Why the next era of advertising is built on signals, not surveillance — regardless of what happens to cookies.
The advertising industry has been built on a simple premise: to reach the right person, you need to know who that person is. Track them. Identify them. Follow them across the web and serve them something relevant. For two decades, this premise drove one of the most sophisticated technological ecosystems ever built around a commercial objective — and it worked, well enough, for long enough, that the industry stopped questioning whether the premise itself was sound.
The foundation it was built on — individual identity assembled from behavioural trails that were never designed to reveal genuine intent, and increasingly restricted by regulation, platform policy, and user choice — was always more fragile than the industry admitted. The question worth asking now is not how to preserve that foundation. It is whether, for the first time, we have the technology to build something genuinely better — something that understands people rather than tracks them.
The model we built — and what it was actually solving
The identity-based model was not a deliberate choice to track people rather than understand them. It was an engineering solution to a genuine limitation. In the early years of programmatic advertising, the computational infrastructure to read complex behavioural signals at scale, interpret semantic meaning, model psychographic depth, and harmonise that understanding across environments simply did not exist. What did exist was the ability to drop a cookie, record a page visit, and match that identifier to another page visit somewhere else on the web. So the industry built on what it had — and built extraordinarily well. Cookies became the currency, identity graphs became the asset, and a generation of adtech was constructed on the implicit assumption that knowing where someone had been was a reasonable proxy for knowing who they were.
It was never a great proxy. It was just the best one available.
To give the model its due: cookie and ID-based infrastructure created real and valuable capabilities. Attribution, retargeting, closed-loop measurement — the ability to understand whether the person who saw the ad became the person who bought the product. These are legitimate problems that identity infrastructure solves well, and the DMP and CDP ecosystem built around them represents genuine engineering achievement that the industry rightly values. A hybrid model — where AI-derived signal intelligence powers upstream audience understanding while identity infrastructure handles downstream measurement and attribution — is a coherent and honest near-term position. The two capabilities are complementary when they stay in their respective lanes.
The problem was not that identity infrastructure existed. The problem was that it got repurposed. Attribution technology became audience technology. A tool designed to measure what happened after the fact was stretched to answer a fundamentally different question: who should we be trying to reach in the first place? A trail of where someone has been is not the same as an understanding of who they are, and the more the industry relied on that imperfect proxy — for audience building, for curation, for upstream media strategy — the more it constructed targeting, planning, and media decisions on a foundation that was never designed to support them. Recognition tells you who came back. Understanding tells you who to go find. The industry needed both, and only built one.
We never had the technology to understand people. Until now.
This is the thesis that changes everything — and the one the industry has not fully confronted. The signals that genuinely reveal who a person is have always existed, and they were always more revealing than any identifier trail. What was missing was the computational sophistication to read them accurately and act on them at scale.
Consider what actually characterises an audience: not the pages they visited, but the content they chose to engage with. The language they use. The topics that recur across their attention. The media environments where they spend time voluntarily. The psychographic patterns that connect their interests into a coherent picture of their values and motivations. These signals are not a proxy for human behaviour — they are an expression of it. They are observable and continuously updated, and they do not disappear when a browser blocks a tracker or a walled garden closes its data. The reason the industry did not build on them was simply that it lacked the tools to process them with sufficient sophistication to act on at advertising scale.
That constraint no longer exists. Large language models trained on the breadth of human language and behaviour can now understand the semantic relationships between an audience's interests and the content environments they inhabit. They can read a brief and extract not just what was said but what was meant — the underlying human picture that a skilled planner would intuit from years of category experience, now derivable systematically and in real time. They can identify patterns in how people behave online — what they say, what they search for, what they engage with, how they describe their own interests — and harmonise those patterns into a coherent, accurate, continuously updated model of who an audience actually is.
This is not incremental improvement on the old model. It is a different model for a different problem. The old model asked where a person had been and what that implied about what they might want. The new model asks what a person's genuine behaviours and signals reveal about who they are right now. The first question produces a trail. The second produces understanding. And upstream audience understanding — before a brief enters a platform, before a budget is committed — is precisely the problem that identity infrastructure was never designed to solve.
Immutable signals. The real building blocks of audience intelligence.
The signals that actually define an audience are not the pages they visited. They are the content they chose to engage with, the language they use, the topics that recur across their attention, the media environments where they spend time voluntarily, and the psychographic patterns that connect their interests into a coherent picture of who they are and what motivates them. These signals are immutable in a way that cookies never were. They do not disappear when a browser blocks a tracker, degrade when a match rate drops, or require an individual to be followed across the web to be observed. They exist in the content people consume, the language they use, the environments they choose — continuously available, continuously updated, and far more revealing of genuine intent than any identifier trail.
