AI Recommendation Engines: Rules vs Learning, Plugin or Custom?

An AI recommendation engine is the system that automatically fills blocks such as "similar products," "frequently bought together" and "recommended for you" on your online store, based on visitor behavior and product information. Set up well, it makes a difference in two places: it raises basket value by suggesting complementary products, and it drives repeat purchases by reminding customers at the right moment. For most stores it makes sense to start with an off-the-shelf solution; a custom build becomes worth discussing once your own business rules, such as margin, stock and campaigns, have to shape the recommendations.

What is the difference between rule-based and AI-powered recommendations?

Recommendation systems essentially combine three approaches. Two of them learn from data, and one is written by people:

• Behavior-based recommendations (collaborative filtering): The "customers who bought this also bought" logic. It looks at which products are viewed together, added to the cart together and purchased together.
• Content-based recommendations: Finds similar products from attributes such as category, brand, color, price range and description. It works for newly added products that have no sales history yet.
• The rule layer: Rules the business defines, such as removing out-of-stock items, pushing a particular brand or prioritizing products in a campaign.

A purely rule-based setup uses only the third layer, so it does exactly what you tell it and nothing more. A learning-based engine uses the first two to decide what to show, and good setups keep the rule layer on top as a guardrail.

The documentation for Amazon Personalize, Amazon's managed recommendation service, summarizes this structure well: in most scenarios, recommendations rely mainly on data about how users interact with items, meaning events such as clicks and purchases, and in some cases they also use item and user information such as price or category. The same documentation gives "recommended for you," "frequently bought together" and "customers who viewed X also viewed" as e-commerce examples.

Similar products vs complementary products: why does it matter?

This is the distinction people confuse most often when planning recommendation blocks. Shopify's Search & Discovery app splits it clearly in two: related products, which resemble the selected product and could be bought instead of it, and complementary products, which are bought alongside it.

Here is why it matters. Showing similar products to an undecided visitor on a product page widens the choice, but it does not raise basket value. Complementary products do. Someone buying a coffee machine should be offered coffee and cleaning tablets, not another coffee machine.

In Shopify's case, automatic related product recommendations are generated from purchase history, product description similarity and related collections, but the description similarity method only works on English-language storefronts. If your store runs in another language, knowing this helps you set the right expectations. The app lets you hand-pick up to 10 complementary and 10 related products per product, and your theme needs to support these sections for the recommendations to appear.

Where should recommendation blocks go?

Each placement serves a different purpose and should be measured separately:

• Product page: Below the product information, not above it. Similar products for the undecided visitor, complementary products for the one who has decided.
• Cart page: Low-priced complementary items. Offering a small product that takes a customer over the free shipping threshold is the most productive use of this spot.
• Post-purchase email: Consumables, spare parts or accessories for what the customer just bought. This is the main source of repeat sales, and it is planned together with your email and SMS marketing.
• Home page: A selection from the category a returning visitor last browsed. For first-time visitors, best sellers are enough.

Should you use an off-the-shelf solution or a custom build?

There are three paths, and the difference between them is a trade-off between flexibility and effort.

Your platform's built-in feature or a ready-made plugin: The built-in recommendation features of platforms such as Shopify, ikas or WooCommerce, or plugins from their app stores, can go live in a few days. No technical team is required. The limit is how far you can influence the recommendation logic; pushing high-margin products or dropping products with high return rates may not be possible in every plugin.

A managed cloud service: Services such as Amazon Personalize train the recommendation model on your data, and your site pulls the results through an interface. Flexibility is high, but you need software work to build the data flow and place the results on your site.

A fully custom build: The recommendation logic is written for your catalog, margins, stock and return data, according to your own rules. You have full control; in return, it takes effort to build, test and monitor continuously.

A practical roadmap: start with an off-the-shelf solution and measure each block's effect on clicks and sales for at least a few months. If a business rule the ready-made solution will not let you apply is genuinely costing revenue, moving to a custom build becomes a decision based on data rather than guesswork.

Is your data ready before you set it up?

No matter how good the engine is, messy data will not produce good results. Check these before you start:

1. Is the catalog clean? Are category, brand and attribute fields filled consistently, and have duplicate records of the same product been removed?
2. Are events being tracked? Are product views, add-to-cart and purchase events recorded correctly and with the product ID?
3. Is there a fallback for new products and new visitors? For a product or visitor with no history, define a fallback rule such as best sellers or popular products in the same category.
4. Have you thought about cookie consent? If you serve visitors in Turkey, the Turkish Personal Data Protection Authority's Cookie Guidelines state that cookies used for personalization the user has not explicitly requested must rely on explicit consent. Design the setup so visitors who decline see general, non-personalized recommendations, and review the legal position with your advisor for every market you sell in.

How do you measure whether the recommendation engine works?

The most common mistake is counting every product bought after a click on a recommendation block as "revenue generated by the engine." Some of those customers would have found and bought the product anyway. For a more honest measurement, run an A/B test that shows the block to some visitors and not to others, and compare the two groups on these metrics:

• Average order value
• Revenue per visitor
• Repeat purchase rate within a set period
• Block click-through rate and add-to-cart rate from the block

Run the test for at least a few weeks, during a period that does not overlap with a campaign.

If you would like to integrate a recommendation engine into your store, or check whether your current platform is suited to it, see our AI integration services and e-commerce development service, and our pricing page for an itemized view of costs.