Multi-Armed Bandit

The Multi-Armed Bandit element sends a push notification to users, just like the Push element. Instead of a single preset, you provide several content variants. Optimization runs continuously on live data, without manually restarting the test or keeping a fixed traffic split. Pushwoosh shifts more traffic toward the variant that gets opened most.

Multi-Armed Bandit is an experimental feature

• Variants: only existing push presets are supported. Email, SMS, or in-app content variants aren’t supported yet.
• Success metric: only push opens count as a win right now. Choosing a custom event as the success metric is coming soon.
• Selection: based on overall variant performance across all users, not on individual user preferences. Pushwoosh avoids sending the same variant to the same person twice in a row.

Behavior and settings may change as this feature develops.

Add the Multi-Armed Bandit element after the event or condition that should trigger it, then configure the element.

Add content variants

Tip

To use Multi-Armed Bandit, first create two or more push presets in your account. Learn more about creating push presets. You need at least two variants to save the element.

1. Double-click the Multi-Armed Bandit element to open its settings.
2. In Content variants, select at least two different push presets.
3. Click Add variant to include more presets. There is no upper limit.

Choose success metric

In the Reward section, pick what counts as a win for a variant:

• Push opened: a variant wins a send when the user opens that push.
• Event happened: pick a custom event as the success signal instead of opens. This option is coming soon.

Set Reward window to define how many hours Pushwoosh waits after sending before checking whether the push was opened. The default is 2 hours.

Note

If a user opens the push after the reward window closes, the open still appears in your regular message statistics. It does not count toward Multi-Armed Bandit optimization. For learning, that send is treated as not opened. Pick a window long enough to capture the opens that matter. A longer window also means Multi-Armed Bandit waits longer before it can learn from each send.

Set message type

Choose Marketing message or Transactional message:

• Marketing message: For campaigns and promos. All rules and limits apply. Respects: opt-outs, frequency capping, and the silence period. Excludes: the global control group users.
• Transactional message: For operational or service messages. Rules and limits don’t apply. Bypasses: opt-outs, frequency capping, and the silence period. Sent to: the global control group users.

Learn how message type affects delivery.

Configure message delivery settings

Set frequency capping

Note

Frequency capping options are available only when Marketing message is selected. When you choose Transactional message, frequency capping doesn’t apply and these options are disabled.

Use Frequency capping to limit how often users receive push messages, preventing over-messaging and reducing churn. In the Multi-Armed Bandit element settings, choose one of the following options:

• Use Global frequency capping settings

  Apply the project-wide limits configured in your Global frequency capping settings.

  For example, if the global limit is set to 3 messages in 9 days, additional messages exceeding this limit will be skipped.

• Ignore Global frequency capping

  The user will receive this message even if they’ve exceeded the channel’s message limits. Use this option with caution to avoid over-messaging.

• Use custom frequency capping

  Set a custom message limit for this message. If the user exceeds this custom cap, the message will be skipped, and the user will proceed to the next step.

  Important: Custom frequency capping does not isolate the message from Global frequency capping. All messages sent on the same channel, including those from other journeys or campaigns, are still counted toward the global cap. If the user has already received 3 push messages this week from other sources, this message may still be blocked, even if the custom capping would allow it. Learn more

Set send rate limits

The Send rate setting controls how quickly messages are delivered to your audience. Adjusting the send rate helps you manage delivery speed, prevent backend overload, and improve overall deliverability.

Choose one of the following options:

• Use global send rate settings Applies the send rate limits configured in your project’s message delivery settings. If no limits are set, all messages will be sent immediately. Use this option when you want delivery speed to follow your project’s default rules. Learn more about global send rate limits

• Send messages without send rate Sends messages as fast as possible, ignoring any global send rate limits. Use with caution to avoid overloading your backend or creating delivery spikes.

• Use custom send rate Overrides the global send rate for this message only. Allows you to specify the number of messages sent per minute, giving you full control over delivery speed. Messages will be sent at the custom rate you define in the message element.

When you finish configuring the Multi-Armed Bandit element, click Save.

How Multi-Armed Bandit picks a variant

Multi-Armed Bandit doesn’t alternate between variants evenly or run a fixed 50/50 split like A/B/n split. Instead, it continuously watches which variant gets opened more often and gradually sends that variant to a larger share of users. It still sends the other variants often enough to keep learning from real data.

