How to Automate Personalized Recommendations in Braze

A hand moving a cursor to a green-outlined panel beside an untouched cream panel on a monitor

To automate personalized recommendations in Braze, build a catalog, train an AI item recommendation on a purchase or interaction event, and reference it with Liquid in your messages. Braze then retrains the model and refreshes user lists on a schedule. Your job is the setup choices, the fallback behaviour and a safe way to change a live recommendation.

Key Takeaways

  • Once a recommendation exists, Braze retrains the model and refreshes user lists on a schedule.
  • The event you configure decides which users receive AI recommendations.
  • Users without enough data get popular items, and the Personalization rate shows how many that is.
  • Each recommendation operates independently, so avoiding duplicates in one message takes Liquid.
  • A failed retrain disables the recommendation, so change a live one by building a replacement.
  • We recommend a wait step after a purchase, so the item just bought has dropped out of the list.

How do you automate personalized recommendations in Braze?

Braze does the recurring work once a recommendation exists. It retrains the model on a schedule, updates the lists, and fills gaps with popular items. Your messages read the result through Liquid, so nobody rebuilds a product list before each send.

The decisions that stay with you are which event feeds the model, which type you pick, and how you change a recommendation that campaigns already depend on. This post takes them in the order you'll meet them. The Braze item recommendations documentation is the source for every product behaviour below.

What has to be in place before the first model trains?

You need at least one catalog before you can create an item recommendation. The property you choose from your event data must map to the id field of the catalog you selected. The engine needs some events whose values match catalog ids to have enough content to work with.

The event you configure also decides who gets recommendations. Braze generates AI recommendations for users who have completed that event. A narrow event gives a narrow audience.

We recommend settling the audience first and the event second, because the audience is the business decision and the event follows from it.

Custom events must have sufficient data before they appear in the event list. A newly instrumented event can't be selected on day one, so tracking work has to land before the build starts.

AI Personalized recommendations work best with hundreds or thousands of items, up to 100,000 catalog items, and typically at least 30,000 users with purchase or interaction data. That is a rough guide. The other types can work with less.

Recommendations draw on interaction data held on user profiles. Which users may receive a message is decided by your consent and subscription rules, and this post doesn't settle that legal question.

Which recommendation type should run the automation?

Four types come from trained models, and rules-based recommendations come from logic you write. The difference that matters for automation is what refreshes by itself and what waits for you.

TypeWho the list is tailored toWhat refreshes itWhere we'd start
AI PersonalizedEach userUser updates, plus model retrainingLifecycle sends where purchase history exists
Recent-items typeEach userUser updates, plus model retrainingLapse and cart reminders
Popular-items typeEveryoneModel retrainingNew users and fallback sections
TrendingEveryoneModel retrainingSeasonal and launch moments
Rules-basedFixed logic you setOnly when you change the logicRestock and first-purchase sends

The popular-items and Trending types are global recommendations that update when the model retrains. Rules-based recommendations run on fixed logic that you set manually. They suit sends that rarely change.

The last column is our judgement, not Braze guidance. Treat it as a starting point and test it against your own audience.

How often do recommendations refresh without anyone touching them?

Recommendations retrain automatically once every week on the paid version and once every month on the other. User-specific lists update on their own schedule, and the paid version updates every 24 hours. A first build is slower: training can take anywhere from 10 minutes to 36 hours, with an email at the end.

Between updates, items a user has just bought may still appear. The option to exclude prior purchases and interactions is only available for AI Personalized, and messages sent after 24 hours won't include the item on the paid version. A message sent minutes after a purchase can therefore recommend the thing the customer just bought.

In a Canvas, we put a wait step at least as long as the update interval ahead of any recommendation-led message that follows a purchase. We also schedule a first build two days before launch, so the 36 hours of training time has room.

What happens when a user has too little data?

Users without enough data for a personalized list get popular items as a fallback. Braze also tops up short lists: if a model fills fewer than 30 items, it appends popular items after the personalized ones. Every send has something to show, because the fallback fills the gap.

