Identity Resolution: The Quiet Foundation of Segment Accuracy

Segments are only as accurate as the identities beneath them. When one person exists as four partial profiles, across devices, emails, and guest checkouts, segments overcount, undercount, and misclassify all at once. Identity resolution, matching those fragments into one person with known confidence, is the unglamorous work that makes every segment above it trustworthy.

Counting fragments, calling them people

A typical customer appears in a database as several ghosts: a phone profile from mobile browsing, a laptop profile from a work email, a guest checkout with a typo'd address, and a loyalty account created in store. Segmentation built on fragments miscounts everything. Reach inflates, frequency caps fail because each ghost looks like a stranger, and behaviour splits so no single profile crosses the thresholds that define your best segments.

For more on this, see this related guide and our earlier explainer on the same theme.

The visible symptoms are familiar: customers targeted as prospects while holding loyalty accounts, win back campaigns landing on people who bought last week under another identity, and segments whose sizes never match the finance team's customer counts. Teams debug the segments for months before suspecting the layer underneath.

Deterministic, probabilistic, and honest confidence

Deterministic matching links fragments on shared hard identifiers: the same email, phone, or account login. It is precise and limited, because fragments often share nothing hard. Probabilistic matching adds softer evidence: device patterns, addresses, names, behaviour timing, and produces a match with a confidence score rather than a certainty. Both are legitimate; pretending the second is the first is where accuracy dies.

Keep confidence attached to the identity. A segment built for a high value offer should draw on high confidence identities only, while broad awareness segments can tolerate looser matching. When every identity is treated as equally certain, teams either over restrict everything or, more commonly, trust merges they should not, and two real people get fused into one confident fiction.

Where matching goes wrong

Shared identifiers are the classic trap. Families share emails and devices, students share addresses, and a phone number gets recycled to a new owner. Aggressive matching fuses households into a single super customer with incoherent tastes, and segmentation then targets that fiction with total confidence. Split errors and merge errors both hurt, but merge errors hurt more because they corrupt two customers' data at once and are harder to detect.

Readers also found this companion piece from our archive useful.

Watch for the household question explicitly. For many businesses the household, not the individual, is the real buying unit, and resolving to household level on purpose is a strategy. Resolving there by accident, while believing you have individuals, is a bug. Decide which unit your segments describe and test the matching against that decision.

Auditing identity accuracy

Sample and inspect. Pull a hundred resolved identities and read them like case files: do the fragments plausibly belong to one person or one household? Measure obvious incoherence, mismatched names, impossible geographies, purchase patterns that alternate between two lives. That manual audit, repeated quarterly, catches what dashboards smooth over.

Also track resolution metrics over time: profiles per known customer, segment size deltas after matching changes, and complaint signals like customers seeing another person's recommendations, the canary of merge errors. Identity resolution is infrastructure. It earns its budget by making every segment, report, and personalisation decision above it slightly less fictional.

Frequently asked questions

What is identity resolution in segmentation?

It is the process of recognising that multiple partial profiles, across devices, emails, and checkouts, belong to the same person or household, and merging them with a known confidence level so segments count real customers instead of fragments.

What goes wrong when identity resolution is poor?

Segments overcount and misclassify: loyal customers get prospecting ads, recent buyers get win back offers, frequency caps fail because each fragment looks like a stranger, and segment sizes stop matching any other count in the business.

Which is worse, merging two people or splitting one?

Merging is usually worse. A split wastes some targeting efficiency; a bad merge corrupts two customers' data simultaneously, produces incoherent profiles, and can surface as one customer seeing another's information, which is a trust incident.

How can we audit matching quality?

Manually review samples of resolved identities for plausibility, track profiles per known customer and segment size shifts after changes, and monitor complaints about wrong personalisation. A quarterly case file review catches errors dashboards hide.