AI item recommendations

Learn how to create an AI item recommendation for items in a catalog.

Use AI item recommendations to calculate the most popular products or create personalized AI recommendations for a specific catalog. After creating your recommendation, you can use personalization to insert those products into your messages.

Prerequisites

Before you start, you’ll need to complete the following:

• You must have at least one catalog to use any of the recommendation types described below.
• You must have purchase or event data on Braze (custom events or the purchase object) that includes a reference to unique product IDs stored in a catalog.

Creating an AI item recommendation

To create an item recommendation:

1. Go to Analytics > AI Item Recommendation.
2. Select Create Prediction > AI Item Recommendation.

You can also choose to create a recommendation straight from an individual catalog. Select your catalog from the Catalogs page, then select Create Recommendation.

Step 1: Add recommendation details

Give your recommendation a name and optional description.

Step 2: Define your recommendation

Select the recommendation type. All recommendation types use the last six months of item interaction (purchase or custom event) data. The interaction mentioned below refers to either a purchase event or a custom event chosen in Step 3.

• Most popular: Calculates up to 30 items from the catalog that all users in the workspace interact with most often, such as the most purchased products.
• Most recent: Creates a list of up to 30 products a user has interacted with most recently.
• AI Personalized: Uses transformers, a new kind of deep learning, to predict each user’s next most likely set of items to interact with. We calculate up to 30 of the next most likely items ranked from most to least likely. This type of recommendation does not use large language models (LLMs) to combine your data with that of any other Braze customer.
• Trending: Calculates up to 30 items from the workspace that had the most recent positive momentum when it comes to user interactions.

Step 2a: Exclude prior purchases or interactions (optional)

To avoid suggesting items that a user has already purchased or interacted with, select Do not recommend items users have previously interacted with. This option is only available when the recommendation Type is set to AI Personalized.

This setting prevents messages from reusing the items a user has already bought or interacted with, provided the recommendation has been updated recently. Items purchased or interacted with between recommendation updates may still appear. For the free version of item recommendations, updates happen weekly. For the pro version of AI item recommendations, updates happen every 24 hours.

For example, when using the pro version of AI item recommendations, if a user purchases something and then receives a marketing email within 30 minutes, the item they just purchased might not be excluded from the email in time. However, any messages sent after 24 hours won’t include that item.

Step 2b: Select a catalog

If not already populated, select the catalog that this recommendation will pull items from.

Step 2c: Add a selection (optional)

If you’d like more control over your recommendation, choose a selection to apply custom filters. Selections filter recommendations by specific columns in your catalog, such as brand, size, or location. Selections that contain Liquid can’t be used in your recommendation.

Step 3: Select the interaction to drive recommendations

Select the event you want this recommendation to optimize for. This event is usually a purchase, but it can also be any interaction with an item.

You can optimize for:

• Purchase events with the Purchase Object
• Custom events that represent a purchase
• Custom events that represent any other item interaction (such as product views, clicks, or media plays)

If you choose Custom Event, select your event from the list.

Step 4: Choose the corresponding property name

To create a recommendation, you need to tell Braze which field of your interaction event (purchase object or custom event) has the unique identifier that matches an item’s id field in the catalog. Not sure? View requirements.

Select this field for the Property Name.

The Property Name field will pre-populate with a list of fields sent through the SDK to Braze. If enough data is provided, these properties will also be ranked in order of probability to be the correct property. Select the one that corresponds to the id field of the catalog.

Requirements

There are some requirements for selecting your property:

• Must map to the id field of your selected catalog.
• If you selected Purchase Object: Must be the product_id or a field of your interaction event’s properties.
• If you selected Custom Event: Must be a field of your custom event’s properties.
• Nested fields must be typed into the Property Name dropdown in dot notation with the format of event_property.nested_property. For example, if selecting the nested property district_name within the event property location, you would enter location.district_name.
• The field can be inside an array of products, or end with an array of IDs. In either case, each product ID will be treated as a separate, sequential event with the same timestamp.

