The store

A fashion brand that sells only printed shirts, with fewer than 100 SKUs. The shirts share the same few cuts; what changes from one to the next is the print. They came to us because their competitors had added recommendations. Their first message was: "Hi, we'd like personalized recommendations."

Before the test

We told them up front that a store like theirs would probably not see a lift. They still wanted to test, and out of respect for their decision, we ran it.

How we tested

A 50/50 A/B test over 45 days, with add-to-cart rate as the primary metric.

What we found

No lift.

Why

Recommendations solve one problem: too many products to look through. When a shopper can see the whole catalog in a few scrolls, the products in a recommendation slot are ones she has just seen, so they tell her nothing new. When every product looks alike, showing "similar items" is the same as showing nothing different. And with so few products, there isn't enough browsing data for the model to learn from.

What we did next

We explained why, charged nothing, and suggested they put the budget into their brand story and what makes their fabrics different, since the cuts themselves don't vary much. When the catalog reaches around 500 SKUs, the test is worth running again.

There is a catalog size below which recommendations can't help. We'd rather tell you before the test than after.