Choosing the unit of randomisation decides what counts as one independent observation in a test.
The common units are users, sessions, page views, devices, stores and regions.
The right choice depends on what the change affects and how users experience it.
User-level assignment is the usual default, because each person sees one consistent version.
It needs a stable identifier, such as a login, since a cleared cookie can put the same person into a different group.
Session or page view assignment gives more units and faster tests, but suits only changes users will not notice between visits.
Units can affect each other, which breaks the test: a new feature in a chat app changes the experience of friends in the control.
This spillover is called interference, and it usually makes the measured effect too small.
Larger units, such as whole cities or friend groups, keep the spillover inside each unit.
The cost is fewer units, so tests on cities or stores need longer runs or larger effects.
Marketplaces are a classic case, since a change that helps buyers in treatment can take sellers away from buyers in control.
The analysis must treat each unit as one observation, not each page view inside it.
Counting page views as independent when users were assigned overstates precision and gives too many false positives.
In practice, teams record the unit in the test plan and check that every metric is calculated at that level.
The usual mistake is assigning by user but working out standard errors as if every page view were independent.