Hello everyone. I've been following the discussions here for a while, particularly the comparisons between vendors like mParticle, Segment, and Lytics, and I appreciate the depth of analysis I've seen. My background is in ERP and inventory systems, where data accuracy is paramount, so the concept of false positives in identity resolution really caught my attention. In my world, a mis-merged customer record can lead to incorrect credit terms, shipping errors, and a cascade of operational issues.
I've been trying to understand how to practically measure the quality of a Customer Data Platform's identity graph. Many discussions focus on match rates and deterministic vs probabilistic, but I'm concerned about the silent errors: the false merges where two different individuals are incorrectly stitched into a single profile. This seems like it would be critical for B2B e-commerce and manufacturing logistics, where entities might share similar names or domains.
I wanted to propose a method for investigating this using SQL, assuming one has access to the raw stitch tables or the event data with resolved identity IDs. The goal isn't to get a perfect percentage, but to establish a repeatable query pattern that can surface potential problem areas. My thinking is to look for profiles with high entropy or conflicting attributes that should be mutually exclusive.
For instance, one could look for resolved profiles where there are multiple, distinctly different email domains from the same company, which might be normal, but also where there are multiple, conflicting job titles that imply different roles, or where geographic locations (like city/state from IP or shipping addresses) are impossibly far apart within unrealistic timeframes. Another check could be for profiles with an unusually high count of merged identifiers from different device types or browsers over a very short period.
I'm curious if anyone has attempted something similar with their CDP's underlying data model. Specifically, what logical rules or SQL patterns have you found most effective for flagging potential false positives? Are there particular thresholds for time or attribute conflict that have proven useful? I'm also wondering if this type of audit is something that CDP vendors can or should provide as a standard quality report, rather than something we have to build ad-hoc.