Read experiment results and check variations

Checked against the product’s code on 7 November 2026.

Start with an existing experiment. Use the experiment guide if you need to create one or review its locked plan.

1. Open Optimization.
2. Open Experiments.
3. Open the experiment.

Read Results, Timeline and Configuration & QA together. A chart alone is not the decision to promote a winner.

Read the result under its plan

Results explorer shows the available saved result and verdict.

1. Review the primary measure.
2. Review its interval, the estimated range around the reported result.
3. Review the safety measures.
4. Check whether traffic split as planned.

Missing or immature evidence can prevent a verdict. A scheduled saved result is not proof that the latest job succeeded.

An analysis view can show fixed horizon, Bayesian, sequential or adjusted calculations. Compare methods on one exact cohort compares methods on the same supported group of visitors. It does not give permission to choose whichever result looks best. The locked plan governs the decision. Only its precommitted deciding method can authorize promotion. For new plans, each guardrail must rule out harm beyond its accepted limit; simply finding no significant harm is not enough. Earlier locked plans keep their recorded checks.

Use exploratory views carefully

The metric explorer and attribute comparisons help investigate results. The CUPED panel uses eligible data from before each visitor first saw a variation to adjust its reading, and compares that with the unadjusted result. Timeline views help inspect how outcomes change with days since exposure.

Treat slices and method comparisons as exploration. A favorable slice does not override the locked primary decision, insufficient evidence or a sample ratio warning.

Refresh an older experiment’s saved results

Editors and Admins can ask the copilot to recompute an older experiment that has no locked measurement plan. A locked plan refuses this request; read its saved Results instead.

Use the screen

Editors and Admins can also do this on the screen.

1. Open the older running or stopped experiment’s details.
2. Choose Recompute legacy results when it is offered. It is offered only for an older, unnamed experiment without a locked plan that is running or stopped.
3. Read the refreshed result or the error that names the experiment.

This screen action needs no approval card. It refreshes the saved result and the timeline. The partial-write warning below still applies if the request fails.

Ask the copilot

1. Ask: “Recompute the saved results for [experiment name].”
2. Read the returned report or refusal.
3. Check whether it is still collecting and when it was computed before relying on it.
4. Reopen the experiment’s Results to review the saved reading.

This rereads stored assignments and goal conversions and writes saved measurements and the daily summary. Traffic and public content stay as they are. It needs no approval card and has no receipt Undo. A failure can follow a partial measurement write, so a failed request does not prove that nothing changed. See Recompute experiment results.

Keep a usual Results view

For an experiment with a locked measurement plan, you can save your preferred method and a single method or side by side view.

Use Analysis lens

1. In Results, open Change how this is read.
2. Find Analysis lens.
3. Choose an available method from its method grid.
4. Under View, choose Focused method or Compare methods.
5. If Save this as my default is offered, choose it.
6. Check for Saved as my default. If it was already shown, the choice already matches your saved view.

This saves your view for this experiment. Use the site-default request below to change the site’s usual view.

Ask the copilot

1. Choose a supported method from the experiment’s available Results methods.
2. Ask the copilot: “Save [method] with a side by side comparison as my usual Results view for [experiment name].” Name a single method view instead if you prefer it.
3. Read the saved receipt.
4. Reopen Results without a shared view link to use that preference.

Editors and Admins can also ask: “Set [method] and a single method view as this site’s usual experiment Results view.” This sets a site default. A personal preference or shared view link can be more specific; an incompatible method falls back visibly.

Neither save changes traffic, the locked plan, figures, thresholds or recorded decisions. These saves need no approval card and offer no Undo. Save the earlier choices again to put them back. See Results preferences.

Check the actual experience

QA means checking that the intended variation is shown before relying on its results.

Configuration & QA shows the lock and current configuration, including any differences it flags.

The selected variation did not appear in the supplied walk described below. Treat this as a check to verify, rather than a working preview guarantee.

1. Open Configuration & QA.
2. Find QA preview. It needs a published page carrying this experiment.
3. Choose the Variation.
4. Choose Open forced preview. The preview is intended to show that variation on a published, bound page with a QA marker. In the supplied demo-store walk, the selected Treatment served the live content without that marker, even after consent and a reload. Verify both the variation and the marker before treating a view as a QA check. Correctly marked forced QA exposures are intended to be excluded from experiment results. Eligible participation also depends on visitor identity and personalization consent.

If you promote a winner, the approved action stops the experiment and copies the winning version into each page’s draft. From that moment every visitor sees the default version until you publish those drafts, which is a separate act. One Undo puts back both the pages and the running split. Continue with the experiment guide for setup, activation, stopping and promotion.