A secondary metric is a supporting metric that provides additional context about an experiment's results but doesn't drive the ship decision. Secondary metrics help you understand why the primary metric moved (or didn't), what else changed, and whether the observed effect has implications beyond the immediate hypothesis.
An experiment might have one primary metric, two guardrail metrics, and a dozen secondary metrics. The secondary metrics don't need to show a statistically significant result for the experiment to succeed. They're diagnostic. They tell you what happened beneath the surface.
What role do secondary metrics play in experiment analysis?
Secondary metrics answer the questions that come after the ship decision.
When the primary metric improves, secondary metrics explain the mechanism. A checkout experiment that increases completed purchases might show, through secondary metrics, that the improvement came from reduced cart abandonment at step three rather than from more users initiating checkout. That distinction matters for what you build next.
When the primary metric doesn't move, secondary metrics can reveal why. The change might have shifted behavior in ways that canceled each other out: more users started the funnel, but fewer completed it. The net effect on the primary metric is zero, but the secondary metrics show that something real happened. That learning informs the next iteration.
When a guardrail metric regresses, secondary metrics help diagnose the cause. A drop in session length might be explained by changes in a specific user segment or product surface, visible only in the secondary metrics.
How should secondary metrics be treated statistically?
Because secondary metrics don't drive the ship decision, they don't need the same false positive protection as success metrics. Confidence's decision framework, described in "Risk-Aware Product Decisions in A/B Tests with Multiple Metrics," applies multiple testing corrections to success metrics (protecting against false ship decisions) and adjusts false negative rates for guardrail metrics (protecting against missed regressions). Secondary metrics typically sit outside both correction families.
This doesn't mean secondary metric results are meaningless. It means you should interpret them as directional evidence rather than as rigorous conclusions. A secondary metric showing p = 0.03 in a set of twenty uncorrected secondary metrics is interesting context, not proof. Treat it as a hypothesis for a future experiment, not a finding you'd bet the product on.
How many secondary metrics should an experiment track?
There's no hard upper limit, but there's a practical one: the more metrics you track, the more noise you have to sift through when interpreting results. A team that adds fifty secondary metrics to every experiment will spend more time reading dashboards than making decisions.
Confidence runs analysis inside your data warehouse, so the computational cost of additional metrics is low. The cognitive cost is what you're managing. Choose secondary metrics that would change your interpretation of the primary result if they moved significantly. If a secondary metric wouldn't alter any decision regardless of its result, it's not worth tracking in the experiment.