Measurement, on the Street's Terms
Eight readers asked this site's answer engine how to measure DOOH and place-based media; it declined all eight, so this is the answer, written down. A DOOH number is an estimation chain — one near-observed event (the play), one auditable claim (the venue), three models (crowd, audience, outcome) — and running the Instrumentation Test in reverse lands the punchline: the channel everyone calls unmeasurable is the one where reach and frequency keep their referent, because a crowd on a concourse can't be instrumented. The workable frequency answer is geographic (a never-redrawing cell grid with per-cell impression ceilings), proof-of-play stays a contract term until playout reporting standardizes, and measurement currencies must become named, versioned, machine-readable data or stay invisible to the agents allocating the next dollar.
In English, please
Digital billboards and screens in public places -- airports, gyms, malls -- can't be measured like websites. A website knows when one specific person clicks; a street screen just plays its ad to whoever happens to be walking past. Nobody logs in to a billboard.
So every audience number in this channel is an estimate. The screen's owner can log that the ad played. An industry list describes what kind of place the screen stands in. But the number of people who saw the play comes from a statistical model -- one play might count as 14 viewers in a busy concourse or a third of a viewer in a quiet hallway -- and the who-were-they layer comes from yet another model, run by measurement organizations.
The essay's twist: that honesty is now an advantage. Online advertising built its reputation on counting clicks exactly, and that exactness is crumbling as privacy rules cut off tracking. Street media never had a click to count, so it has been saying 'this is a careful estimate, and here is how we made it' for roughly ninety years. The web is drifting toward the street's way of measuring, not the other way around.
How do you avoid showing the same person an ad again and again if you can't recognize anyone? You manage places instead of people. Divide the map into small fixed tiles, estimate how many of your target customers pass through each tile at a given hour, and cap how many ad plays you buy there. No one is tracked; the neighborhood is budgeted.
Practical advice from the piece: buyers should demand the recipe behind every audience number and write 'prove my ad actually ran' into their contracts, because no industry-wide proof standard exists yet. And the organizations that produce audience estimates need to publish them in a form software can read, because the ad-buying programs that increasingly allocate budgets can't read a PDF.
On this page
Last week the answer engine on this site declined the same question eight times: how do you actually measure DOOH and place-based media? It declined because I had never written the answer down. Here it is. The short version: every DOOH audience number is a modeled estimate with a methodology attached, the channel has been honest about that since 1933, and web measurement is currently converging on DOOH’s condition rather than the reverse.
The chain, read honestly
A billboard doesn’t click. A screen in a gym or an airport concourse has no cookie, no session, no completion event, and no consent string, because a passer-by presents nothing to it. Buyers raised on web dashboards read that absence as a void and ask whether any DOOH number is real. That is the wrong question. The right question is what each number in the chain proves, because DOOH measurement is an estimation chain with five links, and the chain is completely legible once you stop asking it to impersonate the web.
The play is the nearest thing to an observed event. The media owner’s system logs the creative playing in its scheduled slot; on the programmatic path, OpenRTB 2.6 adds burl, the billing event that fires when the creative has rendered and the impression becomes billable. Note what that proves: the play happened and became billable. It says nothing about who saw it.
The venue is a declared claim. The OpenOOH Venue Taxonomy — version 1.2.1, finalized February 2026, a three-tier hierarchy with immutable IDs — standardizes what kind of place the screen stands in, so a pharmacy, a gym, and a transit concourse are different buys. A venue ID is seller-declared metadata a buyer can audit. It is not a sensor reading.
The crowd is a model. One play on a public screen reaches many people or nobody, so OpenRTB carries the qty impression multiplier: a statistically modeled count of impressions per play, sourced from a declared measurement vendor, with fractional values allowed — 0.32 is as legal as 14.2. Cost math runs through it: (AUCTION_PRICE / 1000) × AUCTION_MULTIPLIER. The multiplier is a vendor’s modeled opinion, carried in the bid request.
The audience is a currency. In the US that means Geopath, the tripartite measurement body founded in 1933, whose modeled, planning-grade audience estimates sit behind much of the country’s OOH inventory. A currency is an estimate the market has agreed to transact on, and this one is mid-modernization: an Ipsos-run pilot starts in late 2026, transition begins 2027, full adoption is anticipated 2028.
The outcome is a model on a model. Footfall and brand-lift studies read modeled exposure against an outcome, per each study’s methodology. Estimates stacked on estimates do not become observations.
