Metrics
A metric is what a number means in your project — defined once, so two tests measuring “Trial Started” measure the same thing.
What a metric is
A metric is one definition of a number, stored on the project: which event, which kinds of purchase, how long a conversion still counts. Without it, every test carries its own private answer, and two of them can report "Trial Started" from two different definitions.
Where you create one
On a test, not on the Metrics page. In a test's settings the Judged on picker lists every live metric except retention — which can never be a primary, see below — and ends with ＋ Define a new metric…, so you are there at the moment you know what the goal means. The Metrics page is the list, and where you archive and restore.
The five kinds
| Kind | What it counts |
|---|---|
| Finishes the flow | Reaching the end of the flow, dismissals excluded |
| Buys something | A purchase, optionally narrowed by kind |
| How much they spend | The same purchases, as an amount rather than a count |
| An event your app sends | One named event, matched exactly as your app passes it to track() |
| Comes back on a later day | The install returned on day 1, 7 or 30 after first seeing the flow |
Event metrics
Type the name exactly as your app passes it to track(). The match is string equality and it is case-sensitive: trial_started and Trial Started are two different events. An event metric cannot be saved without a name — a goal that names nothing can never match.
Purchase and revenue
A transaction is one of four kinds. Leave the filter off and all four count.
trial— an entitlement began without money, usually an amount of zero, which is why a trial goal cannot be written as "a purchase over zero".paid— money arrived for the first time on this subscription, including a trial converting.renewal— money arrived again on a subscription already running.upgrade— the subscriber moved to a different product.
Retention
Day 1, 7 or 30 — a closed set, not any number. It is measured from each install's own first exposure to the flow, never a calendar date, so an install first exposed on a Friday and one first exposed on a Monday get the same amount of time. Only installs whose day has arrived are in the denominator. Retention can never be the metric a test is judged on: a day-7 number is a week behind everything else on the panel. Add it under Also watch.
The conversion window
How long after first seeing the flow a conversion still counts. It runs per install, from that install's own first exposure. The default is seven days and 1 to 30 is accepted; the create form offers 1, 3, 7, 14 and 30.
Pick it from the shape of the conversion. Finishing onboarding happens in one sitting; starting a trial can wait for a payday. A conversion that cannot happen inside the window at all reports zero in every version, forever. Longer catches the slow ones and makes every reading staler — nobody is counted until their own window shuts.
Retention takes no window. The day it names is its window, and a second one is refused.
Goals on a test
A test has one primary metric — Judged on — and at most four under Also watch. Secondaries are reported beside the result and never decide it.
Editing does not move a result
When a test launches, each goal's definition is copied into it by value, with the name it had at that moment. A running test keeps the definition it launched with, and finished results keep reading. Renaming is not a change to what a metric measures, so fixing a typo never claims the definition moved.
Archive, never delete
There is no delete. Archive takes the entry out of every picker, so no new test can be judged on it; nothing already measured changes. Archived entries sit behind Show archived, and Restore saves the same definition again — it comes back with its key, its revision and its history. There is no delete, for the same reason audiences have none.
Related
A metric measures a flow served at a placement to an audience, and the events it counts arrive from the SDK. Delivery failures are health, not a metric.