Campaign Experiments

Experiments let you test campaign changes with controlled comparisons before committing to them at scale. An experiment holds a change in a comparison window until you have enough signal to promote, extend, or roll back. Open Experiments in the sidebar.

Experiment Types

Campaign experiments come in four types:

Geographic Holdouts

Withhold a campaign from specific geographic areas to measure its true incremental impact, then compare performance in exposed versus holdout regions. Geographic holdouts require a guided setup — you define test and control areas before the experiment can be created.

Budget Splits

Allocate budget across two or more arms to see which delivers better results. The platform tracks spend and outcomes for each arm.

Creative Tests

Compare different ad creatives (copy, images, formats) within the same audience and budget.

Audience Tests

Test different audience segments against each other to find the highest-performing targeting.

Creating an Experiment

Most experiments are created directly on the Experiments page:

1. Open Experiments in the sidebar
2. In the Create experiment section, enter a name and choose the type
3. Click Create
4. Configure the arms, hypothesis, and stop conditions in the experiment's tabs

Geographic holdouts can't be created from this quick form — they start from the guided setup that captures test and control areas.

When creating an experiment you can register a hypothesis (a statement, primary metric, and minimum detectable effect) and stop conditions (max duration, max spend, minimum sample size, or a metric threshold) so the experiment has a clear definition of done before it starts collecting data.

Experiment changes are governed, not just authorized: creating an experiment, launching it (the draft→running status change), recording an observation, selecting a winner, and deleting an experiment each run through guardrail evaluation, and the high-impact ones (create, winner selection, delete) file a change set for the audit trail.

Viewing Results

Each experiment detail page is organized into tabs (variants, results, activity, and — for geographic holdouts — geographic assignment). The results tab shows:

• Status — draft, running, paused, completed, or archived
• Arms — the variants being compared with their allocation
• Per-arm metrics — observed performance for each arm
• Significance — whether statistical significance has been reached, the current p-value, and the sample size still needed when the data is inconclusive

The platform records observations against each arm and will not report significance the underlying data hasn't actually reached.

Managing Experiments

From an experiment's detail page you can:

• Pause a running experiment
• Resume a paused experiment
• Roll out winner — promote the winning arm once you have enough signal (manual, or automatic when you configure auto-winner with a confidence threshold)
• Edit the experiment's configuration
• Delete the experiment

Experiments are a building surface and sit in the sidebar's Build group, even though they're documented here under Review. Note also that experiments span more than campaigns — agent-configuration and governance-rule experiments exist as well — but the four types above are the campaign-facing ones.