Privacy made UA more honest, not harder

Five years on from App Tracking Transparency, the standard account is still one of loss: the industry had precise measurement, Apple took it away, and UA has been harder ever since. Half of that is right. The difficulty was real. What the industry had before 2021 was never accurate measurement, though; it was confident measurement, and those are not the same thing.

My thesis is that the privacy changes stripped out a systematic bias the industry had learned to call truth, and that teams which adjusted properly now make better allocation decisions than they ever did with full device-level data. Plenty of measurement leads would dispute that, and the pain of the transition was real enough. The direction of the argument holds anyway.

What deterministic attribution actually measured

Device-level last-touch attribution answered one question with great precision: which ad this device saw last before it installed. The industry then used it to answer a completely different question, which ad caused the install. That gap is the whole story.

Last-touch rewarded the channels that were good at being last. Retargeting collected credit for conversions that would have happened anyway, and so did high-frequency networks and any placement sitting close to the install moment. It under-credited the channels that did the work earlier, along with the ones that reached genuinely new players who then installed organically. Everybody in measurement knew this in principle. Almost nobody moved a budget because of it, because the numbers were so precise that they felt like facts.

Precision is not accuracy. A ruler that runs consistently 20% short still gives you readings to the millimetre, and deterministic attribution was that ruler; the industry allocated tens of billions against it.

The mechanism: how losing the ruler improved the measurement

When SKAN, and later AAK, replaced device-level data with aggregated, thresholded, delayed postbacks, the precise-but-biased ruler broke. Three things followed that the loss narrative rarely mentions.

First, the fiction became untenable, so teams had to ask the causal question directly. Geo holdouts and incrementality tests stopped being academic exercises, and so did media mix models; all of it turned into operating necessity. Those methods carry wide error bars. They also measure the right quantity, and a rough answer to the right question beats a precise answer to the wrong one every time.

Second, the platforms' own modelling grew more honest about uncertainty. Modelled conversions now arrive with the implicit acknowledgement that they are estimates, and that framing, however irritating it is to work with, sits closer to reality than a deterministic install count implying a certainty it never had.

Third, taking away retargeting's unfair credit reshuffled budgets toward channels that acquire new players. Adjust's March report, which showed the paid-to-organic install ratio up sharply in 2025, fits an industry that spends more on acquisition it can now see is incremental and less on re-engagement it once over-credited. The report itself makes no such causal claim.

The trade-off honest measurement imposes

Calling this a free improvement would be dishonest. Three costs are real and they persist.

Coarse data means slow feedback, and a geo holdout takes weeks to read. A team steering by daily SKAN postbacks is flying with a lagging altimeter, so the attention that used to go into optimisation now goes into waiting.

Small advertisers took the harder hit. Incrementality methods need spend volume before they produce anything readable, so a studio spending modestly cannot run a geo test with meaningful power and inherits the uncertainty without the tools to resolve it. The honesty dividend accrued unevenly, mostly to large teams.

And a new bias moved into the space the old one left. When the platforms model conversions, the platform holds the information advantage, which leaves buyers depending on estimates produced by the party that gets paid. That is a different problem from last-touch bias, and to my mind a more troubling one, though at least it is a visible dependency rather than a hidden one.

A framework for measurement in the honest era

The operating principle: separate three kinds of number and never let one masquerade as another.

Platform-reported numbers are steering signals. Fast, biased toward the platform, useful for spotting change inside a channel, and never evidence of return.

Modelled numbers, whether they come from an MMP, an in-house model or a mix model, are estimates with stated uncertainty. Use them for allocation across channels, and make every allocation decision record the confidence interval it rested on.

Experimental numbers from holdouts and geo tests are the only ground truth going. They cost money and they take time, so save them for the decisions that matter most: whether to keep a channel at all, and whether a large budget shift holds up.

The decision rule follows. Any budget change larger than a set threshold, say 20% of a channel's spend, needs at least one experimental reading from the last two quarters behind it. Smaller moves can run on modelled numbers. Daily adjustments run on platform signals, with everyone clear that those signals steer and do not prove.

An illustrative example

Imagine a casual studio in 2020 putting 40% of its iOS budget into a retargeting network that showed a 300% return on last-touch attribution. After ATT the reported return collapsed, and the team read that as measurement loss.

A geo holdout run later showed the retargeting spend had produced a small fraction of the installs it took credit for; most of those players would have reinstalled anyway. The channel had not got worse. The ruler had stopped lying. So the studio moved most of that budget into new-player acquisition on channels the same holdout showed were incremental, and blended payback improved even while every dashboard looked less precise than before.

The lesson is uncomfortable for anyone who built a career on the old numbers. When a measurement system hands you less data, the right response is not to mourn the data; it is to ask what the old data was really telling you, and to notice how much of the industry's confidence turned out to be a rounding error on a biased instrument. Deterministic attribution never told anyone what caused an install. The privacy era made that gap impossible to ignore, and the teams that treat the gap as their starting point rather than a problem to patch are the ones making the fewest expensive mistakes now.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• How privacy loss changed the job of the creative strategist
• Attribution is an allocation rule, not a truth