The Context Economy Has a Governance Problem

The name that was missing

Nathan Woodman, writing in AdExchanger’s Data-Driven Thinking column, just gave a name to a shift the open web has been feeling for two years: the move from an audience economy to a context economy. Audience systems — identity graphs, clean rooms, deterministic IDs — explain who. They are, in his sharp phrase, “the past tense of intent.” Context systems explain what matters right now: where the person is, what surrounds them, what is happening in the world, and what they are most likely to want in this exact moment.

He is right. And he is describing, from the infrastructure end, the same transition this site argued from the signal end five weeks earlier — that the future of advertising is not better segments but a live, governed interpretation of the moment. When an independent columnist and your own thesis converge on the same architecture, that is worth noticing.

It is not only adtech noticing. The same week Woodman’s column ran, a 2026 Databricks executive publication aimed at CIOs and CDOs — The New Architecture of Agentic AI — opened on the identical diagnosis from the enterprise-data side: “the models are smart enough. What they do not have is context.” One of the two sources is vendor-produced — Databricks marketing its own architecture — and the convergence is more telling for it, not less: when an independent ad-tech columnist and a vendor’s own pitch arrive at context is the scarce input in the same news cycle, from opposite ends of the stack, the shift has stopped being a talking point. It is the weather.

But there is a sentence doing quiet, heavy lifting in the context-economy pitch, and it deserves to be pulled into the light before the industry falls in love with it.

The moment is a better product — and a sharper risk

Woodman opens with an anecdote: he spends two hours near a friend who has been searching “retire in Italy,” and that evening his phone serves him Tuscan-villa ads. Nobody was listening. The system “inferred proximity, interest and timing” from models trained on millions of similar moments. His framing: “the kind of relevance that feels like your phone is listening to you without anyone actually needing to listen.”

That is genuinely the better product. It is also the exact moment the context economy becomes something people have every right to be uneasy about. “Relevance that feels like your phone is listening” is a triumph of engineering and a confession in the same breath. The audience economy was creepy because it remembered too much. The context economy is creepy because it infers too much — from proximity, from company kept, from a conversation you had at a dance recital that no microphone recorded.

Here is the uncomfortable symmetry: everything that makes context economics powerful — real-time inference over dense signals, proximity and timing surfaced at auction speed — is also everything that makes it feel like surveillance when it goes wrong. The difference between “the moment the user intuitively knows they want” and “a stranger inferred my private plans from who I stood next to” is not a difference of technology. Both use the same embeddings, the same colocation, the same single-pass inference. The difference is entirely a matter of governance — of what the system is permitted to infer, from whose data, under what consent, with what off-switch.

Which is to say: the context economy’s whole legitimacy rides on the one layer the celebratory version leaves out.

The infrastructure is real — and it is neutral

None of this is a reason to doubt the infrastructure. It is arriving, fast, and it is impressive. The IAB Tech Lab’s ARTF turns the exchange into a container runtime: third-party agent services run co-located inside the host platform, mutating the bidstream through a standardized API that exposes only the data each task needs. Agentic Audiences — LiveRamp’s donated User Context Protocol — carries the payload: dense embeddings compressing identity, contextual, and reinforcement signals into a form fast enough for in-loop inference. Colocation compresses transit time and expands thinking time, opening a compute budget inside the auction that never existed before.

That budget is what makes “interpret the moment in real time” a buildable claim rather than a slogan. But notice what the architecture is and is not. ARTF specifies how a container runs and what data it may touch — it is a control surface, not a conscience. It can enforce “expose only what this task needs” exactly as easily as it can enforce “expose everything the host will sell.” The framework is neutral. What flows through it is a policy choice, and policy is not something a gRPC schema decides.

This is the same lesson the W3C Attribution API debate surfaced last week: the test of any of this machinery is never the information flow — it is who decides. A context system that infers your retirement plans from proximity data is a governance decision wearing an engineering costume. The engineering is settled. The decision is not.

