Deterministic versus probabilistic matching

The difference in one table

What counts as a deterministic match

A match is deterministic when the same value travels from the ad click to the installed app. Three carriers are common.

• The store referrer. On Android, Google Play passes the app a referrer URL from the click that led to the install.^[1] The Play Install Referrer guide covers it in full.
• A device advertising identifier. If the ad platform and the app can both read it, the two records join exactly. On iOS, reading it needs the user's permission through App Tracking Transparency.^[2]
• A click ID. A value created at the click and handed to the app, usually inside the link that opened it.

How a probabilistic match is made

When no exact value survives the trip through the store, the MMP compares what it saw at the click with what it sees at first open. The signals are ordinary ones that many devices share. No single signal identifies a phone. Together, within a short time, they can make one click the likely source.

Time is the control that matters most. Two people on the same network with the same phone model look identical to this method. The longer the gap between click and open, the more such lookalikes pile up. That is why a probabilistic lookback window is kept much shorter than a deterministic one.

Platform rules decide where the method may be used at all. Read the current store policies before you rely on it, and see what you can still measure without the IDFA for the iOS side.

How an MMP ranks the two

An MMP typically runs them in order for each install: deterministic first, probabilistic only when no exact match exists. A statistical match does not outrank an exact one under this ordering. The full ranking, with clicks and views, is in how last-touch attribution picks a winner.

Why the exact value goes missing

A deterministic match fails for ordinary reasons. The person did not allow the identifier to be read. A redirect or link shortener in front of the store link may have dropped the referrer. The click happened on one device and the install on another. In each case the install is real and the click is real, and only the thread between them is gone. Reports often label each install with its match type, so you can filter one type out and see how much of a channel rests on inference.

What to do next

Open your attribution settings and find the window for each match type. Airbridge Core, for one, lets you configure windows per channel and match type, down to minutes.^[3] If your tool has a separate probabilistic window, keep it short. Then break last month's installs down by match type. A channel that leans heavily on probabilistic matches deserves a closer look before you raise its budget.