This article explains what happens when your experiment ends and how to roll out the winning CPP across your campaigns. It also covers how to continue optimizing after you’ve identified a winner.
How The System Determines A Winner
Why We Use A Bayesian Model
CPP experiments run inside the Apple Ads auction, where traffic is not evenly distributed, and daily performance can naturally fluctuate. On top of that, the auction may favor one variation over another based on predicted conversion rates. This creates noisy, real-world data rather than clean, controlled samples.
To handle this, CPP Experiments uses a Bayesian statistical model designed for uncertainty and auction-based environments:
✅ Continuously updates results: the model recalculates probabilities every day as new data comes in. You always see the most current signal instead of waiting until the end of the test.
✅ Handles uneven traffic: If one variation receives more impressions than another (which is common in Parallel experiments), the model automatically accounts for this and evaluates each variation based on the data it actually receives.
✅ Adapts to noise: when results fluctuate, the model stays conservative and waits for more data. When the signal is strong and consistent, it can arrive at a result more quickly.
✅ Makes early signals useful: probabilities shown early in the experiment are statistically valid and help you understand direction, not just final outcomes.
✅ Easy to interpret: for example, “Variation B has a 78% probability of winning” simply means there is a 78% chance it is the best performer on your primary metric.
How A Winner Is Confirmed
The system does not choose a winner just because one variation is ahead. A winner is confirmed only when all of the following conditions are met:
Enough data collected: each variation has reached the minimum required data volume for the primary metric
Enough time passed: minimum runtime is met (7 days for Parallel, 14 days for Switch/Dayparting) to capture weekday and weekend behavior
High probability: the leading variation must reach a high enough confidence level
Stable performance: the advantage must hold steady over time, not just show up briefly
If any of these conditions are not met, the experiment keeps running. If it ends before meeting them, the result is marked as inconclusive.
Why Experiments Take Time
User behavior naturally varies from day to day. For example, a page with a true 5% Download Rate might show 4.3% one day and 5.7% the next. This is expected.
The Bayesian model is built for this reality: it waits for more data when results are noisy and moves faster when the signal is strong and consistent. The estimated duration shown during setup is a best estimate based on historical data, but experiments can finish earlier if the difference between variations becomes clear sooner.
What Happens When An Experiment Ends
When an experiment is completed or stopped, the system automatically restores your setup:
Test ad groups are paused: they stop serving immediately
Original ad group is re-enabled: it resumes running with its original settings, including targeting, bids, keywords, and audiences
Automations are restored: all previously paused rules, Bid Strategies, Budget Allocation, Keyword Expansion Strategies, and CPP Scheduler actions are turned back on
✏️ Note: This happens no matter the outcome. Whether there is a winner, the result is inconclusive, or there is not enough data, your original ad group is always restored.
Applying And Managing The Winning CPP
Once you identify a winning product page, you can roll it out to one or multiple ad groups directly from the Ads dashboard. Find the ad with the winning CPP, then select Actions > Add as Multiple New Ads. For additional ways to create and manage ad variations using the winning CPP, see the 📖 Creating and Editing Ad Variations article.
In SplitMetrics Acquire, you can also use the CPP Scheduler to manage rollouts at scale. The CPP Scheduler helps you plan and coordinate CPP updates across ad groups over time. Once you’ve identified a winner, you can use it to:
Roll out the winning CPP to selected ad groups at a scheduled time
Plan seasonal CPP updates based on experiment results
Coordinate CPP changes across multiple storefronts
You can access the CPP Scheduler from the Optimization Hub in the left sidebar.
Post-Experiment Considerations
Re-Running An Experiment
If your result is inconclusive or there wasn’t enough data, you can run another experiment on the same ad group. The previous experiment must be in Stopped or Completed status before the ad group becomes eligible again.
When planning your next run, it may help to adjust a few things:
Longer duration: if the experiment didn’t collect enough data to reach a clear result
Higher-traffic ad group: if limited traffic slows down data collection
Different primary metric: if the selected metric requires more volume than your current setup can support
Testing A New Hypothesis
A winning CPP in one ad group may not perform the same way in another. Audience behavior, keywords, and storefront differences can all affect results. Because of this, it’s often useful to run separate experiments for:
Different storefronts (countries or regions)
Different campaign types (brand, competitor, or generic)
Different audience segments
To keep track of your overall testing progress, you can use the CPP Testing Coverage metric in the CPP Experiments dashboard, which shows how many of your CPPs have been tested.
Next Steps
You’ve successfully scaled your winning CPP. Next, learn how to manage multiple experiments across your portfolio.
Next article: Manage Your Experiment Portfolio