A customer may browse anonymously on a website, return in a mobile app, and later log in. When no shared identifier conclusively links those records, probabilistic matching estimates whether they belong to the same person. It compares signals such as names, email addresses, addresses, and behavior patterns, then assigns a likelihood rather than relying on a fixed match rule. section 6.1.1
The choice is a trade-off between reach and certainty:
- Deterministic matching applies defined rules to identifiers; a cascade tries rules in sequence and commonly creates a new record if none succeeds. section 6.1.1
- Probabilistic matching can find links those rules miss, but its less rigid criteria increase the risk of incorrectly joining two people's records. Possible techniques include fuzzy string matching, phonetic algorithms, and models trained on labeled records. section 6.1.1
For customer communications, use deterministic identity resolution as the foundation and add probabilistic matching where a customer data platform (CDP) provides it, particularly to connect anonymous activity to known customers or link digital behaviors. Check how much control and visibility the platform offers: CDP-based probabilistic matching can lack transparency, while a composable architecture can give a data team more control over its deterministic rules and monitoring. These are choices about operational accountability, not just match rates. section 6.1.1 section 6.1.3
Measuring a journey across channels
A shoe retailer wants to understand whether a streaming-TV ad contributed to an app purchase. In a data clean room, it can compare protected identifiers from CTV, website, and app activity. A hashed email supports deterministic matching; an IP address combined with device attributes can support probabilistic matching. The resulting cross-channel view can inform ad spending, while secure data preparation and consistent identifiers remain necessary. section 8.4
