Experiment Results

The Experiment Results table

Once your experiment is running and your data source is connected, GrowthBook can compute statistical results. Navigate to your experiment and click the Results tab.

What the table shows

Each row in the results table is one metric. Each column after the first is a variation. The cells show:

Column	What it means
Users	Number of unique users assigned to this variation
Value	The raw metric value for this variation (e.g. conversion rate, average order value)
Chance to Win	(Bayesian) Probability this variation is better than the baseline
Uplift	Relative change vs the control, e.g. +3.2%
Confidence Interval	The range of plausible true uplift values at the configured confidence level

Reading the result: a worked example

Suppose your experiment has two variations — Control (0) and New Checkout (1) — and you are tracking Purchase Rate (a proportion metric):

• Control: 4.8% purchase rate, 10,000 users
• New Checkout: 5.2% purchase rate, 10,050 users
• Uplift: +8.3%
• Chance to Win: 87%
• 95% CI: +1.1% to +15.8%

This means: if you ran this experiment many times, 87% of the time the New Checkout would beat the Control. The true effect is most likely between +1.1% and +15.8%. This is promising but not conclusive — most teams wait for chance to win above 95% before shipping.

Refreshing results

GrowthBook does not poll your warehouse continuously. To see updated numbers:

1. Click Update Data in the top-right of the Results tab.
2. GrowthBook runs the analysis queries against your warehouse and updates the table.
3. You can also schedule automatic refreshes in the Data Source settings.

Guardrail metrics

In addition to your primary goal metrics, you should configure guardrail metrics on every experiment. Guardrail metrics are metrics that must not regress — for example, page load time or revenue per user.

1. In the experiment settings, scroll to Metrics.
2. Add the metric under Guardrails instead of Goals.
3. If a guardrail metric shows a statistically significant negative change, GrowthBook flags it in red — even if your goal metric is positive.

Dimensions

GrowthBook can slice results by a dimension — a categorical attribute of your users (e.g. country, device type, plan tier). To use dimensions:

1. Define a dimension SQL in the Data Source settings. The query must return user_id and the dimension value column.
2. In the Results tab, select the dimension from the Dimension dropdown.
3. GrowthBook re-runs the analysis segmented by each dimension value.

Dimensions are useful for detecting heterogeneous treatment effects — for example, your new checkout may help mobile users but hurt desktop users.

🧪 Wait for sufficient sample size before trusting results

Looking at results early and making decisions when sample sizes are small is called “peeking” — it inflates the false positive rate significantly. Let your experiment run to the planned sample size or duration before drawing conclusions. Use GrowthBook’s sequential testing option if you need to peek safely.

What does "Chance to Win" mean in GrowthBook's Bayesian results?

The p-value of the experiment The probability this variation is better than the baseline, given the observed data The percentage of users who converted in this variation The confidence interval width

How do you refresh the experiment results with the latest warehouse data?

Reload the browser page Click Update Data in the Results tab Re-run the experiment assignment query manually Reconnect the data source

What is the purpose of a guardrail metric?

It is the primary goal the experiment is optimizing for It is a metric that must not regress — it flags regressions even when goal metrics look positive It controls the traffic split between variations It sets the minimum experiment duration

What is "peeking" in the context of A/B testing?

Looking at results early and making decisions before the planned sample size is reached Viewing the assignment query SQL before running the experiment Using dimensions to segment results Checking guardrail metrics after the experiment ends

Submit

Mark as complete