How to build inventory value in a fragmented programmatic ecosystem

More data isn't more value. Denser data is.

The wrong question

First-party data, third-party data, zero-party data. What do they actually mean? Which one is better, and which should you use? How do you create them, and what can you do with them?

I hear these questions in almost every publisher meeting. I've come to think they are the wrong questions. The labels describe where data comes from, not what it is worth. A first-party segment can be nearly worthless, and a well-built contextual signal can outperform it.

I have thought a lot about how to describe the complexity I see in the programmatic ecosystem when it comes to data. There are many variables, concepts and perspectives, and they do not necessarily align. The concept I keep coming back to is Data Density.

Data Density = Addressability × Accuracy

I have boiled it down to two dimensions: Addressability and Accuracy. Data Density is what you get when you multiply them.

Multiplying, not adding, is the point. If either factor is zero, the value is zero. A logged-in user with no meaningful attributes is worth little. So is a perfect segment that only reaches 2% of your inventory. Value lives where both are high at the same time.

• Addressability asks: how much of my inventory can this data actually reach?
• Accuracy asks: how sure are we that the attribute is true for the user we are reaching?

Addressability: how far does your data reach?

Addressability is the reach of your data. It depends on cross-domain support in your IDs or cookies, and on the share of your inventory that is addressable for audience targeting at all.

It is also relative to the campaign. Is it a cross-network programmatic campaign, or an IO campaign on a single site? Is targeting done on the buy side or the sell side? A single-site IO deal can live with a first-party cookie. A cross-network DSP campaign needs an ID the buyer can recognise.

Density decays at every hop

The hard part is that addressability leaks at every step of the chain. An illustrative example of 100 impressions:

1. 100 impressions served
2. ~60 with valid consent (TCF, or pay-or-OK)
3. ~35 with a persistent ID across sessions (Safari and Firefox cut this hard)
4. ~20 matched to an ID the buyer actually uses
5. Fewer still that reach the DSP with the signal intact

The exact numbers differ for every publisher. The pattern doesn't. Every handover between consent tool, ID provider, SSP and DSP loses signal. That is what "fragmented" really means in practice.

The levers

• Consent: consent rate is the first multiplier. Pay-or-OK models in DACH and the Nordics have changed the math here.
• IDs: unified IDs and ID graphs extend reach across domains, but each adds its own match-rate loss.
• Login: the strongest signal, but only for the share of users who log in.
• Non-consented signals: contextual and cohort-based signals keep part of the unconsented traffic sellable without personal data.

Here is the European reality, and where I think we see things US commentators miss: Chrome's third-party cookies never disappeared, but in a GDPR market they were never the foundation anyway. Consent and login are.

Accuracy: how true is the attribute?

Accuracy is about how well the attribute you target is actually represented in the user you reach. Five questions get you most of the way:

• History: how long a user history is the data based on? One session or six months?
• Breadth: is it based on one domain or many?
• Identity: is it tied to a random first- or third-party ID, or to a login?
• Method: is the attribute assigned deterministically (the user told us, or did it) or probabilistically (a model guessed)?
• Confidence: how sure are we that the attribute belongs to this user and not to the household or the shared device?

Time turns context into interest

A single article view is context. The same topic read repeatedly over weeks is interest. Frequency and recency are what separate the two. This is the publisher's natural moat: no third party sees the full reading history on your sites, so no third party can build interest data that is as accurate.

The attributes advertisers value most are geography, socio-demographics, context, interest, and intent. Intent is the hardest to get right, and the most valuable when you do.

Black box vs. transparency

Academic studies of commercial third-party segments have found that even basic attributes like gender and age are often only modestly better than a random guess. A 2019 study in Marketing Science by Neumann, Tucker and Whitfield tested over 90 third-party audience segments from 19 data brokers. Compared with picking users at random, the segments improved targeting accuracy by between 0% and 77%. For gender, the brokers did about as well as random chance. When we built a dedicated lookalike algorithm and benchmarked it against verified logged in users with verified gender, we could predict gender with an accuracy of 78% for men and 87% for women. BUT. This was dependent on the specific publisher having very gender oriented content, for generic news content it becomes almost impossible to predict gender. What this shows is that your content matters for the algorithms. And your login strategies matter for gender targeting.

The lesson for publishers: buyers discount data they can't verify. Accuracy you can't explain gets priced like low accuracy. Declaring how a segment is built (deterministic or probabilistic, how many domains, what lookback window) is not a weakness. It's how you get paid for the quality you already have.

The Data Density matrix

Put the two dimensions on a grid and the vocabulary question answers itself. Each data type lands in a quadrant, and none of them sits in the top-right by default.

So, which is better: first-, second-, third- or zero-party data? None of them on their own. Zero-party data is highly accurate but rarely addressable at scale. Third-party data is addressable but often inaccurate. First-party data can sit anywhere on the grid, depending on consent, ID strategy and how long the history is.

The real question is: which combination moves your inventory towards the top-right, for this campaign?

A worked example

Let's take one segment, "finance-interested 30 to 45 year-olds", through one publisher. The numbers are illustrative, but the shape is what I see again and again.

The publisher serves 10 million impressions a month.

So the publisher has three products, not one:

• ~11,000 impressions of very high accuracy (logged-in, declared): premium price, tiny reach.
• ~109,000 impressions of good accuracy (behavioural interest, modelled age): the core sellable segment.
• ~500,000 impressions on finance content (5% of the total), sellable contextually to everyone, consent or not.

Sold as one blurry "finance audience", these get averaged down. Packaged by density, each finds its price.

How to increase your Data Density

You raise density by working both axes at once. Six levers that matter most:

1. Raise consent. It is the first multiplier on everything else. Test your CMP design, and consider pay-or-OK where your market supports it.
2. Build a login strategy. Logins give you deterministic identity across devices. You don't need everyone. A loyal 10% can anchor your models for the rest.
3. Collect declared data. Short zero-party surveys and preference centres give you ground truth. Use it to calibrate and validate your probabilistic models, not just as a segment on its own.
4. Turn context into interest. Track topics over time, with frequency and recency, across all your domains. This is the data only you have.
5. Keep a contextual baseline. Make 100% of your inventory sellable, including unconsented traffic, so nothing is dark.
6. Make it activatable. Data the buyer can't reach is worth nothing. Package segments as curated deals, pass them through deal IDs or seller-defined audience signals, and support the IDs your buyers actually use.

And be transparent about all of it. As buying becomes more automated, quality has to be declared in a way machines can read. Density will become something you state and prove, not something you claim in a sales meeting.

Your Data Density self-assessment

Take ten minutes with your ad ops and data team and answer these:

• What share of your impressions is consented?
• What share has an ID that persists across sessions, and across your domains?
• What share of users is logged in?
• What share of impressions carries at least one audience attribute?
• How many sessions of history are your interest segments based on?
• Which segments are deterministic, and which are probabilistic? Do your buyers know?
• Is 100% of your inventory sellable, at least contextually?
• Through which paths can buyers actually activate your data: IO, PMP, curation, or open auction?

If you can't answer half of them, that is your first finding.

More data isn't more value. Denser data is.

If you'd like to compare notes on your own numbers, write to me at martin@samhub.io.