A large advertiser in 2016 could describe its best customers in granular detail. Auto intenders in the third week of their research. Lapsed buyers of a competing brand. Households with a new baby. The marketer could find a few hundred thousand more people who looked just like them and start serving ads inside the hour. It felt like the future had arrived.
Almost none of that data belonged to the advertiser. The profiles had no names. The identifiers behind them would be gone within three months. And the same segments were on sale to every competitor in the category. This was the Data Management Platform era, and for roughly a decade it was simply how large brands thought about audiences. It worked well enough that few people questioned the foundation. The foundation is exactly what gave way.
This is part one of a three-part history of customer data infrastructure. It covers the DMP: what it was, how it plugged into the programmatic advertising machine, why it felt powerful, and the structural weaknesses that set up its decline. Part two covers the rise of the Customer Data Platform that replaced it. Part three looks at the agent layer that may come next. To understand why the CDP exists, you have to understand the thing it was built to fix.
Where the DMP came from
The DMP grew up alongside programmatic advertising in the late 2000s. Lotame was founded in 2006, BlueKai in 2008, and DemDex launched in 2009 as a self-described behavioral data bank before Adobe bought it in 2011 and renamed it Audience Manager. These companies were answering a specific problem created by a specific invention: real-time bidding.
Real-time bidding turned display advertising into an auction. When a web page loads, the publisher's ad slot is offered to many advertisers at once, each one deciding in well under a tenth of a second whether to bid and how much. The buying side runs through a Demand-Side Platform, or DSP. The selling side runs through a Supply-Side Platform, or SSP. An ad exchange sits between them and runs the auction. The whole transaction completes faster than a page can finish rendering.
That speed created a question. If a buyer has a hundred milliseconds to decide what an impression is worth, the decision comes down to one thing: who is the person seeing it. A 25-year-old browsing a car review is worth more to a car brand than an anonymous visitor on a recipe page. The DSP needed to know which was which, instantly, at massive scale. The DMP was the answer. It was the audience brain that sat next to the buying machine and told it who it was looking at.
What a DMP actually did
A DMP did three jobs. It collected data, it organized that data into audience segments, and it pushed those segments into the tools that bought and sold ads.
The collection ran almost entirely on the third-party cookie. A small piece of code, a pixel, placed across thousands of websites would drop and read cookies as people browsed. Each cookie was an anonymous identifier. The DMP never learned that the person was Jane Smith of a particular address. It learned that anonymous identifier number 4c9a had visited three travel sites, read two articles about Italy, and priced a flight. That was enough to act on.
DMPs combined three kinds of data. First-party data was the advertiser's own, usually thin, often just website visits. Second-party data was another company's first-party data shared directly. Third-party data was the engine: audience segments collected and sold by data brokers, covering interests, demographics, purchase intent, and life events, aggregated from across the web. BlueKai ran a marketplace where this third-party data was bought and sold at scale. By 2015 it claimed roughly 700 million actionable profiles.
The organizing step turned raw signals into segments. A marketer could define an audience by rules, anyone who visited the pricing page but did not buy, and the DMP would assemble the matching cookies. The technique that made DMPs feel magical was lookalike modeling. You gave the DMP a seed audience, your best converters, and its models scanned for the attributes those people shared, then found a much larger pool of strangers who matched the pattern. A control let you trade precision for reach: a tight 1 percent lookalike stayed very close to the seed, a loose 10 percent went broad. Suddenly a brand could prospect to people who had never heard of it, with a real statistical reason to believe they would respond.
The activation step pushed finished segments out. The DMP synced its audiences into the DSP, so the buying algorithm could bid up impressions for people in a valuable segment and ignore the rest. It pushed suppression lists too, so a brand stopped paying to advertise to people who had already bought. For a decade this was the shape of audience strategy at large advertisers: rented data, cookie identifiers, and thinking organized around campaigns rather than customers.
Why it felt powerful
The DMP deserves a fair hearing, because it solved real problems and the appeal was genuine.
It delivered scale and reach no first-party dataset could match. A brand with a modest website could still address hundreds of millions of people, because the third-party data market pooled signals from across the open web. It made prospecting measurable. Lookalike modeling replaced gut-feel media planning with something you could test, size, and tune. It put precise targeting inside an automated buying system, so a campaign could shift spend toward the segments that responded without a human touching it. And the marquee names lent confidence. When Oracle bought BlueKai in 2014 for around 400 million dollars, and Salesforce bought Krux in 2016 in a deal valued near 700 million, the category looked like permanent infrastructure. Every major marketing cloud had a DMP at its center. Buying one felt less like a bet and more like table stakes.
