Since ATT and the deprecation of cross-app identifiers, most ad platforms have leaned hard into modeling. When a network cannot observe an install or a purchase directly, it estimates one. Modeled conversions are statistically reasonable in aggregate, but they were never designed to answer the question app growth teams actually ask every morning: which specific campaign, ad set, and creative produced this paying user?
What modeled conversions actually measure
A modeled conversion is a probability distributed across many users. The network sees an install it cannot attribute, looks at the broader population of clicks and impressions, and assigns fractional credit based on historical patterns. That works fine when you are reporting quarterly spend to a board. It falls apart the moment you try to make a granular optimization decision.
The core problem is resolution. Modeling answers "roughly how many conversions did this channel drive" but cannot answer "did creative B outperform creative A for high-LTV users in Germany last week." The signal you need for daily UA decisions lives below the resolution that modeling can honestly provide.
Device-level attribution keeps the join intact
Device-level attribution preserves a real, deterministic link between a touchpoint and an outcome on the same device. When a user taps an ad, lands on the store, installs, and later converts, MakeRVN stitches those events through a privacy-safe match that never leaves the device unencrypted. The result is a 1:1 record, not a fractional estimate.
- Deterministic install-to-event joins, not population averages
- Creative- and placement-level granularity that survives down-funnel analysis
- Cohort retention and revenue you can trace back to the original click
- A defensible source of truth when a network's numbers and yours disagree
If you cannot trace a paying user back to the creative that won them, you are not optimizing UA, you are decorating a forecast.
When modeling still earns its place
This is not an argument to throw modeling away. Modeled conversions are useful for filling the unobservable tail, for smoothing sparse data on small campaigns, and for cross-checking your deterministic numbers. The mistake is treating a model as ground truth and then making sharp, irreversible budget cuts on top of it.
A practical hierarchy
Treat device-level attribution as your system of record, use modeled conversions to estimate what you genuinely cannot observe, and reconcile the two on a fixed cadence. When they diverge by more than a few percentage points, investigate the gap instead of averaging it away. That gap is usually where your most interesting learnings hide.
Growth teams that win in the privacy era are not the ones with the cleverest model. They are the ones who kept the most real signal and refused to optimize against noise.