Multi-Armed Bandit Testing: Maximize Conversions During Traffic Spikes
Standard A/B tests split traffic evenly between variations and wait for statistical significance before declaring a winner. That works well for long-running tests on evergreen pages. It works less well when you have a short window of high-value traffic and need to extract as much conversion value from it as possible.
Multi-Armed Bandit mode is built for those situations.
What is Multi-Armed Bandit testing?
Multi-Armed Bandit is an advanced testing approach that shifts traffic toward the best-performing variation while the test is still running. Instead of maintaining a fixed 50/50 split until the end, the algorithm monitors conversions in real time and progressively sends more visitors to whichever variation is converting better.
A small percentage of traffic always continues exploring the other variations, so the system keeps learning. But the majority of your traffic goes to what is actually working right now, not after the test ends.
The result is more conversions during the test itself, not just after it finishes.
When to use it
Multi-Armed Bandit mode makes the most sense when:
- You are running a short-term, high-traffic event like Black Friday, Cyber Monday, or a seasonal sale
- A campaign or piece of content is driving a temporary traffic spike and you want to capitalize on it quickly
- You are testing secondary elements like footer CTAs or download buttons where speed matters more than long-term precision
For most ongoing split tests on evergreen pages, standard Bayesian testing is still the better choice. It is more rigorous and gives you cleaner long-term data. Bandit mode is for situations where the traffic window is short and you cannot afford to wait.
How traffic allocation works
Traffic starts split evenly across all variations. As data accumulates, the algorithm begins weighting traffic toward the better performer sometimes pushing 70 to 90 percent of visitors toward the winning variation while keeping a small exploratory allocation on the others.
The test does not end at a fixed point. It keeps steering traffic toward what is working in real time. If one variation pulls ahead early, most of your visitors will see it. If the picture changes, the allocation adjusts.
Key advantages
No manual intervention required. The algorithm adjusts traffic automatically as data comes in. You do not need to check results during the event or make any changes.
More conversions during the test. Standard tests sacrifice conversions on the losing variation until the test ends. Bandit mode reduces that loss by shifting traffic quickly.
No special technical setup. AB Split Test handles the statistics and allocation automatically. You enable the mode, set your goal, and let it run.
Works during your highest-value traffic windows. Black Friday traffic at 2am is not traffic you want to be running a 50/50 test through.
How to activate Multi-Armed Bandit in AB Split Test
- Go to WP Admin → AB Split Test → Settings → Advanced Settings
- Enable Dynamic Traffic Optimization (Multi-Armed Bandit)
- Save Settings
- Create a new test or open an existing one
- Choose your test type
- Set your primary conversion goal
- Scroll down to the Winning Mode section
- In the Optimization Style dropdown, select Dynamic — Multi-Armed Bandit
- Publish your test
- Monitor real-time results including conversion rates, uplift, and traffic weight columns
- Let the algorithm run for the duration of your campaign
Should every test use Bandit mode?
No. For long-running tests on your main pages like hero sections, pricing pages, product descriptions, use standard Bayesian testing is the right approach. It is more precise, gives you cleaner statistical significance, and produces results you can rely on for long-term decisions.
Bandit mode is for tests where speed and real-time adaptation matter more than textbook significance. High-traffic windows, promotional campaigns, and anything where the opportunity is time-limited.
Available on
Multi-Armed Bandit mode is available on the Ultimate plan.