A few things to know about how it behaves:

• It takes some time and volume for a clear leader to emerge. Early on, variants are sent close to evenly.
• Pushwoosh avoids sending the same variant to the same user twice in a row, so people don’t get the exact same message repeatedly.
• Selection is based on the overall performance of each variant across all users in the journey, not on an individual user’s preferences.

Even after a variant is confirmed as the winner, Multi-Armed Bandit keeps running. Unlike an A/B/n split, it doesn’t lock in the winner and route all traffic to it. The winning variant keeps getting the largest share of sends, while the others still receive a smaller share so Pushwoosh can keep learning. If performance shifts over time, Multi-Armed Bandit adapts and can move traffic to a new leader.

Example scenario: Recovering an abandoned cart

Use Multi-Armed Bandit when users add items to their cart but don’t complete checkout. The example below starts a journey on a Cart abandoned event and tests three push variants to bring users back.

1. Prepare push presets. Create three push presets in your account before you build the journey. Learn more about creating push presets.

For this abandoned-cart scenario, you might create presets like these:

• Cart reminder: “Items are waiting in your cart”
• Cart urgency: “Your cart expires in 1 hour”
• Cart discount: “Get 10% off if you check out now”

2. Create the journey. In Campaigns → Customer Journey Builder, click Create Campaign and choose Build a Journey from Scratch. Add a Trigger-based entry element to the canvas.

3. Set the trigger. Create or enable a Cart abandoned event that fires when a user doesn’t complete checkout within a few hours. In the Trigger-based entry element, select this event as the journey trigger. Learn more about events.

4. Add the Multi-Armed Bandit element. Place the Multi-Armed Bandit element after the trigger, then open its settings.

5. Add content variants. In Content variants, select the three presets you created in step 1.

6. Choose the success metric. In Reward, select Push opened. Set Reward window to 2 hours. This window is long enough to catch opens from users who don’t check their phone right away, but short enough to keep learning fast.

7. Finish setup and launch. Configure message type and delivery settings, click Save, and activate the journey.

After launch

Multi-Armed Bandit may behave roughly like this. Actual traffic split and timing depend on send volume and how users respond to each variant.

First, Pushwoosh sends all three variants close to evenly, since it doesn’t have enough data yet. As opens come in, it starts shifting more traffic toward whichever variant gets opened more. For example, if “Get 10% off if you check out now” consistently outperforms the other two, it might receive about 60% of sends while the other two split the rest.

Pushwoosh keeps sending the lower-performing variants too. This keeps learning going and follows the rule that a user who was just sent one variant won’t receive that same variant again the next time they hit this step.

Review results

After you save the element and launch the journey, open the Multi-Armed Bandit element’s statistics panel to see how your variants are performing.

A banner at the top shows the current status of the experiment:

• Not enough data yet: Multi-Armed Bandit hasn’t collected enough sends to say anything meaningful.
• Still learning: there’s data, but no variant is a clear, statistically significant leader yet.
• Significant leader: one variant is confirmed to outperform the others. It’s marked with a Winner badge.

For each variant, you’ll see:

• Sent: how many times this variant was sent.
• Opened: the share of sends that were opened.
• Traffic: the current share of new sends going to this variant.

Multi-Armed Bandit vs. A/B/n split

Both Multi-Armed Bandit and A/B/n Split optimize your messaging, but they work differently. Use the table below to decide which one fits your use case.

	Multi-Armed Bandit	A/B/n Split
Mechanics	Continuously shifts traffic toward the best-performing variant based on live opens	Splits users into fixed percentage branches you define upfront. Branches can hold entirely different downstream flows
Variants	At least two push presets required. No upper limit on variants	Up to four branches per test
Traffic split	Starts close to even, then shifts automatically as data comes in	Fixed percentage you set, stays fixed until a winner is picked
Duration	Runs continuously, with no defined end	Runs until statistical significance is reached or you stop it manually
Winner	Never locks in: the winning variant gets more traffic, but the others still receive some	Locks in after a winner is picked: losing branches deactivate, and all new users route to the winner
Best for	Ongoing content optimization for a single message step (push presets only, for now)	Testing structurally different flows or sequences when you want a one-time, locked-in decision