configuring-experiment-analytics

Configuring experiment analytics

This skill answers: Who is included in the analysis? and How to measure impact?

Exposure criteria

Exposure criteria determine which users are counted in the experiment analysis.

Include people when

Two options:

1. Feature flag called (default) — users are included when the $feature_flag_called event fires for the experiment's flag. This is the standard approach — it means a user is included only when they actually encounter the feature flag in your code.
2. Custom exposure event — users are included when a specific custom event fires. Use this when you want tighter control over who enters the analysis (e.g., only users who actually visit the page where the experiment runs).

Multiple variant handling

When a user is exposed to multiple variants (e.g., due to flag changes or race conditions):

• Exclude multivariate users — removes these users from the analysis entirely. Cleaner data, smaller sample.
• First seen variant — assigns users to the first variant they were exposed to. Keeps all users in the analysis. Note that "first seen" can introduce other biases as behavior cannot be clearly attributed to a single variant and is not recommended unless necessary.

Bias risk on uneven splits. "Exclude multivariate users" combined with an uneven variant split can introduce bias — multi-variant users are dropped asymmetrically and the smaller variant loses a larger fraction of its assignments. If those users behave differently from the rest, the smaller variant's metrics will be skewed.

The right mitigation depends on experiment state:

• Not yet launched, or only exposed to a few users so far — switch to an even variant split and use the overall rollout percentage to limit test-variant exposure. This removes the bias and preserves statistical power. See configuring-experiment-rollout.
• Live experiment with significant exposures — changing the split mid-run reassigns users across variants, which is bad for user experience and data quality. Switch this setting to "First seen variant" instead — it keeps already-assigned users in their original variant (no reassignment) and removes the asymmetric exclusion.

Filter test accounts

exposure_criteria.filterTestAccounts (default: true) — excludes internal/test users from the analysis.

Resolving experiments

Metric changes require an experiment ID. If the user refers to an experiment by name or description (e.g. "add metrics to the checkout test"), load the finding-experiments skill to resolve it to a concrete ID before proceeding.

Metrics

Metrics are added via experiment-update after creation. The metrics array replaces the entire list, so always get the current experiment first via experiment-get to preserve existing metrics.

Step 1: Discover available events (REQUIRED — always do this first)

Before suggesting or configuring ANY metric, you MUST call read-data-schema to discover what events actually exist in the project. Do NOT skip this step. Do NOT suggest event names based on what you think the project might track — only use events you have confirmed exist.

This applies even when:

• The user provides event names — look them up to confirm they exist and are spelled correctly
• The user asks "what metrics do you suggest?" — look up events first, then suggest from real data
• The context makes certain events seem obvious — they may not exist or may be named differently

Workflow:

1. Call read-data-schema to get the project's events
2. Present relevant events to the user based on the experiment's hypothesis
3. User picks which events to use for metrics
4. Configure metrics with those confirmed event names

Legitimate exception — allow_unknown_events: true: Pass this on experiment-create / experiment-update only when the user is intentionally instrumenting an event that hasn't been ingested yet (e.g. setting up the experiment before the code change ships). Confirm this with the user — never use it as a workaround for "the event lookup didn't return what I expected".

Example:

User: "Let's add some metrics for the checkout experiment"

WRONG: "I'd suggest using purchase_completed as the primary metric..."
  (hallucinated event name — never seen the project's actual events)

RIGHT: *calls read-data-schema* → "Here are the events in your project
  related to checkout: `checkout_step_completed`, `payment_processed`,
  `order_confirmed`. Which of these represents a successful checkout?"

Step 2: Choose metric type

There are four metric types. Each has kind: "ExperimentMetric":

Funnel metrics and the implicit exposure step

Funnel metrics automatically prepend the experiment's exposure event as step_0. So a funnel with 1 step in series is a valid 2-step funnel: exposure → action. This is the correct choice for measuring "what percentage of exposed users did X?"

Examples:

• "What % of exposed users reached /login?" → funnel with 1 step ($pageview filtered to /login)
• "What % of exposed users completed checkout?" → funnel with 1 step (checkout_completed)
• "What % of exposed users went cart → checkout → purchase?" → funnel with 3 steps

Mean vs funnel for the same event

• Mean measures average count/value per user (e.g. "pageviews per user", "revenue per user").
• Funnel measures conversion rate (e.g. "% of exposed users who purchased").

Both can reference the same event — the difference is whether you care about count/magnitude (mean) or yes/no conversion (funnel).

See references/metric-configuration.md for the full rendered ExperimentMetric schema (all four metric types, with required fields per type) plus WRONG/RIGHT JSON pairs for the failure modes that come up most often (ratio with is_set filter instead of math: "sum" + math_property; retention without retention_window_start / start_handling). Read it before assembling a ratio or retention payload — the required fields are authoritative.

Step 3: Primary vs secondary

• Primary metrics — the main success criteria for the experiment. These drive the ship/end decision.
• Secondary metrics — additional measurements for context. Useful for guardrail metrics (e.g., ensuring a conversion improvement doesn't increase error rates).

Interpreting results

See references/interpreting-results.md for guidance on reading experiment results, statistical significance, and when to ship vs end.