Essay · 12 March 2026
The MMP after IDFA: what still binds an install
When Apple restricted the identifier for advertisers, many teams spoke as if measurement had ended. It had not. The join simply changed costume.
A mobile measurement partner still sees a click, a fingerprint-ish bundle of signals, a SKAdNetwork postback, and sometimes a logged-in user who later opens the app. Those are four different strengths of glue. Treating them as one “attributed install” column is the original sin of most growth decks we read in the Signal Mapping Lab.
Probabilistic matching did not disappear; vendors became shy about naming it. IP, device language, OS version, and click-to-install time still produce a score. The score is useful as a remainder after deterministic joins fail. It is not a person. If your finance file treats the remainder as equal to a click ID match, you are arguing with yourself.
SKAdNetwork is the opposite problem: it is official and coarse. A conversion value is a budget of bits, not a journey. Teams that copy a US template onto a United Kingdom grocery app often encode “tutorial complete” in the same band as “first paid order,” then wonder why media looks efficient on weekends when tutorials spike.
What still binds an install, in order of how willing we are to defend it: (1) a signed-in customer ID that existed before the campaign, (2) a publisher click identifier that the MMP stored and the app echoed, (3) a SKAN postback mapped to a published value scheme, (4) modelled remainder, labelled as such. Everything else belongs in an appendix.
The practical move is a weekly census. Export events. Count how many installs sit in each glue bucket. If modelled remainder grows after a privacy prompt change, do not “optimise creative.” Rewrite the source-of-truth paragraph so the board hears the glue weaken in plain English.