Identity Resolution

How identity resolution actually works

Identity resolution is the process of stitching fragmented signals about a person or account into a single, durable profile. A logged-in web session, an email open, a CRM lead, a phone call, and an in-product event may all belong to the same human, but they arrive with different keys: a cookie ID here, a hashed email there, a device fingerprint, a CRM record ID. Resolution reconciles those keys into one canonical entity and assigns a stable identifier that survives across systems and time.

Under the hood, two complementary methods do the work. Deterministic matching joins records on shared, high-confidence keys, most often a hashed email or a known account ID, and it is precise but only fires when a strong key exists. Probabilistic matching infers a likely match from weaker signals, like name plus device plus IP plus behavioral timing, and it scores the likelihood that two records are the same entity. Mature pipelines run deterministic first to anchor the spine, then layer probabilistic linkage to recover the unkeyed long tail without polluting the graph with false merges.

In a warehouse-anchored architecture, this is not a black box vendor decision but a transformation-layer model you own. The identity graph lives as tables in the warehouse, the match logic lives in version-controlled transformation code, and every merge decision is auditable. That ownership matters because identity is the join key that every downstream attribution, segmentation, and revenue model depends on.

Where it breaks in practice

The two failure modes are over-merging and under-merging, and they fail in opposite directions. Over-merging collapses two distinct people into one profile, which corrupts personalization and inflates account-level metrics. Under-merging leaves the same person fragmented across several profiles, which undercounts reach and fractures the customer journey. Probabilistic matching tuned too loosely causes the former; brittle deterministic-only logic causes the latter.

Post-cookie and post-ATT signal loss makes both worse. As third-party cookies and device-level identifiers disappear, the weak signals that probabilistic matching relied on thin out, and pipelines that leaned on them silently degrade. The durable answer is to anchor resolution on first-party identifiers a person knowingly gives you, primarily authenticated email, and to treat third-party signals as decay-prone supplements rather than load-bearing keys.

How we measure resolution quality

Vanity health checks like total profiles count or match rate are misleading, because a high match rate can simply mean you are over-merging. We measure resolution against qualified pipeline: can the resolved profile be traced cleanly from first touch to closed revenue without identity gaps breaking the chain. Precision and recall on a hand-labeled holdout set keep the deterministic-versus-probabilistic balance honest, and merge-and-split audit logs let you reverse a bad decision.

For regulated clients in financial services and healthcare, resolution also has to be compliance-aware. The same graph that unifies a customer must respect consent state and deletion requests, so the identity spine has to carry the legal basis for each linked signal, not just the linkage itself.