That is the entire answer to “what does a DOOH impression mean”: one near-observed event, one auditable claim, three models. The IAB’s own DOOH Measurement Guide, published July 2025, says the landscape remains challenged by inconsistent standards and fragmented practices. All true. Nothing in the chain observes a person seeing an ad. And none of that is the channel’s weakness.
The Instrumentation Test, run in reverse
A month ago, the metric poll essay introduced the Instrumentation Test: a metric survives agentic decisioning only if the event it measures remains unobservable in the system now making the decision, because inference has no work left to do once you can just read the record. On the web, that test is a killing floor. Once the buyer is an agent, its decision path is logged and replayable, so every metric built to infer a hidden human state — attention, reach and frequency, attribution — loses the thing it was inferring.
Now point the same test at a screen. The buying side is going agentic there too; an agent can assemble a screen plan in seconds. But the audience side of a DOOH impression is still a crowd of human heads on a concourse. There is no log of who looked up, no record to replay, nothing to instrument. The event reach and frequency were invented to estimate remains unobservable by physics, not by policy. So the inference still has work to do — in DOOH, uniquely, the classic audience metrics keep their referent.
Sit with the reversal for a second, because it is the punchline of the whole topic. The channel everyone calls unmeasurable is the one where reach and frequency still mean something. The channels everyone calls measured are the ones where those same metrics are going void. DOOH never had a click to hide behind, so it never built the pretense the web is now dismantling: it has been saying “modeled, with the methodology attached” since Geopath’s founding in 1933. Meanwhile signal loss, consent law, and browser-governed measurement are pushing web numbers toward exactly that shape — modeled estimates, methodology attached. The street is not behind on measurement. It arrived early at the condition everyone else is heading toward.
Count places, not people
The most common form of the declined question was reach and frequency: how do you deduplicate reach across fifty screens, or cap frequency against a person who walks past the same billboard twice a day? At person level, you don’t. There is no identifier to join on, and a vendor claiming person-level deduplication across screens is describing a model and owes you its methodology.
The workable answer is geographic, and it is being argued at the standards layer right now: Chris Williams has been making the buyer-side case, and I co-lead the working group where it landed. Cross-channel comparison fails today for a structural reason — there is no consistent geographic definition across media types or countries. Media markets, postal areas, and administrative regions differ by channel, redraw over time, and hold populations that shift by daypart, day, and season. Reach cannot be compared across units that don’t align. The fix in front of the AdCP working group is a hierarchical cell grid: a globally consistent, resolution-laddered, non-overlapping geographic unit that never redraws. A related proposal on the W3C’s private-advertising list would standardize attribution’s unit of analysis the same way, as a geographic cell crossed with an ISO week and an hour of the day. Place and time become the coordinate system rather than the metadata.
On that foundation, frequency governance works without identity. For each cell, take the target audience size — supplied as a named, versioned population surface with a declared time grain, because a concourse at 8am on a Tuesday and the same concourse at 11pm on a Sunday are different denominators — and multiply by the desired average frequency. That product is the impression ceiling for the cell. Buy until the ceiling, then stop. No person is tracked. The place is counted instead.
“Frequency by place works because a cell has a population you can count. That population is a named, versioned surface with a declared time grain, so the ceiling at 8am and the ceiling at 11pm are different numbers.”
— Nathan Woodman, Ether Data
One rule keeps the arithmetic straight: the ceiling is computed and enforced at a single declared resolution. Inside one grid, rolling fine cells up to their parents is exact algebra. Every cell has one parent, and no impressions are created or lost on the way up. The trap is a mixed set, one carrying a parent cell and its own children at the same time, which is what compact encodings produce on purpose. Summed per cell, a mixed set counts the same ground twice. So a set that arrives mixed gets normalized to one resolution before anything is summed, and at that declared resolution non-overlapping cells sum cleanly to the campaign’s full scope. Weighting enters only one step later, when cells reconcile outward to a media market or a postal area. Those units share no index with the cell grid, so the apportionment runs over a declared weight surface, and area is the wrong default weight for a population count, because people are not spread evenly across a cell. The mixed-resolution mechanics are written up in full in Ether Data’s geo interoperability knowledge base, which ships a kb.json and a markdown twin of every page so an agent can grep the methodology instead of reading around it.
