Discovery

AI Product Recommendations: How They Work and What They're Worth

AI product recommendations promise bigger baskets. See what the models actually use, how to measure the lift, and where they fail on Shopify.

A cream ceramic peg figure linked by a saffron cord to one burnt-orange parcel stamped with a tripod, with two plain kraft parcels behind, on a butter-yellow background.
A recommendation is one clear link between a shopper and the item that completes the decision.

Every storefront shows recommendations. Few merchants can say what the block earns. That gap matters more now that the label “AI” sits on almost every recommendation app in the Shopify App Store, because AI product recommendations are not one technique. They are a family of methods that use different inputs and fail in different ways.

This article separates the methods from the marketing: what the models actually run on, what a realistic lift is worth in your profit and loss, the four ways these systems go wrong on a Shopify catalog, and the questions that tell you in one demo call whether a tool will work on your products.

What AI product recommendations actually run on

Strip the branding from any recommendation engine and you find the same question: given this shopper, in this session, on this page, which of these products should come next? The systems differ in which evidence they trust.

  • What this shopper did. Page views, searches, a cart, past orders. Session-based systems weight the last few clicks most heavily, because intent decays fast: someone comparing three air fryers this minute is closer to buying than someone who viewed one last week.
  • What similar shoppers did. Purchase and view co-occurrence: people who bought A also bought B. This is the workhorse of storefront recommendation apps and needs volume to work. Below a few hundred orders a month, the pattern it finds is mostly noise.
  • What the product is. Attributes pulled from titles, descriptions, tags, collections and metafields: fit, compatibility, size, material. Attribute matching is what lets a system suggest a case for a camera rather than another camera, and it is only as good as your product data.

Modern tools mix all three signals and rank the result, often with a language model reading the product data and the conversation. That ranking step is where quality is won or lost. A model that can see attributes can respect a hard constraint such as “fits a 12-inch shelf”. A model that only sees co-purchase counts cannot, however well it ranks.

The five types you will actually deploy

Whatever the model, almost every recommendation on a storefront is one of five types. The type decides where the block belongs and what it can honestly promise.

  • Personalized, for you. Built from this shopper’s own history and session. Belongs on the homepage and category pages, where there is no single product to anchor to. It needs returning traffic to work; on a first visit it falls back to bestsellers.
  • Complementary, cross-sell. Items that complete the purchase: a case for a camera, a filter for a lens. Belongs on the product page and in the cart, where the intent to finish a setup already exists.
  • Alternative, substitute. The same job at a different price, size or colour. Belongs on the product page beside the item being viewed, and not in a full cart, where a cheaper substitute reads as a downgrade.
  • Bestseller or trending. Social proof rather than personalization. Useful on the homepage and collection pages, because it answers what other shoppers bought, which is not the same question as what this shopper should buy.
  • Bundle or kit. Several items sold together at a saving, like the tripod offered with a camera. Belongs at the cart, and works best when the bundle is a real SKU with a real price rather than a discount the engine invents at runtime.

Many apps blur these into one row labelled recommended for you. That is where most bad recommendations come from: a substitute where a complement was needed, or a bestseller dressed up as personalization. Match the type to the page before you judge the algorithm.

What they are worth, and how to check

Recommendation vendors quote lift percentages. Those numbers come from someone else’s catalog, price points and traffic, usually from a test period with a control group. Yours will differ. Four numbers, all from Shopify analytics or a simple export, tell you what a block actually earns:

  1. Attach rate. Orders with two or more line items as a share of all orders. It moves before revenue does, so it is the earliest honest signal that suggestions are being clicked.
  2. Average order value. Revenue divided by orders. BuyScout reports that stores using its AI Store Agent typically grow average order value 15–25% through upsells and cross-sells. That is a vendor figure describing typical results, not a guarantee, and it belongs against your own 28-day baseline.
  3. Revenue per session. This separates a bigger basket from fewer, larger orders. A recommendation block that lifts order value while quietly lowering conversion has not helped you.
  4. Return rate on recommended items. An accessory that does not fit, or a size suggested badly, costs more in returns and support time than the extra line item earns.

