Exposure and denominators
An exposure is one person seeing one answer, even when the rollout counts accounts. Read every rate on this page before you compare a rollout's numbers to an experiment's.
Rolling out. The account rollout unit is part of the current release and is not on every project yet. Until it reaches yours, a feature flag rolls out by person and the question below does not appear when you create one.
An exposure is recorded when a person is served a value. One person, one exposure, on every experience type.
That stays true when the rollout counts accounts. Ten people in one treated account are ten exposures, not one.
Why the unit and the denominator are different questions
They come apart on an account-unit flag, and it is the only place they do:
| Counts | |
|---|---|
| The rollout unit decides who is in | accounts |
| The exposure decides what you divide by | people |
So a flag rolled out to 10% of accounts can easily produce 40% of your exposures, if your large customers happen to land in the treated slice. Both numbers are correct. They are answers to different questions.
A rate computed over exposures is a per-person rate, always. If you want a per-account rate you have to aggregate to the account yourself. Intempt does not silently switch denominators on you, and it will not guess which one you meant.
Why exposure is not counted per account
Because every downstream rate you care about is per person. Conversion, activation, click-through, revenue per user: all of them divide by people. If exposure were counted per account, every one of those rates would quietly change meaning the day you switched a flag's unit, and the number would still be labelled the same thing on the same chart.
Keeping the denominator fixed means changing the unit changes who is served and never what your metrics mean.
What to check when a number looks wrong
Served share does not match the rollout percentage. Expected on an account unit. The percentage is of accounts; the share you are looking at is probably of people. Large accounts skew it up, and people with no account skew it down.
Two people in one company got different answers. That should be impossible on an account unit. Check whether both actually resolve to the same account: a user attached to no company is decided separately from one who is.
Exposures are lower than your active user count. Only people who were actually evaluated are exposed. Somebody who never hit the code path that calls the flag is not an exposure.
Comparing a rollout to an experiment
You can, but know what you are comparing. An experiment's subject is the device, a people-unit rollout's subject is the person, and an account-unit flag's subject is the company. Three different populations, one shared exposure denominator.
That shared denominator is what makes the comparison possible at all. It is also why you should say which unit was in force when you quote a rate from either one.
When no account resolves
An account-unit flag serves the off value to anyone Intempt cannot place in an account. It does not fall back to deciding per person, and this page explains why that is the safe answer.
Introduction to Analytics
Analytics turns the events you track into reports and dashboards: build Insights, Funnel, and Retention reports on user or account behavior, then combine them onto a dashboard to monitor your KPIs.