Reading results
What each part of the results page shows, from the verdict and group table to data quality and the chart.
The results page leads with the verdict and three key numbers, then shows each goal as a table, data quality checks, and a chart under Details. People with access.readResults can read it; by default that is anyone with access.manage.
Results are computed by your scheduled results route, so the first results appear after its first run following Start. Recompute asks for fresh results for one test straight away.
Rounds
Each start of a test is a round, and results are per round. Select a round at the top. A decided test opens on the round it was decided from, and shows the numbers recorded at the decision everywhere on the page. A rollout round (after deciding a test with a holdout) shows "Since shipping" results for the winner against the holdout.
The verdict
The verdict is one sentence and, when there is one, a recommended action, such as "Variant A is statistically better: +12% (CI +4% to +20%)." A guardrail line follows when the test has guardrails, for example "No harm detected so far" or "Too little data to rule out harm." All verdicts are listed in Verdicts.
Goal tables
Each goal has a table with one row per group:
| Column | Meaning |
|---|---|
| People | People exposed to this group in this round |
| Conversions | People with at least one outcome |
| Rate or Average | Conversion rate for yes/no goals; average per person for counts and revenue |
| Lift | Relative change against the Original |
| Confidence interval | The range the true lift is likely in. With more than one variant, the main goal's interval is adjusted for the number of comparisons |
| p-value | Sequential method only |
| Chance to win, Risk | Bayesian method only |
The main goal decides the verdict. Supporting goals are labelled "Supporting (not used for the decision)". Guardrails show OK, Watch, Too little data or Failing. The holdout is shown as "Measured, not compared" during the test.
Revenue goals show the uncapped total alongside the capped decision statistic. If fewer than 20 people had the outcome in the 90 days before Start, no cap is applied and the page says so.
Data quality
Data quality lists anything that makes the results less trustworthy:
| Check | Shown when |
|---|---|
| Traffic split | The people counted per group do not match the configured split (sample-ratio mismatch) |
| More than one group | People who saw more than one group were excluded |
| Mismatched beacons | Browser exposures that did not match the assigned group were ignored |
| Identity conflicts | Some linked identities were seen in different groups |
| Too few people | The compared groups have not reached the minimum yet |
| Layer room | The layer has fewer slices reserved than the requested allocation |
| Page and goal checks | A version page stopped resolving, a placement could not be read, or a click goal is missing from some versions |
A traffic-split problem, or excluded people reaching 1%, turns the verdict into "Data problem". See Data quality.
The chart
Under Details, the chart shows the lift and confidence interval over time for each variant. Content-change markers show when an editor published a change to tested content, so you can see whether results moved afterwards.
Results for a goal that could not be read
If a goal's data source fails or times out, that goal is shown as unavailable for this run with a reason, and the next results run tries again. If the main goal is unavailable, the verdict is "Results unavailable".
The Tests home
The Tests home groups tests by what needs doing (needs attention, running, ready to start, draft, decided, archived), with a one-line status and health label for each. Draft tests name their next setup step.