What Is an Ecommerce Recommendation Engine? Placements, Data and Setup

A recommendation engine decides which products a shopper sees next: on the product page, in the cart, at checkout, after payment and in emails. This guide covers where those recommendations go, the data that drives them, rules versus machine learning, what Shopify and Klaviyo actually do, what never to recommend, and how to prove it works. By industry: selling routines for skincare brands.

What is a recommendation engine in ecommerce?

An ecommerce recommendation engine is software that picks a short list of products for a given shopper at a given moment. It works like a good sales assistant: it knows the catalog, notices what the customer is looking at, remembers what they bought last time and suggests one or two things that fit. The difference is that it does this for every visitor, on every page, and can be tested.

It does two jobs. It filters, removing what shouldn't be shown (sold out, already in the cart, already bought). Then it ranks what's left, so the most likely purchase comes first. Most guides talk only about the ranking; in a real store, the filtering causes most of the complaints.

The engine produces different kinds of recommendations, each suited to a placement:

Recommendations are one part of personalization, and the ecommerce use is mostly cross-selling: adding the right second item to an order.

Where recommendations show up in a store

Each placement comes at a different point in the purchase, so it suits a different kind of recommendation:

The further down the funnel, the fewer products to show. A product page can carry a row of four to eight. A checkout should carry one. Stacking the same suggestion at every step annoys more than it sells: a shopper who declined the refill pack in the cart shouldn't meet it again as an order bump and again after payment. When a pair of products sells together often, a fixed bundle with its own price can work better than a recommendation.

Emails are where recommendations reach people who left. A post-purchase email suggesting the complementary product, or a replenishment email timed to the usual reorder gap, are both triggered email campaigns with a recommendation block inside.

What data drives recommendations

Most guides list browsing data first. For a store, the most useful data comes from orders and the customer record, because it's tied to money and it doesn't disappear with a cookie.

The last three rows are what separates a store's recommendations from a generic model. A subscriber who receives a product every month doesn't need it suggested again. A customer who refunded or disputed an item shouldn't see it in the next email. Both facts live in your billing and payment systems, not in your browsing data, so they need to be connected on purpose. Building segments from these signals is covered in how to segment customers.

Rule-based vs machine learning recommendations

Rules are lists you write: "show the cleanser with the serum", "show best sellers in this collection", "show the refill 25 days after the first order". They're easy to explain and they work for a small catalog or a launch. Their limit is upkeep: every new product needs new rules.

Machine learning engines learn the pairings from behavior. They come in three families, which every guide on the topic describes:

In practice, most stores use both: hand-picked complementary products where you know the right answer, and generated suggestions everywhere else. Shopify's own setup works this way.

Inside Tagada

Build the funnel once, route every customer to the right offer

Checkout, upsells and subscriptions on one flow, with conditions per country, order or customer.

Start free

Build or buy a recommendation engine

Most brands don't need to build one. The tools they already use include recommendations, and the official documentation states what each one does.

The details that matter when you set them up:

Specialized personalization vendors exist for larger catalogs. Judge them on the same points: which placements they cover, how they handle new products and visitors, which exclusions they apply, whether you can pin products by hand and whether they support a proper holdout. Their published case studies come from their own customers; your test is the only number that counts.

A custom build makes sense for a large catalog with heavy traffic and a data team. It means collecting events, computing pairings or training a model, serving results fast enough not to slow the page, and keeping a fallback for when it fails. For most DTC brands, the time is better spent on the inputs: clean product data, connected order and subscription data, and good exclusions.

What a recommendation engine should never show

A bad recommendation costs more than an empty slot. Shopify and Klaviyo already exclude some items by default: Shopify won't display a product that's sold out (related products can show if "continue selling when out of stock" is on; complementary products need inventory above zero), unlisted, priced at $0.00, a gift card, unpublished, or already in the visitor's cart.1Shopify Help Center, Search & Discovery: Product recommendations Klaviyo leaves out items without images, items the recipient already bought, out-of-stock items and the item that triggered the flow.5Klaviyo Help Center, How to use product feeds and recommendations The rest is up to you.

The last line matters for anything health-adjacent: supplements, skincare for a condition, sleep or weight products. A recommendation like "since you bought X for joint pain" turns a purchase into a health inference. Some US state laws treat information inferred from purchases as health data, and ad platforms restrict audiences built on conditions; the rules are covered in what behavioral targeting is. Keep recommendations product-to-product ("goes well with X") rather than condition-based.

The same care applies to how an add-on is sold. A recommended item should never be added to the cart for the shopper or offered with a pre-ticked box; the shopper adds it. The rules are in cart drawer upsells.

How to measure recommendations

A recommendation block always gets clicks. The question is whether it adds revenue that wouldn't have happened anyway. Shoppers who click "You might also like" are often the ones who were going to buy more regardless, so comparing them with everyone else overstates the effect.

How to size and read the test is in what is A/B testing, and the other levers that move conversion are in how to increase ecommerce conversion rates.

Where Tagada fits

Whatever runs your recommendations, the order of work stays the same: decide what each placement is for, connect order, subscription and refund data, set the exclusions, and keep a holdout to prove the result.