It is also worth saying why this channel gets the clean version first. A play happens on a screen at a surveyed address, so the cell that holds it is a fact, with no positional error to declare. The exactness belongs to the channel, not to the unit. Put addressable inventory into the same ledger and the exactness leaves: a mobile impression lands in its cell through a location lookup that carries real error, and both impressions then sit under one cap as though they were placed with equal confidence. Woodman raised that floor on the same working-group thread last month, and it is a cross-channel problem rather than a DOOH one.
Just as important is what this architecture refuses to attempt: real-time deduplicated reach across all media inside the buy decision window. That is the industry’s hardest problem, and no protocol should pretend otherwise. The per-cell ceiling is the guardrail that makes the whole thing shippable without solving it.
Proof is a contract term
The other cluster of declined questions was about proof: how do you verify a DOOH campaign actually ran? The uncomfortable fact is that no single cross-industry technical standard for proof-of-play exists. Playout reporting is an operational practice, much of it self-reported, and the trade bodies describe its current state as variable themselves. The burl event proves a play became billable; the qty multiplier proves a vendor’s model produced a number. The UK previews where this converges — Outsmart publishes OpenDirect (OOH) 1.5.1, an Impression Multiplier Standard with published methodology, and Playout 2024, a centralized industry tool housing media-owner playout data in a published file format — but those are UK-market standards, not global ones. Until that changes, what counts as proof is defined per media owner, in the contract: playout reports, burl reconciliation, audit rights, all specified before spend.
Agents are what make this urgent instead of merely annoying. An agent will happily assemble and execute a screen plan in seconds, and it will enforce exactly the proof obligations someone encoded and nothing more. The agentic standards work has started carrying the channel natively: AdCP’s delivery metrics already include plays, loop plays, screens used, screen time, share-of-voice achieved, and venue breakdowns, and its broadcast guidance states the settlement split that applies to every scheduled channel — playout logs prove the spots ran, and a measurement currency turns those plays into an audience number. Its design rule for the wire is the right discipline for the whole category: pass “the reliable audience number and its basis, not a made-up digital completion surrogate.” Static out-of-home is the frontier case, because a billboard has no play event at all: its currency is estimated impressions from a measurement body, plus panel IDs, posting periods, and share-of-voice as a contracted rate. The wire shape for that is being scoped in the same working group now.
One more consequence of the chain: estimates compound. A modeled multiplier times a modeled currency times a modeled outcome study is a stack of models, and an agent optimizing on that stack without explicit error awareness will optimize the noise. Methodology checks are not compliance theater here. They are what keeps the optimizer pointed at the signal.
The currencies have to become data
The audience link of the chain belongs to measurement bodies, and modernizing the model is only half their work. Geopath’s overhaul is underway — the Ipsos pilot in late 2026, transition from 2027 — but a better model that agents cannot consume changes nothing about how the channel trades. Three moves make a currency real in an agentic market. Name it and version it in the protocol vocabularies, so a bid or a delivery report can say which system produced its audience figure. Publish the reference population as a named, versioned population surface with a declared time grain, which is exactly what a joint industry committee already produces, expressed in a form agents can read. And declare the currency at delivery time, so a reconciliation join can tie the delivered number to the system that measured it.
The entry cost is small. Confirming how a currency should be named and versioned in a standards vocabulary is a weeks-scale contribution, not a governance commitment. The standard defines the rails; the population models, outcome methodologies, and accreditation that run on them remain the measurement bodies’ territory to fill. The door is open now, while the vocabulary is still being set, and the bodies that walk through it get their names written into the layer agents will transact on.
What to do with all this
For buyers: on every buy, require the declared venue taxonomy and version, the multiplier vendor and source type, proof-of-play terms in the contract, and the methodology behind every audience number. Read outcome studies as modeled evidence. The same discipline extends to in-store retail media, which is place-based media wearing a lanyard: same venue questions, same multiplier questions, same proof questions, inside the store.
For screen networks and media owners: transparency about how your numbers are made is a sales asset, because buyers increasingly gate spend on it, and agents will gate it mechanically.
For measurement bodies: assert your currency into the agentic standards while the vocabulary is still forming. A currency that exists only as a PDF and a panel is invisible to the systems that will soon allocate the next dollar.
And for the eight readers who asked: measure DOOH like the modeled channel it is. Read what is observed: the play. Audit what is declared: the venue. Respect what is modeled: the crowd, the audience, the outcome. Govern frequency by place instead of person. The street never promised you a click, and it turns out the click was the anomaly. The reference layer for every spec named here lives in the DOOH and place-based standards deep dive; this essay is the argument for reading it without apology.