What probabilistic matching still does after IDFA

Probabilistic matching is a family of guesses: IP, user agent, timestamp proximity, sometimes a little more. Before ATT, many teams treated it as a backup singer behind IDFA. After ATT, some treated it as a replacement lead vocal. That promotion was a category error.

A guess can still be useful. If a click and an install share a narrow time window on a rare device configuration, you may be looking at the same human. If they share a carrier-grade NAT in Bangkok at lunchtime, you may be looking at a thousand humans. The method did not change. The prior did.

Where it still earns its keep

Android-heavy catalogues in Thailand still see substantial GAID coverage on opted-in traffic. Probabilistic methods then fill the cracks: OEM browsers that strip parameters, delayed installs after a shared family device, web-to-app hops that never carried a click ID. Used as a minority method, it can close gaps without pretending to be a census.

We also see it help in fraud review. A cluster of “matches” that are too clean — identical timing, identical UA strings, rotating IPs in a tight ASN — is often a farm. Paradoxically, probabilistic signals are sometimes better at showing industrial behaviour than at naming a single customer.

Where it becomes theatre

Any report that quotes probabilistic match rates to one decimal place without a confidence band is selling calm. Ask for the features used, the collision rate on your traffic, and whether the vendor’s model was trained on US iOS behaviour. A model that learned suburban LTE will misread dual-SIM Android traffic on BTS-adjacent networks.

Jurisdictions matter. Some probabilistic techniques sit awkwardly beside consent language you already showed the user. If your privacy notice promised “no fingerprinting,” do not quietly re-enable it because SKAN feels incomplete. That is not a measurement decision; it is a legal one. Talk to counsel before you talk to the MMP.

How we teach it in the lab

In the Attribution Measurement Lab we treat probabilistic credit as a labelled uncertainty, not a second last-click column. Students have to write a sentence a director can read: “We assign these installs as probable matches with an estimated collision risk of X; we do not use them for bidder optimisation.” If they cannot write the sentence, they cannot ship the number.

If you want the longer version with workbook exercises, the flagship outline is the place. For a Bangkok-specific window discussion, see the Thailand note.