When you can process these signals with the sophistication of modern AI — reading semantic relationships between interests and content environments, interpreting behavioural patterns at scale, deriving psychographic depth from observable online behaviour — you arrive at something qualitatively different from what the old model ever produced. Not a better approximation of who someone is. Genuine understanding, derived directly from how they actually behave, stable across the regulatory and technical shifts that continue to erode identifier-based infrastructure. That understanding does not degrade with privacy legislation. It does not break when a walled garden closes its data. It is grounded in behaviour, not in the technical scaffolding built to observe it.
The harmonisation problem the industry has not solved
Even if you accept that signal-based intelligence is more durable and more revealing than identifier data for upstream decisioning, you immediately encounter a structural problem that better signals alone cannot fix. Every channel speaks a different language. YouTube has its own audience taxonomy. Meta has its own. Every DSP, every CTV supply path, every audio environment has its own. The understanding you derive from behavioural signals does not map neatly from one environment to another without a layer specifically designed to translate it.
The industry's response to this problem has been to rebuild audiences inside each platform separately — losing coherence at every handoff. The result is not five expressions of one audience but five different audiences, each shaped by the assumptions, incentives, and native data of the platform that built it. The original intent gets diluted and reinterpreted at every step downstream, so that the audience running on YouTube is not the same audience running on Meta, which is not the same audience running in the DSP — even when all three were intended to represent the same brief. The fragmentation is structural, not accidental, and it compounds with every channel added to the plan.
This is why unified audience logic requires more than better signals — it requires a layer that sits above all platforms, defines the audience once in a platform-independent form grounded in live behavioural intelligence, and translates that definition into each channel and environment's native targeting framework without losing its meaning. Not rebuilt inside each system, but translated from a single upstream source. This is the layer we are building — not inside the platforms, but above them — grounded in the conviction that audience intelligence derived from genuine human signals, validated against psychographic frameworks already trusted by the market and translated natively across every channel and environment, is more accurate, more durable, and more honest than anything the identifier model ever produced. When the brief, the audience, and the inventory selection all derive from the same upstream source, outcomes are not just better. They are explainable. And explainability is what turns a result into a repeatable process rather than a lucky outcome.
The 1P data problem agencies aren't talking about loudly enough
There is a second structural shift compounding the identifier problem, and the industry is only beginning to reckon with it fully. First-party and zero-party data is being inhoused. Brands have spent the last several years recognising that their customer data is among their most valuable commercial assets, and the rational response has been predictable: hold it closer, build internal data infrastructure, reduce the number of third parties with open access. The era of brands freely sharing CRM data with agency partners as a matter of course is ending — not dramatically, but steadily and structurally.
Clean rooms emerged as the architecture for navigating this transition — privacy-safe environments where brands and agencies can match datasets and extract insights without exposing raw records. The capability is real, but clean rooms require structured data infrastructure on both sides, legal and procurement alignment that can take months to negotiate, and ongoing technical maintenance. They work well for large enterprise relationships with significant shared commercial incentive and the resources to sustain them. They are not a practical solution for the majority of agency relationships running across dozens of clients with varying data maturity, budget levels, and category contexts. The agency sitting between a brand that has inhoused its data, a platform that has restricted its identifiers, and a regulatory environment that continues to tighten is being squeezed from every direction simultaneously — and this is not a temporary dislocation but a structural realignment of where data lives and who controls it.
An audience model built on immutable behavioural signals does not require a client to share their CRM or a clean room agreement to be in place before a campaign can be planned. It derives understanding from what is genuinely observable — independently, upstream, before any data partnership is needed. Clean rooms remain valuable where they work, particularly for connecting modelled intelligence to validated real buyer data and for closed-loop measurement, and the industry should continue to develop them. But as the primary targeting strategy for most agency relationships, the clean room model assumes a level of infrastructure, legal alignment, and commercial maturity that the majority of the market cannot consistently access. Unified audience logic built on behavioural signals does not depend on structural conditions that are actively deteriorating — and it becomes more valuable with every year that the data-sharing model tightens.
Addressing the counter: identifiers do make targeting better
This argument will attract serious counter-argument, and it should. Cookies and identity-based targeting do make audience targeting more precise in specific and important contexts. Retargeting a known customer is genuinely more efficient than reaching a modelled lookalike. Suppression lists save budget by excluding people who have already converted. Attribution models that close the loop between ad exposure and purchase provide the kind of accountability that brands and CFOs rightly demand. For downstream measurement, CRM-connected activation, and closed-loop optimisation, identity-based infrastructure remains the right tool — and we are not arguing against it in those contexts.
The distinction the industry has conflated for two decades, however, is that precision is not the same as understanding. An identifier gives you precision within a known population — people who have already interacted with your brand in some form, whose behaviour you have already observed. That is genuinely valuable for managing existing customer relationships. It is far less useful for the upstream problem: reaching the right new people, understanding what they care about before they have ever interacted with your brand, curating media for an audience that has no prior relationship with your business. That problem requires genuine comprehension of who people are, not a record of where they went. Identifier-based precision is also precision on a fragmenting addressable universe — Safari and Firefox already block third-party cookies by default, consent rates are declining, and match rates across walled gardens continue to fall. The precision is real, but the ceiling is structurally constrained in ways that no amount of engineering sophistication can fully offset.
