WooCommerce

AI Product Recommendations in WooCommerce: Personalization Without Overdoing It

ยท Krzysztof Balicki

Wicker basket, a box and a tablet on a wooden counter

AI product recommendations in WooCommerce work when they have something to learn from (orders, browsing), sit in a few well-chosen places in the store and pass an A/B test before they get all the traffic. That’s the short version. I’m writing this for the store owner, operations director or IT manager who wants to outsource this kind of project and needs some way to judge a developer’s offer. We’ll cover what a sensible rollout looks like, what to ask before signing a contract and where to start in the first place.

What are AI product recommendations and how do they differ from regular related products?

They are suggestions calculated by a model based on what customers do. Not just upsells and cross-sells that someone once clicked together in the admin panel. And manual links in WooCommerce have a problem: someone on the team has to pick products for every single item. With a large catalog, some of them are out of date after a few months. AI-powered related products change along with the data. New purchases come in, the season changes, and the store starts showing something different on its own.

These blocks usually end up on the product page, in the cart, on the thank-you page after checkout and in transactional emails. But to be honest: with a small catalog and a few dozen orders a month, simple rules like „accessories from the same category” are often more than enough. The model has nothing to learn from, and that’s that.

What data do you need for AI store personalization?

Order history and browsing information. You don’t need customers’ personal data for this. From orders, the model learns what gets bought together, who comes back and how sales rise and fall over the year. From browsing - which products and categories someone viewed and what was left in abandoned carts.

You can get orders and products through the REST API in the WooCommerce documentation. It returns orders and products as JSON, and authentication uses a consumer key and consumer secret generated in the store’s admin panel. Tracking browsing requires cookie consent. Without it, recommendations can rely on orders and the context of the page the customer is currently viewing.

Personalization without overdoing it: where to draw the line

A good recommendation makes the decision easier. A bad one pushes a product the customer just bought, or hides what they were actually looking for. In practice, the same three mistakes keep coming back: recommending something that was just purchased, several „you may also like” blocks crammed onto one page, and suggestions so personal that the customer starts wondering how the store knows so much about them (and that doesn’t help sales, it hurts them).

That’s why business rules have to sit on top of the model. You exclude certain categories, skip what’s out of stock, boost higher-margin items and hide the ones you don’t want to promote at all. And then there’s performance. Calculating recommendations must not slow down the product page, so load them asynchronously or from cache.

What a good AI product recommendation rollout looks like, step by step

It starts with a goal and one place in the store. Not with choosing a model, that comes later. An order that works:

  1. Goal and metric, for example average order value or the share of orders with a product added from recommendations.
  2. Data audit: number of orders, catalog completeness, quality of categories and attributes.
  3. One starting location, usually the cart or the product page.
  4. Business rules: exclusions, stock levels, margin.
  5. A/B test on part of the traffic.
  6. Decision to expand to more locations or to make fixes.

A/B testing product recommendations compares the new block with what you currently have in related products. You set the duration and the metric before the start. Not after the fact, when it’s tempting to pick whichever number happens to look good. And one more thing: nobody honest will promise a specific conversion lift up front. Only a test in your own store will show that result.

What should you ask a developer before hiring them?

A few questions quickly show whether a developer is thinking about data, privacy and measuring results, or just about where to paste a widget. In my view, before choosing an offer you should ask:

  • where the model gets its data and on which servers it is processed,
  • what the store shows a customer who did not accept cookies,
  • how the A/B test will be run and who will evaluate the result,
  • which rules the store team can set on their own,
  • how the solution affects product page speed,
  • who maintains the solution after WooCommerce and WordPress updates.

A ready-made plugin works for a standard catalog and typical needs, because it’s quick to set up. But if the store has unusual pricing rules, an ERP integration or its own margin logic, a custom solution for your store will do better, built around your data and processes.

Where to start with store personalization next quarter

With a data review and one recommendation block, in the cart or on the product page. Before you sit down with a developer, prepare:

  • REST API access with limited-permission keys,
  • a short catalog description: number of products, categories and seasonality,
  • your current cross-sell and upsell rules,
  • your current cookie policy and how consent is collected.

With this set in hand, the conversation is about specifics, not generalities. Want to check whether AI product recommendations make sense in your WooCommerce store? Describe your catalog and current product links in our project quote request form. We’ll reply with a proposed first step and a test plan.

Frequently asked questions

Do AI recommendations work in a small WooCommerce store?

It mainly depends on the number of orders, because the model learns from repeat purchases. Are your sales small? Then it’s better to start with simple rules based on categories and attributes. You can add the model later, once some history builds up.

Do product recommendations require cookie consent?

Tracking browsing and building a customer profile - yes, that requires consent. However, recommendations calculated from store-wide orders and the context of the page being viewed can work without profiling a specific person. In both cases, it’s a good idea to describe how it works in your privacy policy.

How do you check whether recommendations are working?

With an A/B test on part of the traffic, where the new block goes head to head with your existing related products. You set the metric and duration before the start. And you decide on expanding after the test ends. Not after the first three days, when the chart just happens to be heading nicely upward.