AI Product Recommendations: Types and How They Work

Product recommendations suggest items a customer could buy, using store rules, product similarities or information about that customer. AI can rank those suggestions from browsing and purchase patterns. A shopping assistant can also ask what the customer needs and recommend from their answers.

A bestseller carousel gives everyone a starting point. Personalised recommendations change with the customer, so the same popular products aren’t always the right response.

Types of Product Recommendations

Product recommendations differ by the information used to select them. These hypothetical store examples show the difference.

Type	What selects the products	Store example
Bestsellers	Sales across the store or a category	A homewares store shows its most purchased mugs to new visitors.
Related items	Similar attributes, shared categories or merchant choices	A customer viewing a cotton quilt sees other cotton quilts in the same size.
Frequently bought together	Products repeatedly purchased in the same order, or a retailer’s chosen pairing	A camera page suggests a memory card customers often buy with that camera.
Personalised	The individual customer’s browsing, previous orders or stated needs	A customer who browses waterproof jackets sees suitable rainwear when they return.

A recommendation is personalised when the selection changes with the individual customer. Putting someone’s name above a bestseller list doesn’t personalise the products. Related items can change with the product page while still showing every visitor the same suggestions.

Recommendations are one part of ecommerce personalisation. Frequently-bought-together blocks can support cross-selling, but a purchase pattern alone doesn’t prove two products are compatible.

How AI Product Recommendations Work

AI product recommendations rank catalogue items using signals about products and customer behaviour. Browsing shows interest, while completed orders show what customers actually bought. Product attributes help the model find similarities between items.

One approach matches products with similar descriptions or attributes. Another finds patterns across customers who viewed or bought similar items. This second approach is called collaborative filtering. A system can combine these methods. Stock and category restrictions then limit which results it displays.

For a hypothetical outdoor store, repeated views of waterproof jackets signal interest in rainwear. Previous orders can suggest which items customers tend to buy next. Catalogue attributes distinguish waterproof jackets from lightweight windbreakers. A useful recommendation still needs the customer’s size and an available product.

Behaviour-based models need repeated activity across products and customers. There’s no single minimum traffic or order count that applies to every model. A bought-together model needs orders containing product pairs. A personalised model needs enough information about the individual customer as well as patterns across the store.

Adobe says most behavioural recommendation types need several days of active shopper activity. Its readiness indicators also depend on catalogue size and how widely customers interact with products. Catalogue-based similarity recommendations can work without purchase history. Low-traffic stores can begin with selected related products and switch on behavioural recommendations when their data supports them.

Missing attributes create another problem. If every jacket has a vague description, the model has less information for distinguishing materials or weather protection. Improve the product data before expecting a different algorithm to fix poor suggestions.

Where Recommendations Belong on Your Store

Choose recommendation placement by the decision the customer is making at that point. Each block needs a clear job.

Location	Useful recommendation type	What it helps the customer do
Home page	Bestsellers for new visitors, personalised items for returning customers	Find a starting point in the range
Product page	Related alternatives or compatible frequently-bought-together items	Choose between products or find a missing accessory
Cart	A small selection relevant to the basket	Add a useful item without losing the checkout path
Email	Recently viewed items, personalised suggestions or bestsellers	Return to a product or discover something relevant

Avoid stacking several carousels on one page. Repeated products make the blocks harder to distinguish and take space from product details. Start with one recommendation block that answers the customer’s next question. Keep cart suggestions away from the controls needed to complete the order.

Use clear labels that describe the selection. Call a shared bestseller list ‘Bestsellers’, even when the block sits in a personalised page layout.

Options on Shopify, WooCommerce, Magento and Klaviyo

Store platforms have different starting points, so separate built-in product relationships from additional recommendation services.

Shopify’s Search & Discovery app lets you select related and complementary products for product pages. Shopify also generates related suggestions using purchase history, descriptions or related collections. Manual selections can appear ahead of automatic suggestions or replace them. The theme needs the relevant recommendation sections. Additional Shopify apps can supply placements or selection methods beyond that starting point.

WooCommerce generates related products from shared categories or tags. You can manually assign up-sells and cart cross-sells, with display dependent on the theme. Woo’s Product Recommendations extension adds recommendation engines with filters and placement controls. Its product page warns that engines might not work as expected with block themes. That constraint matters before you choose it for a store using the Site Editor.

Magento stores can begin with product-level relationships for related items, up-sells and cross-sells. Adobe Commerce also has Adobe AI-powered Product Recommendations. That service isn’t a standard Magento Open Source feature. For Magento Open Source, an extension or a separately connected recommendation service is the route to additional automated selection. Keep the distinction between those editions when comparing options.

Klaviyo product recommendations use product feeds inside email product blocks. A feed can select bestsellers or personalise products using customer behaviour. Klaviyo’s ‘Products a customer may also like’ uses collaborative filtering to predict a next purchase.

In a flow email, add a dynamic product block and select its feed. Browse-based suggestions need Viewed Product tracking. Choose a fallback for recipients with limited history, then preview the email for different recipients. A welcome flow can use bestsellers while a browse-abandonment flow can use recently viewed products.

Conversational Product Recommendations

Conversational product recommendations use answers about the customer’s needs to narrow the catalogue. An assistant can ask about the intended use or budget before suggesting products. This gives a new visitor a way to express preferences without building a long browsing history first.

In a hypothetical furniture conversation, a customer asks for a small dining table. The assistant asks how many people it needs to seat and the maximum width. Those answers rule out unsuitable tables. A later request for a lighter finish changes the suggestions while keeping the size requirements.

shoppa is an AI shopping agent that helps customers find, compare and buy products in the chat on your site. It searches your catalogue by meaning and keyword, with exact matches first. Customers can refine the results over several messages, and shoppa carries the earlier requirements forward.

Its suggestions use your live catalogue data and store documents. Customers can compare products side by side on the attributes that matter, with the trade-offs stated. They can add an item to their basket and pay in the same chat. Order support and returns continue there after the purchase.

Choose conversational recommendations when customers need help explaining what will suit them. Keep passive recommendation blocks for customers who want to browse without starting a conversation. Both approaches still depend on accurate product information.

Judging Whether Recommendations Earn Their Place

Judge product recommendations by what happens after customers see them, including whether more of those sessions end in a purchase.

Track click-through rate as recommendation clicks divided by recommendation impressions. Use a consistent definition of an impression, such as a block that actually enters the viewport. Track revenue from recommended items with a declared attribution window. Keep that attributed revenue separate from evidence of additional sales.

Average order value (AOV) is order revenue divided by completed orders. Read it beside conversion rate and revenue per session. AOV can rise while fewer customers buy, so a larger average basket alone doesn’t show the recommendation helped.

Run an A/B test with one group seeing the existing block and another seeing the proposed replacement. Keep the placement and other page content unchanged when testing the selection method. Assign customers consistently and choose the success measure before the test starts. Use enough completed orders to judge the difference, rather than stopping at the first favourable result.

For a small store, begin with a frequently visited product category so the test receives usable activity. Inspect whether the suggested items are in stock and compatible with the main product. Keep the version that improves the chosen sales measure without making checkout harder. Use the separate guide to increasing average order value when you need to compare recommendations with other order-value tactics.