Vishnu Vardhan Reddy Kaithapuram takes a look at why data governance is no longer a back-office concern, but the only game in town for advertisers...

For most of the last five years, the advertising industry told itself a tidy story about identity. The third-party cookie would be deprecated. Mobile identifiers would fade. In their place, browsers and platforms would ship privacy-preserving primitives that handed measurement back to advertisers in a cleaner, safer form. A great many strategy decks were written on that assumption. 

The story did not play out that way. Chrome reversed its plan to remove third-party cookies and now leaves them on by default. Most of the measurement machinery promised under Privacy Sandbox has been wound down or retired for low adoption. And yet the identifier still eroded, just not on anyone's published roadmap. Safari and Firefox have blocked third-party cookies for years. Apple's tracking prompt hollowed out mobile identity. Consent rates fell, and a widening patchwork of privacy law kept moving the legal ground under everyone's feet. 

So here is the situation engineers actually inherited. The clean central replacement never arrived. What arrived instead was fragmentation, partial signal, and a measurement problem that no browser vendor was going to solve on our behalf. I come at this from a particular angle. My work has centred on data governance, which at heart asks two questions about any number a pipeline produces. Can you trust it, and do you know where it came from? Those used to be back-office concerns. In a post-identifier world they are the whole game. 

The first thing that changes is that aggregation stops being a final step and becomes the starting point. The older generation of reporting systems assumed a spine of stable identity. You logged an event per user, kept the rows, and joined them later to whatever question you wanted to answer. When the identity spine is unreliable, that design quietly breaks. Aggregation-first systems flip the order. You commit to cohorts, time windows, and counts early, and you accept that you will not be able to drill back to the individual to repair a number afterwards. 

The second thing is that data is both late and modelled, and naive systems treat it as neither. Conversions do not arrive when the click does. They trickle in across attribution windows, through delayed post-backs and offline matches, sometimes days after a report has been declared finished. Worse, when observation is blocked by a missing consent or an opted-out device, the gap gets filled by estimation, so a modern conversion count is part observed and part inferred. This is exactly where silent errors live. Having spent my career on data lineage and anomaly detection, the discipline built to catch these failures, I have come to think the fix is not a better dashboard but a system that knows where every number came from and notices when one starts to behave abnormally. Provenance and automated error detection are no longer governance niceties. They are load-bearing parts of any measurement system you intend to trust. 

That leads to the tradeoff nobody enjoys stating out loud. Privacy-preserving measurement spends a budget. Whether the mechanism is added noise, minimum cohort thresholds, or suppressed low-count cells, the rule is the same. The finer you slice, the more the protection costs you, either as larger error per slice or as data you are not permitted to show at all. You cannot promise an advertiser infinitely granular breakdowns and a strong privacy guarantee in the same breath. The honest system makes the cost visible rather than dressing estimates up as precision. 

This is where buyers have real leverage, and most are not using it. The right questions are not about dashboards. They are about how a number was produced. How much of this figure was observed and how much was modelled. What is the attribution window, and do numbers restate as late conversions arrive. At what cohort size do you stop reporting, and what happens to the data below that line. Is noise being added, is it disclosed, and could two queries be differenced to back out an individual. A vendor who can answer those calmly is doing the work. A vendor who treats the questions as rude is selling confidence rather than measurement. 

None of this is a counsel of despair. Measurement after the identifier is not less rigorous, it is differently rigorous. The mental model that holds up is to stop reading these numbers as counts and start reading them as estimates with provenance. Every figure carries a story about how much of it was seen, how much was inferred, how late it might still change, and how much protection was spent to release it. The teams who have built governance and lineage systems already think this way out of necessity. The commercial side benefits from thinking that way too, because in a post-identifier world the most valuable thing a measurement system can tell you is not just the number. It is how the number was made. 

Vishnu Vardhan Reddy Kaithapuram is a Software Development Engineer at Amazon Advertising in New York, writing in a personal capacity. He holds an MS in Data Science from the University at Buffalo, and his research focuses on machine learning for automated data governance. He has spoken at the Marketplace Risk New York Conference.