Then subtract what you gave away. A recommendation that moves a shopper from a 60 US-Dollar item to a 45 US-Dollar alternative raises nothing. A discount attached to every suggestion can lift conversion and cut margin at the same time. The honest test is contribution margin per session, before and after, over the same length of time. Twenty-eight days is a workable window, and Shopify has no built-in storefront A/B test, so most merchants compare periods and watch for seasonality.

One more caution: recommendations are not incremental if they only show the shopper what they were going to buy anyway. A block that repeats the product page’s own related items adds nothing. Value comes from surfacing something the shopper had not considered, at a moment when the suggestion saves a decision.

Where they go wrong on a Shopify store

Most failures are data failures wearing an algorithm costume. Four recur:

  • Cold start. A new store, a new collection, or a product with no history has nothing to learn from. Co-purchase models show bestsellers instead, which is safe and dull. Attribute matching and a conversation both work on day one; history-based ranking does not.
  • Out-of-stock and stale catalog. If the engine does not read inventory at the variant level, it will recommend a color that sold out on Tuesday. Check the block on a product whose popular variant is out of stock.
  • Wrong relationship. Alternatives belong on the product page, complements in the cart. An engine tuned on one will misbehave in the other: a cheaper substitute under a full cart reads as a downgrade, and a case for every item on a collection page is noise.
  • Confident nonsense. A language-model layer can produce a recommendation that sounds specific and is wrong: a filter that does not fit, a bundle discount that never existed. The fix is grounding, not better prose: every claim should trace back to your product data or policy.

Two existing articles go deeper on the hard parts: Conversational AI product recommendations covers the design of the conversation itself, and the Shopify catalog checklist covers the product-data audit that decides whether any engine can be accurate.

You can judge a tool faster than its sales deck expects. Four questions, asked against your own catalog:

  1. Can it see inventory and variants? Ask it to recommend from a product where one variant is sold out. If it suggests the sold-out one, its data layer is shallow.
  2. Can it state why? A suggestion should come with a reason a shopper can check: “works with your model”, “same size as your last order”. Recommendations without reasons get ignored.
  3. Can it apply a constraint? Name a real filter from your catalog, whether width, compatibility, skin type or pet size, and see whether the shortlist respects it or merely mentions it.
  4. What does it cost per order, and where does the data go? Compare the monthly fee against contribution margin, not revenue, and confirm the tool is clear on data handling. BuyScout states GDPR and AI Act compliance as a baseline.

The fourth question weighs more in the US than it did two years ago, but the first three decide whether the tool works at all.

Where a conversation changes the job

A recommendation row infers intent from behavior. A conversation asks for it. BuyScout® AI Store Agent is a free app for Shopify stores that puts an AI store agent on the storefront, working from your product data and policies plus the page, browsing, cart and conversation context. That changes what a recommendation can be:

  • It can answer a constraint directly. “Which of these fits a 12-inch shelf?” is a filter a thumbnail row cannot run.
  • It can compare on the attribute that matters. Two projectors side by side on brightness is a different job from showing three more projectors.
  • It can suggest at the moment of hesitation. When a camera goes into the cart, the agent can offer the tripod bundle at 10% off, the example BuyScout uses on its own site.
  • It can explain fit. Size and fit guidance answers the doubt that otherwise turns a recommendation into a return.

BuyScout reports that stores using the agent typically lift conversion 10–30%, grow average order value 15–25% and cut repetitive support tickets 41%. Those are the vendor’s own figures, described as typical rather than guaranteed; your catalog, prices and traffic decide what you actually see.

Start smaller than you think. Open ten product pages and read the six items the block shows. Fix the obvious misses first: out of stock, wrong collection, a cheaper substitute for the same product. Then record 28 days of attach rate, average order value and revenue per session, install the tool, and compare the same window.

BuyScout® AI Store Agent is free forever and installs from the Shopify App Store in about a minute. Put it on one product page and ask it a question a real customer would ask, whether a constraint, a comparison or a fit question. That is a faster test than another app trial, and it costs nothing to run.

Other articles

All articles

Try it Now

Help shoppers find the right product faster

Free forever. Set up in 1 minute.