Example mappings

The following example mappings both refer to this sample catalog:

• custom event
• purchase object

1
2
3
4
5
6
7
8
9
10
11
12
13
{
  "events": [
    {
      "external_id": "user1",
      "app_id": "your-app-id",
      "name": "added_to_cart",
      "time": "2024-07-16T19:20:30+01:00",
      "properties": {
        "product_sku": "ADI-BL-7"
      }
    }
  ]
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
{
  "events": [
    {
      "external_id": "user1",
      "app_id": "your-app-id",
      "name": "added_to_cart",
      "time": "2024-07-16T19:20:30+01:00",
      "properties": {
        "transaction_id": "2ff3f9a9-8803-4c3a-91da-14adbf93dc99",
        "products": [
          { "sku": "ADI-BL-7" },
          { "sku": "ADI-WH-9" }
        ]
      }
    }
  ]
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
{
  "events": [
    {
      "external_id": "user1",
      "app_id": "your-app-id",
      "name": "added_to_cart",
      "time": "2024-07-16T19:20:30+01:00",
      "properties": {
        "transaction_id": "13791e08-7c22-4f6c-8cc6-832c76af3743",
        "purchase": {
          "product_skus": ["ADI-BL-7", "ADI-WH-9"]
        }
      }
    }
  ]
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
{
  "purchases": [
    {
      "external_id": "user1",
      "app_id": "11ae5b4b-2445-4440-a04f-bf537764c9ad",
      "product_id": "ADI-BL-7",
      "currency": "USD",
      "price": 100.00,
      "time": "2024-07-16T19:20:30+01:00",
      "properties": {
        "color": "black",
        "checkout_duration": 180,
        "size": "7",
        "brand": "Adidas"
      }
    }
  ]
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
{
  "purchases": [
    {
      "external_id": "user1",
      "app_id": "11ae5b4b-2445-4440-a04f-bf537764c9ad",
      "product_id": "shoes",
      "currency": "USD",
      "price": 100.00,
      "time": "2024-07-16T19:20:30+01:00",
      "properties": {
        "sku": "ADI-RD-8",
        "color": "red",
        "checkout_duration": 180,
        "size": "8",
        "brand": "Adidas"
      }
    }
  ]
}

Step 5: Train the recommendation

When you’re ready, select Create Recommendation. This process can take anywhere from 10 minutes to 36 hours to complete. You will receive an email update when the recommendation is successfully trained or an explanation of why the creation may have failed.

You can find the recommendation on the Predictions page, where you can then edit or archive it as needed. Recommendations will automatically retrain once every week (paid) or month (free).

Analytics

You can view analytics for your recommendation to see which items users were recommended and how accurate the recommendation model was.

1. Go to Analytics > Item Recommendation.
2. Select your recommendation from the list.

At the top of the page, you can find statistics about your recommendation, such as precision and coverage.

These metrics are defined in the following table.

The next section shows a breakdown of items in the catalog, split into two possible columns:

• Personalized items or Most recent items: This column lists each item in the catalog in descending order of most often recommended to users. This column also shows how many users were assigned each item by the model.
• Most Popular items: This column lists each item in the catalog in descending order of popularity. Popularity here refers to items in the catalog that users interact with most often in the entire workspace. Most popular is used as the fallback when personalized or most recent cannot be calculated for an individual user.

The Recommendation overview shows a summary of your chosen recommendation configuration, including when the recommendation was last updated.

Using recommendations in messaging

After your recommendation finishes training, you can personalize your messages with Liquid to insert the most popular products in that catalog. The Liquid can be generated for you by the personalization window found in message composers:

1. In any message composers that support personalization, select to open the personalization window.
2. For Personalization Type, select Item Recommendation.
3. For Item Recommendation Name, select the recommendation you just created.
4. For Number of Predicted Items, enter how many top products you’d like to be inserted. For example, you can display the top three most purchased items.
5. For Information to Display, select which fields from the catalog should be included for each item. The values for these fields for each item will be drawn from the catalog associated with this recommendation.
6. Select the Copy icon and paste the Liquid wherever it needs to go in your message.

AI item recommendations tiers

The following table describes the differences between the free and pro version of the AI Personalized, Popular, and Trending recommendation types:

<sup>1. This is the frequency at which user-specific item recommendations are updated (all models except Most Popular items, which updates when the model retrains). For example, if a user purchases an item recommended based on AI item recommendations, their recommended items will be updated according to this frequency
2. Available recommendation types are AI Personalized, Most recent, Most popular, and Trending.</sup>

Frequently asked questions

What causes “Most popular” items to be mixed into other models’ recommendations?

When our recommendation engine curates a list for you, it first prioritizes personalized selections based on the specific model you’ve chosen, like “Most recent” or “AI Personalized”. If this model can’t fill the complete list of 30 recommendations for whatever reason, some of your most popular items among all users are then added to make sure each user always has a full set of recommendations.

This happens under a few specific conditions:

• The model finds fewer than 30 items that match your criteria.
• Relevant items are no longer available or in stock.
• Items don’t meet the current selection criteria, perhaps due to a change in stock or user preferences.