution discussions, among measurement scientists and protocol engineers who push back hard on the operational details. The proposal has gotten sharper every round.
Identity had a good run, and where it survives, in authenticated relationships, in first-party data, it still matters. But the industry has spent a decade trying to rebuild a deterministic "who" in a world that keeps taking it away. The agents are here. Feed them the board and the dice, the where and the when, structured and normalized, and measurement stops being an archaeology project and becomes an engineering discipline.
You can win Catan without knowing who you are playing against. You cannot win without understanding the board.
Thank you to Chris Williams, whose arguments in the working thread and in his joint industry committee work shaped the middle of this piece, and who was the first to share it on LinkedIn.
About me
I am Nate Woodman. I have spent twenty years building through ad tech's platform shifts: one of the first agency trading desks, people-based identity as Chief Data Officer at Havas Media, incrementality experiments I started at IPONWEB so advertisers could optimize to the KPIs that drive their business, and now agent-operated go-to-market systems.
I am available for Fractional CBO and CPO engagements and for agentic GTM implementation, and I am open to the right full-time senior role. The fastest way to reach me is to book time directly. The longer story is at proofindata.com/work-with-me.