Without a unique identifier like the IDFA, you are left matching patterns — device model, OS version, time window, IP range — to judge that a click and an install were probably the same person. That approach is probabilistic attribution.
Why it exists
Before Apple's App Tracking Transparency (ATT), a unique identifier could link a click and an install with certainty. Once users can decline tracking, that ID often isn't available, so probabilistic matching became a fallback way to estimate the connection.
Deterministic vs. probabilistic
Deterministic matching: a unique ID links click and install with certainty. Accurate, but requires user consent.
Probabilistic matching: contextual signals suggest a likely match without consent. Less privacy-invasive, but can be wrong.
Why to be careful
Mismatched attributions from probabilistic matching can inflate or deflate a channel's apparent performance. This is part of why channel-attribution methods that don't need individual tracking — like MMM — are getting renewed attention.
In place of a unique identifier it combines contextual signals — device model, OS version, screen resolution, timezone, IP range — to estimate the probability that a click and an install belong to the same person. Each signal alone is shared by huge numbers of people, so the method leans on how rare the combination is.
How does it differ from deterministic matching?
Deterministic matching links one to one through a unique ID like the IDFA: accurate, but it requires user consent. Probabilistic matching estimates without consent and is never 100% accurate. Mismatched conversions inflate or deflate individual channels' apparent performance.
How much should you trust probabilistic results?
Treat them as directional rather than as grounds for ranking channels against each other. Accuracy varies with traffic characteristics, so the error does not land evenly across channels. Cross-check against your own server-side conversions or holdout results before deciding anything.