Recognition tells you who came back. Understanding tells you who to go find. Both matter, but only one of them scales into the future without a structural ceiling.
Privacy-first is not a constraint. It is the point.
The privacy conversation in adtech has been framed as a compliance problem from the start — how to preserve targeting capability as regulators restrict what can be tracked. That framing is backwards. Privacy-first is not a constraint to be engineered around. It is the direction the world is correctly moving, and it validates the thesis rather than threatening it.
The identity-based model was built on behavioural trails that were never designed to reveal genuine intent, and the regulatory responses across every major market — GDPR, CCPA, App Tracking Transparency — represent a structural correction to that reality. These are not temporary compliance burdens. They are a permanent recalibration of the relationship between digital behaviour and commercial exploitation of it, and that recalibration is not reversing regardless of what any single browser vendor decides to do with cookies. A model built on understanding rather than tracking is inherently aligned with where that recalibration is heading. Behavioural signal intelligence derived from what people choose to do publicly does not require following an individual across the web or assembling profiles from trails people never intended to leave. It observes what is genuinely observable — the content people engage with, the language they use, the environments they choose — and derives understanding from patterns rather than from surveillance.
This is not a privacy-by-compliance model. It is a privacy-by-design model, one that does not need to be retrofitted for a privacy-first world because it was built for one from the start. The industry has spent years searching for the next identifier — a new technical mechanism to preserve the old model under tightening conditions, whether through universal IDs, cohort models, or probabilistic graphs. Each represents a sophisticated attempt to continue what was always done under different constraints. The more honest question is whether the old model was ever the right one for the upstream problem the industry most needed to solve — not attribution and measurement, but genuine comprehension of who an audience is and what they care about before money is spent. A model built on that comprehension does not need a new identifier, because it never needed the old one for that purpose.
Synthetic audiences — the logical extension
Follow this thinking to its natural conclusion and you arrive somewhere the industry has not fully reckoned with yet. If an audience can be understood deeply enough through immutable signals and behavioural patterns — if the model is rich enough to characterise not just who an audience is but how they think, what they respond to, and why — then something previously impossible becomes possible. You can ask the audience questions before the campaign runs. Not by surveying real people, but by building synthetic representations of audience segments derived from the same unified intelligence that powers targeting, and using those representations to simulate how the audience would respond to a message, a creative, or a product positioning. Synthetic users modelled not from demographic assumptions but from the same immutable behavioural signals that define the audience in the first place.
The AI infrastructure to do this exists today. The missing piece has always been the audience intelligence underneath — a unified model grounded in genuine behavioural depth rather than identity approximation, and that is precisely the piece unified audience logic provides. This is not a thought experiment. The architecture that makes it possible is what we have been building, and synthetic audience modelling is the natural next step of that foundation — not a separate capability bolted on top, but the same intelligence extended into a new dimension. The audience definition that decides where to run also informs how the audience thinks. The line between planning and research dissolves because both are drawing from the same upstream understanding. Brief, audience, media, and message become coherent expressions of a single source of intelligence rather than four disconnected processes handed off between tools and teams.
When you can remove the individual from the equation entirely — while retaining genuine depth of understanding about how that audience thinks and behaves — you arrive at something the industry has been searching for since the first privacy regulation passed: advertising that is relevant without being intrusive, precise without being surveillance-dependent, and effective without requiring people to be tracked. That is what a genuinely privacy-first advertising model looks like. Not a model that finds clever new ways to follow people, but a model that never needed to follow them in the first place.
Not a replacement. A progression.
This argument is not a claim that identity-based infrastructure should be abandoned. Attribution, measurement, and closed-loop tracking are legitimate capabilities that identity infrastructure handles well, and the industry should continue to develop them. What this argument is, precisely stated, is a claim that the industry repurposed identity infrastructure to solve a problem it was never designed for — that upstream audience understanding required genuine comprehension of human behaviour all along, and that the technology to build that comprehension at scale now exists for the first time.
The honest path forward is not a choice between identity-based and signal-based models but a recognition that they answer different questions, and that each should be deployed where it genuinely belongs. The model the industry built was the best available answer to a real problem given the constraints of the time. Those constraints have changed, and the answer can change with them. The marginal gains in advertising are not downstream in execution — they are upstream, in the quality of the audience understanding that everything else is built from, in the coherence of the intelligence that travels across every channel and environment without fragmenting, and in the depth of the model that eventually knows an audience well enough to simulate how they think.
The platforms built their advantage on knowing where people had been. The next era will be built on understanding who they are — and that shift is not a threat to the industry. It is the industry finally solving the problem it always meant to solve.