The fallback does mean a message can look personalised to the sender and generic to the recipient. The Analytics page shows the Personalization rate, the percentage of users who performed the configured event in the past 24 months and have personalized recommendations stored on their profile. Read it before the first send, and decide whether the share on fallback deserves different copy.

How do you keep the same item out of two sections of one email?

Recommendations operate independently and don't know what the other models are recommending. Two sections in one message can therefore show the same item. Braze's documented answer is to use Liquid to track which item IDs you've already displayed and skip them in later sections.

The documented starting point for reading a recommendation is a single assignment, with your own recommendation name in the placeholder:

{% assign items = {{product_recommendation.${recommendation_name}}} %}

If a recommendation only returns one item, referencing the second position gives an empty field. Design the layout for a short list, and preview with a recipient who sits on the fallback.

What breaks an automated recommendation, and how do you change one safely?

Treat an active recommendation as a dependency of every message that references it. Read the documented behaviour for changing one closely.

SituationDocumented behaviourWhat we do
Property name changed on an active recommendationUsers keep seeing the previous version while the model retrainsDon't edit it. Build a replacement
Retraining failsThe recommendation is disabled, and Liquid that references it fails, so messages don't sendKeep the original untouched until the replacement has trained
Old item ID path after a renameEvents with the old path are effectively ignored for the new modelAgree identifier changes with developers before release

The right-hand column is our practice. Braze's own guidance covers the safe sequence below.

When identifiers drift between your product feed and your events, that is a data governance question, and CustomerIK's customer data management service covers schema design and validation rules for how data enters and moves through your stack.

The sequence Braze documents has three steps:

  1. Create a new item recommendation with the desired property name configuration.
  2. Verify that the training completes successfully.
  3. Update your messaging to reference the new recommendation instead of modifying the active one.

We add a fourth step: move one Canvas across first, send to an internal audience, then move the rest. Retire the old recommendation only when nothing references it.

Take a hypothetical retailer whose developers rename the product identifier on the purchase event during a site release. Marketing edits the live recommendation to match.

If the retrain fails, every campaign that references it stops sending together. Had the team built a second recommendation instead, the original would have kept serving while the new one trained, and a failed build would have cost nothing but time.

How does this connect to the rest of the programme?

Recommendations are one input to a wider personalisation set-up. Our comparison of AI and rules-based item recommendations helps you choose between types, and Connected Content covers engines that live outside Braze. The mechanics of Liquid sit in our post on personalisation with Liquid tags.

Two pieces cover the failure side. Common personalisation pitfalls lists what goes wrong in live programmes, and what resolves when in real-time personalisation explains timing. Once a recommendation block is live, email A/B testing is how you learn whether it earns its place.

As a Braze implementation partner, CustomerIK recommends giving every live recommendation a named owner, changing it only by building a replacement, and reading the Personalization rate before each new audience goes in.

Frequently Asked Questions

1. How long does a new recommendation take to train?

Training can take anywhere between 10 minutes and 36 hours. Braze emails you when it finishes, or with an explanation if it failed.

2. How often does a recommendation retrain?

Once every week on the paid version and once every month on the other. The global types update when the model retrains.

3. Can a catalog selection that contains Liquid be used in an AI recommendation?

Selections with Liquid cannot be used in AI recommendations, so build the selection without Liquid first.

4. Will a customer be shown something they just bought?

It can happen between updates. The exclusion option is only available for AI Personalized, and it depends on the recommendation having updated recently.

5. What do users with no history receive?

They receive popular items from the linked catalog as a fallback.

Sources

CustomerIK is a Braze implementation partner working across onboarding, technical integration, marketing operations and customer data management.

If you're planning a recommendation build or changing a live one, let's talk.

Preferences

Privacy is important to us, so you have the option of disabling certain types of storage that may not be necessary for the basic functioning of the website. Blocking categories may impact your experience on the website.

Accept all cookies
Accept all cookies

These items are required to enable basic website functionality.

Always active

These items are used to deliver advertising that is more relevant to you and your interests.

These items allow the website to remember choices you make (such as your user name, language, or the region you are in) and provide enhanced, more personal features.

These items help the website operator understand how its website performs, how visitors interact with the site, and whether there may be technical issues.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.