One Customer, Three Records: Identity Resolution Errors Poison Segments Quietly

Guest checkout, a work email, a phone number as login: the same person becomes three customers, each looking lukewarm, and every segment built on top inherits the fracture.

Before a store can segment its customers it has to decide who its customers are, and that decision is quietly wrong in most databases. The same person checks out as a guest in March, creates an account with a personal email in June, and orders from a work email in November. Three records, each with one order, each classified as a one time buyer of modest value. Merged, they are a loyal repeat customer in your top decile who has never once been treated like one, because no system ever saw her whole. Identity resolution, the matching of records that belong to the same person, is unglamorous infrastructure with glamorous consequences: RFM scores, lifetime value, churn models, and lookalike seeds all consume identity as an input and faithfully amplify its errors. A store with a fifteen percent duplicate rate does not have slightly wrong segments. It has a systematically understated loyal base, an overstated one time buyer problem, and marketing that keeps paying acquisition prices to reacquire people it already has.

Match on the strong keys first

Resolution quality is a hierarchy problem. Exact email matches are strong but miss every multi email customer. Phone numbers are strong second keys, shared less often than addresses and typed at checkout constantly. Name plus address matching catches households and movers but risks merging family members into one chimera customer who buys for three people and responds to nothing personally. The defensible order is deterministic keys first, email, phone, account ID, then probabilistic matches with a confidence threshold and a human review band, and never the reverse. Probabilistic first feels modern and produces confident nonsense at scale, because a model that merges aggressively will always find the database tidier and the customers stranger.

Merges should be reversible. Every merge decision kept with its evidence, so when the chimera surfaces, a support agent reporting that Mrs. Alvarez does not recognize half her order history, the records can be split back along the recorded seams. Systems that merge destructively turn a data error into a permanent customer experience error.

Related reading on this site: SegmentSage and Your Segments Overlap. The Customer in Five Audiences Gets Five Stories. cover the neighbouring problems in more depth.

Measure duplication like a business metric

Put a number on it and report it beside the metrics it corrupts. Sample based audits work: take a few hundred multi order email addresses and count how many appear under several records, or track the share of orders attached to records created the same day as the order, the guest fracture signature. When leadership watches the duplicate rate next to the repeat purchase rate, something clarifying happens: improving resolution often moves the repeat rate more than a quarter of retention campaigns do, because the loyal customers were there all along, filed under three names.

Prevention beats cleanup and lives at checkout. Account prompts that recognize a returning email and offer the existing record, guest checkout that still captures a durable key, phone as a first class identifier where customers accept it, each new fracture avoided is a merge nobody has to risk later. The cheapest identity resolution is the duplicate that never gets created.

Households are the honest hard case

Some shared signals are genuinely shared lives: partners ordering to one address, parents buying for adult children at the same house, a family phone plan number on four accounts. Resolution rules that treat the household as one customer destroy personalization for everyone in it, gifts become recommendations, surprises become spoilers, a real harm in categories built on gifting. The mature answer is two level identity: person records for communication and preference, household linkage for delivery, fraud, and value views, with the levels kept distinct in the data model instead of collapsed by whichever query ran last. It is more work than one customer table. It is also the difference between a database that describes your customers and one that describes your shipments.

Frequently asked questions

What is identity resolution in ecommerce?

It is the process of recognizing that multiple records, a guest checkout, accounts under different emails, belong to the same customer, and merging them so history, value, and preferences are complete. Without it, loyal customers fragment into several apparently one time buyers.

What are the best matching keys for merging customer records?

Deterministic keys in order of strength: account ID, exact email, and phone number. Name and address matching can follow with care because households share them. Probabilistic matching should sit behind a confidence threshold with a review band, and merges should be recorded so they can be reversed.

How can a store measure its duplicate customer rate?

Through sample audits, such as checking how many known multi order customers appear under several records, and by tracking signals like the share of orders placed on records created that same day, which indicates guest checkout fracture.

Why not merge everyone at the same address into one customer?

Because households contain distinct buyers whose preferences, gifts, and communications should stay separate. The sound model keeps person level records for marketing and a household level link for delivery and value views, rather than collapsing them.