Identity Resolution Errors: When Two Customers Become One Segment Member
Every segment in a store is built on a quiet assumption: that the system knows which records are the same person. Identity resolution, matching emails, devices, accounts, and orders into one customer, is where that assumption goes to die. Merge two different people into one profile and every segment touching that profile is wrong in a way no amount of clever segment logic can fix.
How bad merges happen
Shared devices and shared emails do most of the damage. A family tablet, a couple using one email for orders, a workplace address used by a whole team: deterministic matching on email alone welds those people into a single customer with bizarre behaviour, three sizes, five addresses, contradictory preferences.
Probabilistic matching adds its own errors at scale. Similar names at the same address, a parent and adult child with the same name in the same house, get merged by systems tuned for match rate. Vendors report resolution accuracy in aggregate, which hides the fact that errors concentrate exactly where households share attributes, which is a large share of a consumer base.
Symptoms you can see in segment data
Look for impossible customers: order counts no human sustains, category affinities that span unrelated departments with no basket ever mixing them, a single profile subscribed and unsubscribed on the same day, return rates that blend a careful buyer with a serial returner at one address.
Segment level, merges inflate your best segments. A merged household reads as a high frequency, high value customer, so champions segments quietly fill with households rather than people, and per customer value reporting drifts away from what finance counts. Nobody notices until a campaign to champions produces response patterns that make no sense per person.
Splitting is harder than merging
Most systems merge eagerly and split reluctantly, because splitting requires deciding which events belonged to whom with incomplete evidence. That asymmetry means errors accumulate in one direction, and the correction tooling, where it exists at all, is manual and slow.
Given that, the rational policy is conservative merging: require strong keys for a merge, an account login, a verified email plus a second identifier, and accept some duplicate profiles as the cheaper error. A duplicate splits a customer's history across two records and understates their value. A bad merge fabricates a person who does not exist and acts on the fabrication.
Audit before you build on top
Before any segmentation programme launches or relaunches, sample the merged profiles with the highest event counts and read them by hand. Fifty profiles, an hour, and you will know whether your identity layer describes people or households with good PR.
Publish the known limitation next to the segment definitions. A team told its segments are household level for a third of the base makes different decisions from a team told its data is person level truth.
Keep reading
Related guides on this site: Lookback Windows: The Quiet Setting That Changes Who Is in Every Segment How Small Can a Segment Be Before It Is Just Noise Wearing a Name Practical answers about personalization.
Frequently asked questions
What is identity resolution?
The process of matching records from different systems, devices, and channels into a single customer profile. Segments, lifetime value, and personalization all depend on its accuracy.
What is the most common identity error in ecommerce?
Merging different people who share an email or device, especially within households. The merged profile shows impossible behaviour and corrupts any segment it joins, often inflating high value segments.
Are duplicate profiles or bad merges worse?
Bad merges are usually worse. A duplicate understates one customer's value across two records, while a bad merge fabricates a nonexistent person and drives real actions, offers, and suppressions from the fabrication.