Every online store owner has seen the "you may also like" row on Tokopedia or Shopee and wondered how much it actually sells. The answer, based on what large e-commerce players have published over the years, is a lot. Amazon has long attributed a meaningful share of its revenue to recommendations, and every marketplace since has copied the pattern for a reason.

Here is what most SME owners get wrong about it: they assume a product recommendation engine requires machine learning, a data science team, and a budget they do not have. So they skip it entirely, and every product page on their store becomes a dead end. Customer views item, customer buys or leaves, transaction over.

The truth is that a recommendation engine is a spectrum, and the cheap end of that spectrum captures most of the value. Let me demystify how the logic actually works, and show you where a store your size should start.

What a Recommendation Engine Actually Does

Strip away the buzzwords and a product recommendation engine answers one question: given what this customer is looking at or has bought, what is the most likely next thing they will want?

There are only a few fundamentally different ways to answer it:

  • Hand-curated rules. A human decides: anyone viewing this rice cooker should also see the measuring cup, the spare inner pot, and the steamer basket. Zero math, pure merchandising knowledge.
  • Co-purchase statistics. Look at your order history. Which items appear in the same basket unusually often? Phone case with screen protector. Paint with brushes and thinner. This is "frequently bought together," and it is just counting.
  • Collaborative filtering. The step Amazon made famous: customers who bought what you bought also bought X, even when the items have no obvious connection. The system finds patterns across thousands of customers that no merchandiser would spot, like a specific coffee grinder correlating with a specific brand of kitchen scale.
  • Content-based matching. Recommend items with similar attributes: same category, same brand, adjacent price point. "Similar products" rows are usually this.

Notice that only the third one is really "machine learning" in any meaningful sense. The first two are things you can do with a spreadsheet and your existing order data.

Why It Works: The Basket Math

The commercial logic is simple. Suppose your store does 1,000 orders a month at an average basket of Rp250,000. You do not need dramatic results from cross-selling. If recommendations nudge just 6 percent of orders to add one Rp75,000 item, that is an extra Rp4.5 million a month, roughly Rp54 million a year, from customers you already paid to acquire.

That last point is the one to sit with. Every rupiah of extra basket size comes at zero additional marketing cost. You already won the hard battle, which is getting the customer to your store with intent to buy. Recommendations simply stop you from leaving money on the table at the exact moment the wallet is open.

There is a second, less obvious benefit: recommendations rescue dead ends. A product page for an out-of-stock or poorly matched item normally loses the visitor. A good "similar products" row converts a would-be bounce into a browse.

Start Manual: The 80 Percent You Can Build This Month

Here is my honest advice for a store doing under a few thousand orders a month: do not touch machine learning yet. Start with curated pairing rules, because you will capture most of the lift at a fraction of the cost, and you will learn what your customers actually respond to.

The playbook:

  1. Pull your last 6 to 12 months of orders and find the top 20 products by revenue. These pages get the traffic, so they get the recommendations first.
  2. For each, define two lists. Complements: things used together with it (the cross-sell). Alternatives: things bought instead of it, at similar or slightly higher price (the upsell and the dead-end rescue).
  3. Mine your own baskets for surprises. Even a pivot table over order data will show you pairs that co-occur more than chance. Some will surprise you, and those are your best recommendations precisely because customers had to work to find the second item on their own.
  4. Place them where decisions happen. Product page for complements and alternatives, cart page for small add-ons under about 20 percent of the basket value. The cart is prime real estate: the buying decision is already made, so a Rp30,000 add-on faces almost no resistance.
  5. Measure one number: average items per order, before and after. If it does not move within two months, your pairings are wrong, not the concept.

Most store platforms already support this. Shopify has manual related-product options and cheap apps for it, WooCommerce has linked products built in, and if you run a custom storefront, this is a small feature for your developer, days of work, not months.

A practical anecdote: a home goods store I advised added hand-picked complements to only its top 15 product pages. Items per order went from 1.3 to 1.5 in about ten weeks. No algorithm, just a merchandiser's knowledge finally put in front of the customer at the right moment.

When to Graduate to the Automated Version

Manual rules have a ceiling. They cover your top products, they go stale as your catalog changes, and they only encode what you already know. You are ready for automated recommendations when:

  • Your catalog is too large to curate, roughly 500 or more active SKUs.
  • You have real data volume, thousands of orders, because collaborative filtering is statistics and starves on small samples.
  • The manual version already proved the lift, so you know the ROI case for going further.

At that point you rarely need to build from scratch. Recommendation is a mature problem: your platform's app ecosystem, or a modest custom service built on your order history, will do. Treat it like any other investment with a payback calculation, not like a research project. The same discipline applies as with any data-driven tooling; a forecasting system follows an identical crawl-walk-run path, which I covered in A Retail Chain Used Forecasting to Tame Stockouts.

One warning from experience: an automated engine trained on thin data produces embarrassing suggestions, and irrelevant recommendations are worse than none because they erode trust in your whole store. This is why the sequencing matters.

The Takeaway

A product recommendation engine is not one technology, it is a ladder. The bottom rungs are curated complements and "frequently bought together" rules built from your own order history, and they deliver most of the revenue lift for almost none of the cost. The top rungs, collaborative filtering and real personalization, only make sense once your catalog and order volume outgrow human curation.

So the action item is unglamorous: this week, open your order history, find your top 20 products, and write down what should be recommended next to each one. Then get those rows onto your product and cart pages. Measure items per order. Everything more sophisticated can wait until this stops working.

The stores that win are not the ones with the fanciest algorithm. They are the ones that stopped letting every product page be a dead end.