AI product recommendations are systems that predict which products a shopper is most likely to want next, using browsing and purchase behavior, product attributes and context such as device or season. For marketers, they are no longer only a “you may also like” strip: they shape home pages, emails, search results, ads and, increasingly, AI shopping assistants that answer questions in natural language.
How do recommendation engines work?
Most systems combine several methods. Collaborative filtering finds patterns across shoppers: people who bought A also bought B. Content-based filtering compares product attributes, such as color, material or brand. Newer systems add deep learning models that read sequences of actions in a session and, in some cases, language models that understand product descriptions and reviews.
The output is a ranked list of products for a given slot, such as a product page, cart or email. The best systems retrain often, so a sudden trend or stock change shows up quickly.
Which tools do stores actually use?
The choice depends on the platform and the size of the catalog.
- Shopify: built-in Search & Discovery app for related products, plus apps such as Rebuy, LimeSpot and Klaviyo for email recommendations
- WooCommerce: plugins and external services, including Klaviyo, Clerk.io and Algolia Recommend
- Enterprise: Google Cloud’s Vertex AI Search for commerce, Amazon Personalize, Salesforce Commerce Cloud Einstein, Adobe Commerce with Adobe Sensei features, Bloomreach, Dynamic Yield
- Email and CRM: Klaviyo, Braze, Insider and Salesforce Marketing Cloud offer product blocks fed by behavior
Many of these are configured by marketers, not data scientists. That shifts the work from building models to choosing strategies, slots and rules.
Where do marketers use recommendations?
Recommendations show up across the customer journey. Each slot has a different job.
On site
- Home page: popular or trending items for new visitors, personalized picks for returning ones
- Product page: similar items and frequently bought together
- Cart: low-cost add-ons and accessories
- Search: re-ranking results by likely interest
- Empty or out-of-stock pages: close alternatives instead of a dead end
Off site
- Abandoned browse and cart emails with related products
- Post-purchase emails with complementary items, timed to use
- Dynamic product ads in Meta and Google, fed by the catalog and pixel data
Conversational
- AI shopping assistants on site that answer “what’s a good running shoe for flat feet?” and suggest products
- Product data that feeds answers in tools like ChatGPT shopping and Google’s AI Mode, where clean product feeds and reviews matter
What do marketers control?
Even with automated models, the marketer sets the frame. Poor settings produce poor recommendations no matter how good the model is.
- Strategy per slot: similar, complementary, trending, recently viewed, personalized
- Business rules: exclude out-of-stock items, set margin floors, boost new collections
- Diversity: avoid showing five near-identical items
- Cold start: what new visitors and new products see before there is data
- Placement and design: where the block sits and how many items it shows
- Testing: A/B tests between strategies or against no recommendations
What are the limits and risks?
Recommendations amplify what already sells. Without rules, a few best-sellers fill every slot and new products never get exposure. That is good for short-term clicks and bad for range and margin.
Privacy rules also apply. Personalization based on tracking depends on consent under GDPR, KVKK and similar laws. Shoppers who decline cookies still need sensible, non-personal recommendations.
- Filter bubbles: shoppers see more of the same and less of the range
- Data quality: wrong categories or missing attributes produce odd suggestions
- Over-personalization: recommendations that feel invasive, especially on sensitive products
- Attribution: vendors often credit the widget for sales that would have happened anyway
How do you know if they work?
The honest test is an experiment: show recommendations to one group and not to another, or compare two strategies. Clicks on the widget are not proof of value.
Track revenue per visitor, average order value and conversion rate between groups. Watch margin too, since cheap add-ons may lift order count but not profit.
What product data do recommendations need?
Engines learn from two sources: behavior and catalog data. Behavior comes from tracking events such as product views, add to cart and purchases, each with a product ID. Catalog data comes from your product feed: titles, categories, attributes, prices, stock and images.
Weak catalog data is the most common reason recommendations look random. A fashion store that stores color only inside the product title, not as a field, cannot ask the engine to suggest “the same dress in another color”. A home goods store with three different spellings of one category splits its own data.
- Consistent categories with one clear hierarchy
- Attributes as separate fields: color, size, material, style, price band
- Accurate stock levels synced at least daily
- Clean product IDs that match across site, feed and analytics
- Good images, since some engines also use visual similarity
How is generative AI changing recommendations?
Language models add a new layer. Instead of a grid of products, a shopper can describe a need and get a short list with reasons. Shopify, Amazon with Rufus and many third-party tools now offer this kind of assistant.
For marketers, this raises the value of rich product descriptions, clear specifications and real reviews, because the assistant reads them to decide what to suggest. A product with a vague two-line description is hard for any assistant to recommend with confidence.
How Alien Road applies AI product recommendations
In a typical e-commerce engagement, we start with a catalog and data audit: product attributes, categories, stock feeds and consent setup. Gaps there limit any recommendation engine, so fixing them comes first.
We then map slots across the site and email flows, choose a strategy for each and set business rules for stock, margin and diversity. Tests start on the highest-traffic slots, usually product pages and cart.
Measurement uses holdout groups where the platform allows, and reports compare revenue per visitor, order value and margin between groups. We report what the tests show, including slots where recommendations made no difference.
Where to start this week
List every place your store shows recommendations today and the strategy each uses. Many stores find the same “related products” setting everywhere. Change one slot, such as the cart, to complementary low-cost items with an in-stock rule, and run it for two weeks against the current version. Compare average order value between the two groups before changing the next slot.