The moment has four dimensions

A moment is not one thing. It is at least four: who is there, what they are consuming, where they are standing, and when in the week it is. Run Woodman’s recital anecdote back through that grid and every dimension is doing work — the who (a friend, in proximity), the what (a conversation about Tuscany), the where (a recital hall), the when (a Saturday afternoon). The engine needed all four to make the moment legible. Take any one away and the Tuscan villa never surfaces.

Now look at what the stack actually carries. Identity graphs, contextual signals, the dense vectors inside Agentic Audiences, the signal containers this site has argued for — all of it encodes the who and the what. The where and the when ride along as raw fields when they ride at all — a lat-long in a bid request, a timestamp in a log. Coordinates, not representations. Nothing in the stack yet represents place and time the way embeddings represent meaning.

The reason sits inside the embeddings themselves. The vectors in today’s stack are text models: they encode the words that describe an audience, a page, a brief. And text embeddings are famously lossy in exactly three places — numbers, geography, and time. Ask a text model whether 2:14 a.m. is near 2:41 a.m. and it reasons about the characters, not the clock. A second family inverts this. Spatiotemporal embeddings are made of what text models lose: the vector encodes the place and the hour themselves, learned from the structure of census, workforce, and mobility data rather than from descriptions of it. The category already has public proof on the spatial axis — Google’s Population Dynamics Foundation Model ships graph-learned embeddings for some 28,000 U.S. postal codes, built from aggregated search trends, busyness, weather, and air quality. The hour of the week as a native coordinate is the newer frontier — and it reached a standards body this week: a proposal on the W3C’s private-advertising list would standardize attribution’s unit of analysis as a geographic cell crossed with an ISO week and an hour of the day. Place and time as the coordinate system, not the metadata.

Here is where this stops being an engineering note and becomes this essay’s argument. A text space and a place-time space do not share a geometry. Computing similarity between them is not hard — it is undefined, until someone defines what comparison means across the two. That is a standardization question, not an integration ticket: agents need an agreed way to weigh a semantic signal against a spatiotemporal one inside the same decision, and that agreement has to be written somewhere all sides can read it. It is exactly the kind of definition a signals working group exists to produce — and I can report from inside one that it is on the table, not hypothetical. In the AdCP Signals & Measurement working group, where I contribute as a founding member, the time axis has already started entering the schema: the signal definition recently gained a last_updated field — an addition I authored — freshness becoming a first-class property of the signal object. Metadata first, coordinates next. The signals adaptor this site builds in the open is one of the reference implementations that work anchors against.

And the governance thesis does not relax when the moment gains dimensions — it doubles. Where and when carry their own protected inferences: a place can be a clinic, a courthouse, a shelter; an hour-of-week pattern can reveal worship, shift work, a school run. A four-dimensional moment is four dimensions of inference someone must be permitted to make. The consent gate does not get smaller as the moment gets richer. It is the only thing standing between “one agent reads every dimension of the same moment in a single pass” — which is the version of agentic advertising worth agreeing on — and a system that knows where you stand at 2 p.m. on Sundays and never asked.

What the container model already knew

This is why the signal container framing matters more, not less, as the context economy arrives. A signal container was never only about making the moment activatable. It packages intent with its policy — consent, provenance, expansion rights, identifier constraints, and an audit trail — so the moment cannot travel without the rules that govern it attached. The policy layer is not paperwork stapled to the signal. It is the thing that separates a context economy from a surveillance economy running on the same rails.

And that architecture is no longer only this site’s argument — it is partially operational in the standard I help build. In AdCP’s current signal schemas, restricted_attributes and policy_categories are fields on the signal definition itself, with consent basis and data-subject rights traveling in the governed plan around it. The policy rides at the protocol layer, not in a side letter. That is what “the moment cannot travel without its rules” looks like when it stops being prose and becomes a schema — and it is why the signals adaptor this site ships in the open treats the governance fields as load-bearing, not decorative.

Woodman’s own list is closer to this than the celebratory framing lets on. “Containerized bidding provides the environment,” he writes; “embedding interoperability provides the language; secure data collaboration provides the memory.” Environment, language, memory. What is conspicuously absent from that trio is consent — the fourth thing that decides whether the memory was ever ours to trade. A context economy with environment, language, and memory but no governance is not a new economy. It is the old surveillance economy with a lower latency.

