Read results

Fabric Experiments uses exposure-first attribution: conversions are interpreted in the context of who was exposed to which variant.

Local reports are available through the CLI:

fx report homepage-cta

Hosted reports use the aggregate API:

GET /api/v1/orgs/{orgId}/experiments/{experimentId}/aggregate?metric=signup

The response includes one row per variant with exposure count, conversion count, conversion sum, failure count, and computed timestamp. analytics adds:

• median, mean, and P90 time from first exposure to first matching conversion;
• mean conversion depth plus counts of subjects reaching 1+, 2+, and 3+ events;
• CUPED-adjusted rates and lift when the pre-period covariate has usable variance;
• optional segment rows when segmentKey=<dimension> is supplied.

Studio renders frequentist and Bayesian views together. Bayesian probability uses the documented Beta(1,1) prior; CUPED uses the per-subject conversion count in the configurable pre-exposure window (14 days by default). A missing or constant covariate is reported as unavailable rather than displaying a false precision improvement.

Segment values are never inferred from the full targeting context. Applications must explicitly send a low-cardinality dimension:

createClient(manifest, {
  subjectId,
  analyticsDimensions: { country: 'US', plan: 'team' },
})

Agent readouts

When an experiment reaches a terminal state, audit-forward can dispatch the analyst workflow. The analyst fetches aggregate rows, asks the configured LLM for a short readout, and posts to Slack if FX_SLACK_WEBHOOK_URL is set.