# Read results (/docs/experiments/results)

Open an experiment to see its results. The top shows the **verdict**, the **recommendation** and the exposures. Below that are the primary metric per variant, exposures per day, segments, secondary metrics and **Summary · what agents see**. Results always cover every exposure since launch.

The examples below come from a test of an annual-plan discount banner: `control` vs `test`, 50/50, with "upgraded within 14 days" as the primary metric.

## The numbers [#the-numbers]

| Number | Meaning | Example |
| --- | --- | --- |
| **Exposures** | People counted in each variant | 1,475 vs 1,525 |
| **Rate** | Conversion rate, or average per person for mean metrics | 30.8% vs 39.1% |
| **Lift** | How much higher than control, relatively | +26.9% |
| **Interval** | The range the true lift falls in with 95% probability | +15.0% to +39.8% |
| **P(beats control)** | Chance this variant truly beats control | 100% |
| **P(best)** | Chance this variant is the best of all variants | 100% |

Read it as: *test converts 39.1% of exposed trials vs 30.8%. The true lift is very likely between +15% and +40%, and there's essentially no chance control wins.*

A wide interval means you know the direction but not yet the size. With three or more variants, a variant can beat control and still not be the best, so the verdict uses P(best).

## Verdicts [#verdicts]

| Verdict | When |
| --- | --- |
| **Needs more data** | A variant has fewer than 100 exposures. |
| **Winner** | A variant has at least a 95% chance of being best. |
| **Keep control** | Control is best with 95% probability, or every variant has at most a 5% chance of beating it. The change doesn't help. |
| **Inconclusive** | None of the above, yet. |

"Keep control" is a clear result, not a failure. Variants with fewer than 100 exposures are marked *small sample*.

The verdict and recommendation use only the primary metric. Secondary metrics get their own verdicts. With several secondary metrics, one may look like a winner by chance, so treat a surprise there as an idea for the next test.

## Split check [#split-check]

If you set 50/50 and got 70/30, something is broken: maybe the flag is read on only one code path, a redirect loses events in one variant, or bots are getting in. Datahop checks the split. When it's off, you see a warning banner and the recommendation tells you to check your implementation before trusting the results. Below 50 exposures the check is skipped.

## How long to run [#how-long-to-run]

For conversion metrics, the recommendation estimates how many more days you need to detect your minimum lift at today's traffic:

> Keep running — roughly 6 more day(s) to reach 3,609 exposures per variant.

Treat it as a guide, not a stopping rule. Before you decide, give every exposed person their full metric window, and run at least a full week if behaviour differs by weekday.

## Segments [#segments]

An average can hide very different stories. **Segments** show the primary metric again for each [cohort](/docs/people/cohorts) you pick, with its own lift and verdict. Add up to 10 with **Add cohort segments** when you create the experiment, or **Edit segments** at any time, even after it ended.

| Segment | Control | Test | Lift | Verdict |
| --- | --- | --- | --- | --- |
| Everyone | 30.8% | 39.1% | +26.9% [+15.0%, +39.8%] | Winner |
| New users | 17.9% | 28.9% | +61.2% [+27.5%, +102.8%] | Winner |
| Power users | 58.7% | 62.2% | +5.9% [−6.7%, +20.6%] | Inconclusive |

The banner works, and almost all of the effect comes from new users. Power users already upgrade at 59%, so a discount for them may cost revenue without changing much. A sensible decision: ship it, then target the banner flag at people *not in* Power users.

The overall verdict tells you *whether* it works. Segments tell you *for whom*.

<Callout type="warn" title="Pick segments the test can't change">
  Segments use cohort membership as it is when you look, not when people entered.

  - **Time windows move.** "Signed up in the last 7 days" means as of today, so after two weeks the people who were new at launch have left the segment.
  - **Outcomes can leak in.** If a cohort is defined by the thing you're measuring (say, *upgraded users*), the test itself moves people into it.

  The cleanest segments are fixed before the test starts: plan, country, platform, signup month.
</Callout>

## Conclude [#conclude]

When you've decided, click **Conclude**. Pick the winner (the dialog suggests the one the results point to), write the reason, and leave **Roll out winner to everyone** on or turn it off. Then click **Ship**.

With roll-out on, everyone gets the winning variant right away through the same flag, so your `getFeatureFlag(…) === 'test'` code keeps working until you clean it up. The decision, the reason and who made it are saved on the experiment.

A good reason mentions the segments, for example: *"+26.9% upgrades, driven by new users (+61%). Power users flat. Follow-up: exclude Power users from the banner."*