TrackingDesk

Glossary

Identity resolution

Deciding that separate records — a click, a session, an order, an email — belong to the same person. Everything downstream inherits whatever this gets wrong.

Also called: identity stitching, profile unification

You have a click from Tuesday, an anonymous browsing session from Thursday, an order on Friday, and a support ticket next week. Identity resolution is the process of deciding which of those are the same person.

It matters because every number built on top of it inherits its errors. Split one customer into three and your repeat-purchase rate collapses, your acquisition costs look better than they are, and your retention reporting is fiction. Merge three people into one and you have a different set of wrong numbers plus a privacy problem.

ad click click ID anonymous session device, network, timing order hashed email support ticket hashed email one profile everything downstream exact match — deterministic inferred — probabilistic wrong merge = one person's behaviour attached to another's profile
Two kinds of edge, two kinds of failure. A deterministic match is right or absent; a probabilistic one can be wrong while looking identical to a correct match — and every number built on this profile inherits that.

Two approaches, and the distinction is not cosmetic.

Deterministic matching uses something exact: a login, a hashed email, an order reference, a click ID carried through the funnel. When it matches, it is right. It simply fails to match a great deal of the time, because most visitors never identify themselves.

Probabilistic matching infers the link from signals that merely tend to co-occur — device characteristics, network, timing, behaviour. It has far better coverage and it is sometimes wrong, in ways that are difficult to detect after the fact because the output looks identical to a correct match.

The judgement call is which error you would rather have. Under-matching gives you fragmented, conservative numbers. Over-matching gives you clean-looking numbers that are quietly wrong, and it can attach one person’s behaviour to another’s profile — which is a privacy incident, not a data quality issue.

For measurement, prefer deterministic and accept the gaps. For anything that changes what an individual sees, be very sure before you merge.

Do not confuse with

Close enough to get mixed up, different enough that the mix-up costs something.