Recommendations for a new category launch: no history, no co purchases, no problem to hide behind

For the launch window, the recommendations in and out of the new category should be hand picked by whoever owns the range: these products with those, this new line beside the established one it was bought to sit with. Hand picked feels like a failure of the personalisation programme, which is why teams avoid it and let an empty engine show bestsellers from unrelated categories on the launch pages instead. One of those choices looks like merchandising. The other looks like a bug.

Set the handover date when the editorial period is agreed, not when someone remembers. The trigger is data volume: enough sessions and orders in the category for co purchase patterns to be more than anecdote. Naming the trigger in advance stops the editorial set fossilising, which is the known failure mode at the other end, a launch rail still showing someones picks from two ranges ago.

Editorial first, and say so internally

Between editorial and behavioural data sits attribute matching, and for a new category it does real work: recommend by shared use, material, price band, and compatibility attributes the range team defines at setup. A new category launched with its attributes properly built can run sensible frequently bought logic from day one, because the attributes encode what the buyers already know about what goes with what.

This is the moment to insist on the attribute work, while the category is small enough to do completely. Retrofitting attributes onto a mature category is a project everyone postpones. A launch has no excuse: the products are few, the team is attentive, and every attribute skipped becomes a recommendation gap that gets blamed on the engine in the first quarterly review.

Attributes carry the interim logic

The new category has no history, but its shoppers do. Use the traffic honestly: visitors arriving from the adjacent category are telling you the intended relationship, and launch recommendations on the established category pages, the new line presented as the extension it is, borrow an audience that already trusts the context. Search queries in the first weeks are the other borrowed signal: the words shoppers use for the new products are the words the category pages and product data should adopt immediately.

Resist importing a lookalike categories co purchase graph wholesale. Adjacent is not identical, and recommendations trained on the neighbouring category will confidently pair the new products with partners that make sense over there and none here. Borrow audiences and vocabulary. Do not borrow conclusions.

Borrow signal from the neighbours

Handover is not a switch. Behavioural signals earn weight as they accumulate: clicks first, because they arrive first and mean least, then add to basket patterns, then co purchase, which is the strongest and the slowest. A staged weighting lets the team watch each signal agree or disagree with the editorial set, and disagreement is information: either the shoppers see a pairing the range team missed, or the data is still too thin to trust, and the counts tell you which.

Keep one editorial override permanently: the ability to exclude a pairing that is statistically real and commercially wrong, the loss leader the engine pairs with everything, the product with the returns problem the graph cannot see. An engine with no override will eventually recommend something the whole company can see is daft, and the override is cheaper to keep than the meeting that follows.

Let the data take over in stages

Keep reading: How should recommendations change when a shopper keeps returning the same category?, How do you keep recommendation carousels from showing the same five products everywhere?.