Recommendation engine

What a recommendation engine does

A recommendation engine suggests products that fit a customer or an item. It works on similarity derived from behaviour, not from rules. Two basic methods are common. One calculates similarity between items: someone who bought A often buys B as well. The other calculates similarity between customers: customers who behaved similarly to you buy this item.

The two methods deliver different results. The item based method is simpler, more stable and works with few orders already. The customer based method is more precise but needs more data and is harder to explain.

When results become usable

The decisive size is not the number of visitors but the number of orders. A system learning connections from purchases can do nothing with a hundred visitors. Patterns that repeat and are not chance emerge only with some thousands of orders.

A second factor is the breadth of the assortment. In a narrow assortment even a simple rule finds usable recommendations. In a broad assortment a good system works considerably more strongly, because the choice is harder for the customer.

How the effect is measured

• Click rate on the recommendation: shows whether the suggestions are noticed at all. On its own it says little.
• Share of orders containing a recommended item: the more telling figure, because it concerns the actual purchase.
• Basket value of the group involved: if it rises, the recommendation worked. The comparison is against a group without recommendations, not against last month.
• Coverage: which share of the assortment is recommended at all. A system always showing the same twenty items only helps those.

Which effects cancel the benefit

The first effect is reinforcement. A system that only confirms what was already bought suggests the same thing again and prevents discovery of other items. Over time the variety in the basket falls.

The second effect is suggestions at the wrong moment. A recommendation immediately before completion distracts from the purchase. A pointer to matching accessories before buying brings more than a list of similar products beside it.

The third effect is visible mistakes. A suggestion offering an already bought item again, or accessories for a returned order, looks inattentive and lowers trust. The fourth effect is the missing explanation. Anyone who cannot say why something is shown treats the adaptation as chance rather than service, and resistance appears where predictability is missing.

One rule follows in practice: use recommendations where the benefit is immediately obvious, such as matching accessories, and leave the rest of the page untouched. Few good suggestions work better than many arbitrary ones.

Frequent questions

• Do I need a tool for it? Not for simple rules. From some thousands of orders a system pays off.
• How much data does it need? Some thousands of orders for patterns to be solid.
• Does it replace manual cross selling? No, curated selection stays better for goods that need explaining.