Skip to main content
Product recommendations are a tool that automatically matches relevant products to different situations. You can use them to build personalized collections for returning customers, surface bestsellers for new visitors, or display complementary products next to a specific item on your website.

​
Where you can use recommendations

You can deploy recommendations in three primary contexts:
  • Email campaigns and newsletters — embed personalized product blocks in any email.
  • Recommendation widgets on your website — show dynamic product collections directly on storefront pages.
  • API operations — pull recommendation output through the API for further processing in any channel.

​
How to create a recommendation algorithm

Setting up a new algorithm takes four steps:
1

Open the recommendations section

Go to Content → Product Recommendations → Create recommendations preset.
Screenshot 2026 07 27 At 15 11 17
2

Select an algorithm

Choose the algorithm type that matches your objective (bestsellers, similar products, related products, and so on).
Screenshot 2026 07 27 At 15 14 10
3

Configure the settings

Fill in the settings specific to the algorithm type you selected. The screenshot below uses Custom recommendations settings as an example.
If the algorithm builds recommendations from a scheduled product segment, refresh that segment before launching the algorithm. That way, the algorithm will use the current segment contents when generating recommendations.
Screenshot 2026 07 27 At 15 18 09
4

Launch and wait for recalculation

Start the algorithm and wait for its first recalculation to finish.
Screenshot 2026 07 27 At 17 11 03
If the algorithm builds recommendations from a scheduled product segment, refresh that segment before launching the algorithm. That way, the algorithm will use the current segment contents when generating recommendations.
Algorithms are not available for selection in campaigns or widgets until their initial recalculation finishes. Algorithms left in draft status also remain unavailable.

​
How recommendations work in multi-brand projects

In multi-brand projects, recommendation algorithms are available across all brands. However, when recommendations are generated for a campaign, only products from the brand where the recommendation preset was created are taken into account. Cross-brand recommendations are not allowed. The brand used for recommendation output depends on the channel:

​
How regional data is handled

If your catalog includes regional data, recommendations are generated in two steps:
  1. During recalculation — regional availability and pricing are considered. All other regional attributes fall back to the master feed values.
  2. At output time — regional product data matching the customer’s zone is inserted into the final recommendation.
This means a customer in one region can see prices and stock relevant to their location, while the underlying algorithm still operates on a consistent catalog.
ExampleThe Products similar to recently browsed products algorithm is running. In the similarity settings, the custom product field Company is selected — and its value can differ from zone to zone. Similarity is calculated from the master feed value, while the email displays the value from the customer’s zone.
  1. The customer views Product 1. In the master feed, its company is A; in the customer’s zone, it’s B.
  2. The algorithm looks for similar products using the master feed value, so it selects products where the company is A.
  3. Product 2 makes it into the selection: its master feed company is also A, but in the customer’s zone it’s C.
  4. In the campaign, Product 2 appears with the value from the customer’s zone — company C.
Result: the customer sees a recommendation labeled company C, even though the matching was done on company A. This is expected: similarity is calculated from the master feed, and display values come from the zone.

​
Recommendations and product groups

Use product groups to combine variants of the same product. If you sell one shirt in five colors, group those five items together — otherwise recommendations can show the same shirt five times in a row. During calculation, grouped products are treated as a single unit and their interactions are summed. At output, the algorithm displays the most popular product from that group, so customers never see two near-identical items side by side.
ExampleA group contains two products:
  • Sneakers in size 8 — 5 views, 2 orders.
  • Sneakers in size 9 — 1 view, 1 order.
When calculating recommendations, the algorithm treats these as a single product: sneakers with 6 views and 3 orders. The recommendation then shows the size 8 pair, since it’s the more popular of the two.

​
Real-time algorithms

Some algorithms update dynamically within seconds based on customer behavior on your site.
Avoid real-time algorithms in triggered flows that send immediately after an on-site action. The data they rely on may not be ready yet, so recommendations can come through empty or off-target.

​
Types of real-time recommendations:

1. Based on a product list
  • Often Purchased With Items in Product List
  • Similar products to product list recommendations
Recalculated when products are added to or removed from the list. Changing an item’s quantity does not trigger a recalculation. 2. Based on products and categories viewed in the last session
  • Bestsellers from previously viewed categories
  • Products similar to recently browsed products
Recalculated whenever the customer views a product or category. 3. Based on recent orders
  • Custom post-purchase recommendations
  • Post-purchase cross-sell
Recalculated whenever an order changes. The customer’s last order is the most recently modified order — not necessarily the most recently placed one.