The trust argument this season has been making lands here too. Trust is the scarce asset because a machine cannot mint its own standing. In the context economy, the same scarcity reappears one level down: an inference engine can manufacture relevance, but it cannot manufacture the permission to have inferred what it inferred. Permission, like trust, has to be granted from outside the system — by a person, through a consent that means something, enforced by a governance layer that can actually say no. Colocation gives the machine a bigger budget to think. It gives it exactly zero additional right to.

Permissions are a relevance signal

Here is the part that should change how operators think about the governance layer, because it is not a moral argument — it is a performance one. The instinct is to read access controls defensively: keep the system away from the data it should not touch. But the same constraints that prevent exposure also tell the system what matters. Hanlin Tang, who leads neural-network engineering at Databricks, puts it precisely in that same enterprise publication: in an agentic system, governance “is not just security. It is also quality. Permissions are a signal of what to focus on and what to ignore.” If five versions of a customer table sit in the catalog and the agent may only reach the current one, it cannot reason from yesterday’s numbers — the permission is doing quality work, not just compliance work.

Carry that back to the moment. A context system with no governance layer does not merely risk inferring things it had no right to infer. It also drowns in signal it cannot rank. Consent, provenance, and scope are not the tax you pay to reach the moment — they are how the system knows which moment, from whose data, is the one worth acting on. The consent gate in the container is not standing between the signal and the moment. It is the thing that tells the signal which moment it is for. Read that way, governance stops being the brake on the context economy and becomes the engine of its relevance — which is the reason it stops being overhead at all.

If that still sounds abstract, here is a miniature from inside the machinery, from the same week as this essay’s four-dimensions addition: a scope-isolation fix I proposed to AdCP’s conditional-fetch path. The protocol lets a buyer’s agent ask “has this signal feed changed since I last looked?” and be told unchanged — cheap, fast, cacheable. The catch: the cache token is scope-keyed by design, because an account with negotiated overlays is not looking at the same feed as the public one. An implementation that ignores the scope key can answer unchanged from someone else’s scope — and the account-specific pricing the caller was entitled to simply never arrives. Notice what kind of failure that is. Nothing leaked. No privacy audit would flag it. The buyer just got the generic answer instead of the one it had negotiated the right to see — a relevance failure, caused by a governance boundary going unenforced. That is “permissions are a relevance signal” with the abstraction removed: what a system is allowed to see is information about what it should be seeing, and the boundary check is not the brake on the exchange — it is the only mechanism by which the right signal reaches the right buyer at all.

For operators, then

If the context economy is real — and it is — the operator move is not to marvel at the relevance. It is to treat the governance layer as the product, not the compliance tax. Three questions, none of them about targeting:

1. What is the system permitted to infer, and from whose data? Proximity, timing, and company-kept are inferences about people who never opted into being your signal. The container’s policy layer is where that line gets drawn — or doesn’t.
2. Can the off-switch actually be exercised, invisibly? The Attribution API’s one genuinely honest feature is that opting out is undetectable to the site. A context system without an unpunishable off-switch is not offering a choice; it is offering the appearance of one.
3. Does the provenance travel with the signal? When a moment is activated three hops from where it was observed, the only thing standing between “relevant” and “creepy” is whether the source, method, and consent came along for the ride. That is a container property, and most of the market does not have it yet.

Woodman’s closing line is the one worth keeping, because it is true: “The future of advertising will not favor those who know the most about audiences; it will favor those who also understand the moment and can act on it instantly.” I would add only the clause the season’s writing keeps insisting on. It will favor those who understand the moment, can act on it instantly — and can prove they were allowed to. The first two are engineering, and the engineering is nearly done. The third is governance, and the governance is the whole game.

The context economy is coming whether or not the industry builds the governance layer. Whether it is a better economy or just a faster one depends entirely on which of those the winners decide to optimize.