For a marketer judged on reach and cost per impression, the DMP genuinely moved the numbers. The trouble was that the numbers it moved were campaign metrics, and the foundation underneath them was rented.
The cracks in the foundation
The DMP did not fail because of one event. It failed because of structural weaknesses that were present from the start and got harder to ignore every year.
The data was rented, never owned. A DMP's third-party segments were a subscription. The day you stopped paying, the audiences were gone, and nothing accumulated as an asset. Worse, the same segments were sold to everyone. If your DMP could find in-market auto buyers, so could every rival automaker drawing from the same brokers. You were renting an edge that, by definition, was not exclusive.
The profiles were anonymous and had no persistent identity. A DMP knew cookies, not people. The same individual on a laptop, a phone, and a work computer looked like three unrelated strangers, because each browser held a different cookie. There was no stable thread tying a person's history together. Worse, the thread was deliberately short-lived. Third-party cookies were widely treated as expiring within roughly 90 days, and they were deleted constantly by users and software. When a cookie vanished, the profile attached to it vanished too. You could not build a relationship on an identifier designed to disappear.
Connecting the pieces was lossy. To use a DMP audience in a DSP, the two systems had to agree that a cookie on one side was the same person as a cookie on the other, a process called cookie syncing. It never worked completely. Match rates between platforms commonly landed somewhere in the 40 to 60 percent range, and often lower, so a meaningful slice of any audience simply failed to carry across. Marketers were paying for segments and losing a large fraction of them in the plumbing.
And data leaked. Once a third party could read a publisher's audience through a pixel, that audience could be targeted elsewhere for far less money. The textbook example: a luxury brand advertises on a premium publisher at a high price, captures the readers it reached with a pixel, then finds those same readers on a cheap exchange and serves them again at a fraction of the cost. The publisher's most valuable asset, its audience, walked out the door. Publishers came to distrust the very pixels that fed the DMP machine, and began blocking them.
Underneath all of it sat a quieter problem. The DMP was built for campaigns, not customers. It optimized the next burst of media spend. It had no concept of a customer relationship that lasted years, no view of lifetime value, no memory of a known person across channels. It was an excellent audience engine for an advertising worldview, and a poor fit for any company that wanted to actually know the people it served.
What broke it, and what came next
Two forces turned chronic weakness into terminal decline.
The first was privacy. Apple introduced Intelligent Tracking Prevention in Safari in 2017 and moved to full third-party cookie blocking by 2020. Firefox began blocking third-party cookies by default in 2019. GDPR took effect across Europe in 2018, the California Consumer Privacy Act in 2020, and Apple's App Tracking Transparency in 2021. The Cambridge Analytica scandal broke in 2018, and within days Facebook shut down the third-party data targeting that had been part of the same ecosystem. Each step starved the DMP of its raw material, the third-party cookie and the cross-site tracking around it.
The second force was a change of mind. Marketers grew tired of paying for data they did not own, segments their competitors also had, and platforms that, in the words of one industry account, had been oversold and over-promised. They began investing in first-party data: information collected directly, with consent, from their own customers. Owned, not rented. Tied to a known person, not an anonymous cookie. Kept for years, not 90 days.
The DMP era then ended, and not slowly. Salesforce shut down Audience Studio, the platform built from Krux, with an end-of-life date of February 1, 2024. Oracle closed its entire advertising division, BlueKai included, with support ending September 30, 2024, after revenue fell to roughly 300 million dollars in fiscal 2024, down from about 2 billion dollars two years earlier. The 2020 incident in which a misconfigured BlueKai database exposed billions of web-tracking records had already made the model's risks impossible to overlook.
Something had to replace the rented audience engine, and the answer was a platform built on the opposite principles: owned first-party data, a persistent identity for each known person, a profile that lasts a customer lifetime, and a single record any tool could use. That platform is the Customer Data Platform, and its rise is the subject of part two.
Council summary
This post argues that the DMP was a capable audience engine built on a foundation it never owned: third-party cookies, anonymous profiles, and broker data every competitor could also buy. Its decline was structural, not a single misstep, and privacy regulation plus a marketer shift toward first-party data only finished what the design flaws had started. The named figures were checked against primary reporting: Oracle paid a little above 400 million dollars for BlueKai in 2014, Salesforce bought Krux in a deal valued near 700 million in 2016, and Oracle's ad revenue fell to roughly 300 million dollars in fiscal 2024 from about 2 billion two years earlier. The original draft described that peak vaguely as "in the billions," which was corrected to the verified 2 billion dollar figure. The reader takeaway: rented data can move campaign metrics for a while, but it cannot build a durable customer asset, which is precisely the gap the CDP was created to close.
Comments