For about eighteen months after ATT launched, every conversation about mobile measurement was the same conversation. The deterministic data is gone. SKAdNetwork is broken. Probabilistic attribution is a mess. Modeled conversions are vendor magic. Nobody knows what they're doing.
Most of that was true. It was also a transitional state, not a permanent one. The industry has now had enough time to settle into a working set of approaches. They are not as good as the pre-ATT setup but they are good enough to actually run businesses on. This piece is about what that actually looks like in 2025.
SKAdNetwork is mostly figured out
SKAN 4 changed the math enough that the worst complaints from the SKAN 2/3 era do not really apply anymore. You get conversion postbacks across multiple windows. You can configure conversion values with reasonable granularity. The privacy threshold issues that nuked smaller campaigns are still real but better understood.
What teams have actually figured out is the configuration discipline. SKAN works when you have a clear conversion schema mapped to actual business value, not when you try to recreate your full event taxonomy in 64 buckets. The teams getting good results from SKAN are the teams that decided what they actually need to measure and configured for that, rather than trying to preserve the pre-ATT measurement surface.
It is also worth saying out loud: SKAN is not the only signal. It is one signal among several. Treating it as the source of truth was always a mistake. Treating it as one input into a triangulated view of campaign performance is much closer to the actual practice now.
Probabilistic attribution did not die, it just got narrower
Probabilistic attribution at the user level largely went away with ATT and is mostly not coming back. What did not go away is probabilistic measurement at the campaign and channel level, which is a different thing.
The MMPs that survived the transition all do some form of media-mix modeling now. The data goes in at the channel and campaign level. The output is an estimate of marginal contribution. This is not new methodology — MMM has been around for decades — but the application to mobile UA at the daily and campaign level is relatively new and has improved a lot in the last two years.
Teams that have built MMM into their UA workflow can run more aggressive testing and reallocation than teams that are still trying to make individual click attribution work. The MMM users do not know which click drove which install, but they know which campaigns produce incremental volume and at what cost. For making decisions, that is what you actually need.
Incrementality testing is finally normal
Incrementality testing was always the right answer. For years it was something only the top-percentile teams ran because the operational overhead was real. That has changed. The MMP platforms have built incrementality testing into their products. Several specialist tools have made the test design and read-out cheap enough that mid-tier teams run it routinely.
If your team is not running incrementality testing on at least your largest channels, you are operating on attribution data that probably overstates your ROAS by some unknown amount. The amount varies by channel and creative but it is rarely small. Teams that actually measure incrementality typically discover their attributed ROAS overstates incremental ROAS by twenty to fifty percent on the largest channels.
This is not a vendor pitch. The tooling is widely available and the operational overhead is much lower than it used to be. The reason more teams do not do it is mostly cultural. Performance marketing as a discipline is built around attributed ROAS as the success metric. Switching to incremental ROAS as the success metric requires changing how teams report up. Some teams have made that change. Many have not.
Modeled conversions are the new normal
Both Apple and Google ship modeled conversions in their attribution surfaces now. The networks layer their own modeling on top. The MMPs combine all of it.
The result is that pretty much every conversion you see in your dashboards is modeled to some extent. The question is no longer "is this real or modeled" but "how much modeling and what's the model doing."
The teams that have adapted well are the teams that treat modeled conversions as estimates with confidence intervals rather than as ground truth. They look at modeled numbers across multiple sources, compare them, and form a view based on the convergence or divergence rather than picking one number to trust.
The teams that have adapted badly are still treating one source's modeled numbers as truth and trying to reconcile other sources to that one. This produces endless reconciliation work and a false sense of precision.
What the working stack actually looks like
For a typical mobile app team running serious paid UA in 2025, the working measurement stack looks something like this:
An MMP for the unified attribution layer, set up to ingest SKAN postbacks and channel-level cost and event data.
SKAN configuration that maps a small number of conversion buckets to actual business value, not an attempt to preserve full event tracking.
An MMM tool, either inside the MMP or standalone, that produces channel-level marginal contribution estimates updated weekly or daily.
Regular incrementality tests on the largest channels, run as standard practice rather than as ad-hoc validation.
An internal data warehouse that pulls all of this together and lets the team look at modeled numbers from multiple sources side by side rather than picking one source to trust.
This is not a particularly exotic stack. Most teams have most of these pieces. The teams that get good results have all of the pieces working together rather than in isolation.
What I think the next phase looks like
The measurement landscape will probably keep tightening. Apple has not finished restricting fingerprinting. Google's privacy sandbox for Android is going to land in some form. The browser-side privacy work continues to bleed into mobile patterns even where the technical implementation differs.
Teams that have built a measurement practice around triangulation across multiple imperfect signals will be fine through whatever comes next. Teams that are still hoping for some signal to come back to where it was pre-ATT are setting themselves up for repeated disappointment.
The settled state of post-IDFA measurement is not amazing. It is workable. For most teams that is the realistic ceiling and the teams that have accepted that ceiling are running better practices than the teams that are still in denial about it.
