AI product recommendations on Shopify: how they work and what to expect

What are AI product recommendations on Shopify?

AI product recommendations are suggestions a system picks for a specific shopper, product page or cart, instead of a fixed list you set by hand. On Shopify they range from the platform's own “You may also like” block to apps that read behavior, catalog data or a live conversation to choose what to show next.

The label “AI” covers very different machinery. A widget that ranks products by co-purchase counts and an assistant that asks “Is this for oily or dry skin?” before suggesting a moisturizer are both sold as AI recommendations. They fail in different ways, so it pays to know which one you are looking at before you judge the results.

This guide walks through what Shopify gives you out of the box, the three main approaches a product recommendation engine can take, and the practical problems that decide whether shoppers act on a suggestion: new products with no history, sold-out items, and suggestions that never explain themselves.

How does Shopify's built-in product recommendation work?

Shopify offers two recommendation intents. Related products are similar items a shopper might choose instead, and Shopify generates them automatically. Complementary products are add-ons bought alongside the item, and you set them yourself in the free Search & Discovery app. Themes show them in blocks such as “You may also like” and “Pair it with.”

According to the Shopify Help Center page on product recommendations, auto-generated related products draw on purchase history first, then on the product description (only for stores with an English storefront), then on related collections when neither is available. You cannot edit the generated list, but you can add up to 10 related and up to 10 complementary picks per product and choose whether to show only your picks or your picks plus the generated ones. The same page notes that complementary products must be active and in stock to appear.

For developers, the Shopify Ajax Product Recommendations API exposes the same logic at a locale-aware recommendations endpoint. It takes a product ID, an optional limit from 1 to 10 (default 10) and an intent of related or complementary (default related). The shopify.dev theme guide to product recommendations confirms that only related recommendations are auto-generated; complementary ones need manual setup.

Rule-based vs collaborative vs conversational: what's the difference?

Rule-based recommendations follow conditions you write, such as “show the travel size with every full size.” Collaborative filtering learns from what many shoppers bought or viewed together. Conversational recommendations ask the shopper about the need first and then choose from the catalog. Each needs different inputs and breaks in different situations.

Most Shopify stores end up with a mix. Manual complementary picks are rules. The auto-generated related list leans on purchase history, which is a collaborative signal. A conversational layer, such as an AI shopping assistant in a chat window, sits on top and handles the shopper who does not know which product page to open in the first place.

Why do recommendations fail on new products and small stores?

This is the cold start problem. Systems that learn from co-purchases need history, and a new product or a young store has very little. Until enough orders accumulate, collaborative suggestions fall back to generic signals such as collections or overall popularity, which often look random to the shopper reading them.

Research on recommender systems has described cold start as a challenge that affects almost all of them, because new users and new items arrive with too little information to recommend from; see this peer-reviewed study on promoting cold-start items for a technical treatment. Shopify's own fallback order (purchase history, then description, then collections) is a practical answer to the same problem.

Content-based and conversational approaches suffer less, because they work from product attributes rather than order history. A product launched this morning already has a title, description, variants and tags. If those fields are complete, a system that reads them can recommend the product sensibly from day one.

• Fill in the fields a recommender can read: product type, tags, variant options (size, shade, material) and a description that states who the product is for.
• Set manual complementary picks for new launches so they appear before sales history exists.
• Keep size charts and ingredient or material lists on the product page in text, not only in images.
• Review suggestions on your newest products first; that is where generated lists are weakest.

What should happen when a recommended product is out of stock?

A recommendation should never send a shopper to something she cannot buy. Sold-out products and sold-out variants should be filtered out before the suggestion is shown, or clearly marked with an alternative offered in the same breath. The check has to run at the moment of the answer, not from a cached copy of the catalog.

Shopify filters inactive and out-of-stock items from the complementary block, per the Help Center. The Ajax API reference returns an available flag for each product and variant, so a theme or app can decide what to do. The tricky case is the variant: the product is in stock, but not in the shopper's size or shade.

AI tools add a second risk. A language model that describes a product from memory, or from a catalog snapshot taken last week, can state a price or stock level that is no longer true. That is a hallucination in the commercial sense, and it is why price and stock should be read live for every answer. Our article on AI chatbot hallucinations about prices and stock covers the checks in detail.

1. Read stock at the variant level, not only the product level.
2. Drop sold-out variants from suggestions, or show the closest in-stock alternative with a short reason.
3. Read price and compare-at price live when you show the card, never from the model's text.
4. Log requests for sold-out items; repeated asks are a restock signal.

Why should every recommendation explain why it fits?

A suggestion with a reason gives the shopper a way to judge it. “Matte finish, made for oily skin, which you mentioned” is something she can agree or disagree with. A bare thumbnail in a carousel asks her to do the matching herself, and she may simply scroll past it. The reason should come from real product data.

Baymard Institute's research on cross-sell modules found that 68% of desktop sites in its product page benchmark fail to provide enough information for cross-sells (published March 2021), such as a fully visible title, price or variations. If the basic facts are missing, shoppers have no basis for clicking. A reason tied to their stated need adds the piece a grid cannot.

Good reasons are short, specific and checkable. They point to an attribute the store has published (finish, fabric weight, width, metal) and connect it to something the shopper said. They do not invent benefits, certifications or ingredients that are not on the product page.

• Weak: “Customers also loved this!”
• Better: “Same fit as the jeans in your cart, in a lighter fabric for warm days.”
• Weak: “Perfect for you.”
• Better: “Comes in a wide width, which you asked about, and in your size it's in stock.”

How Tycoon approaches product recommendations

Tycoon is an AI sales associate for Shopify stores, currently in early access. When the shopper's need is unclear, it is built to ask one short question built from the store's own catalog, such as size, skin type or occasion, and then suggest one to three products, each with a reason.

If the shopper has already said what she needs, it skips the question. Each suggestion appears as a product card with real images, the live price and the right variant already selected. Price, compare-at price and stock are read live from the store for every answer, and every draft is checked whole before it is sent, so a figure that does not match live data or approved facts is blocked and rewritten.

On the Tycoon plan, nightly learning looks at what worked in that store, such as which recommendations led to a purchase and which questions helped shoppers decide, and tunes which product to recommend first. It learns only from that store. You can see the details on the AI product recommendations feature page and in how it works.

How do you know if product recommendations are working?

Compare revenue from shoppers who saw the recommendations with a similar group who did not. Clicks and “influenced revenue” reports can credit the tool for sales that would have happened anyway. A holdout test, where a small random share of visitors never sees the feature, isolates the difference the feature actually makes.

Shopify's own recommendation blocks report into the Analytics page of the admin, according to the shopify.dev theme guide. That tells you how often the blocks get clicked, which is useful, but it is not the same as incremental lift. For the method, read our guide to measuring AI chatbot revenue with a holdout test.

• Track add-to-cart rate from suggestions, not only clicks.
• Watch average order value and conversion rate together; one can rise while the other falls.
• Check return reasons for recommended items; “didn't fit” or “not as expected” points to weak reasons or poor data.

Frequently asked questions

Sources

1. Shopify Help Center: Product recommendations (Search & Discovery) · checked October 20, 2026
2. Shopify.dev: Ajax API, Product Recommendations reference · checked October 20, 2026
3. Shopify.dev: Show product recommendations (themes) · checked October 20, 2026
4. Baymard Institute: 6 list item attributes to include for cross-sell recommendations (2021) · checked October 20, 2026
5. Promoting Cold-Start Items in Recommender Systems (PLOS ONE, 2014, via PubMed Central) · checked October 20, 2026