Build a recommendation

Amplitude Activation lets you create recommendations for personalization campaigns. Recommendations can increase engagement, reduce churn, and create cross-selling opportunities.

Build a new recommendation

To create a new recommendation, open Cohorts & Audiences and click Recommendations in the left rail. Click Create Recommendation and follow these steps:

1. Select the type of recommendation to create:
   • Top Trending: Generates a list of items with the highest increase in popularity over a specific period. Use this type to identify new and emerging trends.
   • Most Popular: Creates a list of the most popular items based on usage across all users. Use this type to highlight content or products that are trending and in high demand.
   • AI Based: Uses an AI-based model to provide personalized content to each user. This type accounts for the user's past behavior and preferences to curate a tailored list of items. This recommendation type is the most sophisticated and can significantly improve user engagement and conversion rates.

2. Define your outcome. The outcome is the goal you want to reach or the metric you want to improve. In the Define starting cohort section, choose the cohort this recommendation trains its algorithm on. By default, Amplitude selects all users active in the last 90 days. For best results, use a more narrowly tailored cohort. To narrow the cohort, select conditions such as performed events and filters. If you're defining a Top Trending or Most Popular recommendation type, skip to step #6.
3. For an AI Based recommendation type, create your item catalog by choosing the items you want to recommend to reach your outcome. Under Define items to be recommended, click Select event... to choose the exposure event.
4. Click Select property... to designate the item to recommend to the user. You don't select the item itself. Instead, choose an event property associated with the exposure event. The recommendation chooses the recommended item from the values of this event property.
5. Choose the items to recommend in the Current list of items that will be recommended section. By default, Amplitude Activation chooses from the 50 most frequent values of the event property you selected in the previous step, based on 30-day uniques.

You can also configure your recommendation to use a static list of property values. Toggle Create with Static List and select the candidates from the list of options. You can exclude specific values from a dynamic recommendation and choose to exclude converted items. Click Next > to continue to the Save section. Skip to step #11 to complete your new AI Based recommendation.

6. For a Top Trending or Most Popular recommendation type, use the Select time range of trend section. Specify the time frame for the user to complete the outcome event. The default setting is the past seven days. The Top Trending type includes an offset option with a default of two days.

7. Under Define your outcome, choose the outcome event for this recommendation within a specified time frame. The default time frame is one hour for the outcome. Click Next > to move to the Items tab.
8. Create your item catalog by choosing the items you want to recommend to reach your outcome. Under Define items to be recommended, click Select event... to choose the exposure event.
9. Click Select property... to designate the item to recommend to the user. You don't select the item itself. Instead, choose an event property associated with the exposure event. The recommendation chooses the recommended item from the values of this event property.

For example, a music app might want users to buy concert tickets from within the app. The app might show users who followed a playlist a concert popup, based on the genre of the playlist they followed. In this case, select the genre event property attached to the follow_playlist event. More generally, the event property often has a name like SKU, ID, or Name.

10. Specify the number of items to recommend to each user. By default, Amplitude Activation chooses from the 50 most frequent values of the event property you selected in the previous step, based on 30-day uniques.

You can also configure your recommendation to use a static list of property values. Toggle Create with Static List and select the candidates from the list of options. You can exclude specific values from a dynamic recommendation as well. Then click Next > to move to the next step.

11. In the Save tab, give your new recommendation a name and description.
12. Use the Control slider to configure the percentage of users to include as a control. Amplitude randomly selects these users to receive a random set of items as a recommendation. Use this control group to measure the lift this recommendation generates. To exclude the same control group across multiple recommendations, click Link to an Existing Recommendation and choose an existing recommendation.
13. When you’re finished, click Build. Amplitude takes about an hour to generate your recommendation and sends you an email when it’s ready.

Understand your recommendation

After your recommendation is complete, click the recommendation to view basic information about it. The recommendation opens to the Overview tab.

The confidence score represents Amplitude’s confidence that this recommendation generates statistically significant lift relative to a random list of items.

Below the confidence score, Amplitude lists items ranked by their frequency of appearance in the recommendation. The item at the top of the list is the item this recommendation suggests most often and the item most likely to result in a conversion.

The Performance tab shows four statistics across the top:

• Accuracy: Amplitude Experiment conducts regular training runs using recent user data, then builds a model based on that data and recent user activity. Each training run excludes data from a percentage of users. When the model is complete, Amplitude Recommend runs it against this holdout group to estimate accuracy.
• Lift against Baseline: When Amplitude deploys a model, Amplitude Recommend shows recommendations to some users and random items to others. This statistic is the ratio of conversions in the group that saw recommendations to those that didn't.
• Recommendation CR and Control CR: These are the conversion rates in the population of users that saw a recommendation and the population of users that saw random items, respectively.
• Significance: The higher this number, the more confident Amplitude Recommend is of the result. Predictions and recommendations have their own built-in mini-version of Amplitude Experiment. The test group, control group, difference in conversion rates, number of exposures, and number of conversions all go into a standard significance calculation.

Common mistakes in creating a recommendation

• Using the wrong cohort. Recommending all users can be sufficient, but sometimes you need a more specific cohort. Think about your goal and which users are the best candidates to achieve it. If your goal is to optimize for a second purchase, for example, select users who have already purchased once as the starting cohort.
• Specifying the wrong outcome. Your outcome event dictates the rankings of the selected items. If you optimize for “product purchased WHERE amount > $100,” the recommendation prioritizes expensive items. Confirm that you want this behavior before launching the recommendation.
• Wrong exposure event timing. If you choose an exposure event that occurs after the outcome event, the recommendation doesn't train properly.
• Wrong exposure event context. The context for the event property depends on the exposure event it’s configured from. The event property’s name can easily mean different things on “product clicked” vs “button clicked.”
• Wrong event property. For example, Product Name and Product ID convey the same basic information, but in very different formats. Make sure the one you select matches the way data is stored in your CMS.
• Outcome event doesn't have enough unique actions: Ensure that the outcome event has at least 50, and ideally 100, unique user actions every day. If conversion numbers fall below this threshold, the model doesn't detect signals and recommends the same content for all users.