Cold start recommendations: what to show shoppers you know nothing about

October 21, 2026

Short answer: The cold start problem is usually described as a data problem, but shoppers experience it as a relevance problem. A first time visitor lands on a store and is shown recommendations trained on everyone else's behaviour, which may have nothing to do with why they came. Most leave before the system learns a single useful thing about them, so the store's personalisation quietly serves only the returning minority.

The landing page already told you something

Cold does not mean clueless. The page a visitor arrives on, the campaign or search term that brought them, the product they opened first, and their rough location and device all narrow what sensible recommendations look like. A visitor who lands on a trail shoe from a running article should meet running gear logic, not the store's global best sellers.

Session behaviour compounds fast. Two product views and a filter choice are enough to shift recommendations from generic to plausible, if the system updates within the session rather than overnight. Stores running daily batch recommendations are personalising for yesterday's visitor, which for a one visit shopper is nobody.

Popularity is a starting point with known biases

Best sellers are the default cold start answer because they are safe, but popularity carries biases: it favours discounted lines, older products with long histories, and whatever the last promotion pushed. A cold start shelf built purely on sales rank shows new visitors the store's past, not its best current answer to their need.

Curated entry shelves beat raw popularity for considered categories. A small set chosen by merchandisers to represent the range, new in pieces, proven gifts, the reliable mid price option, gives a first time visitor orientation that an algorithm ranking by units cannot.

Ask, but ask cheaply

Explicit signals are gold when they cost the shopper nothing. A two tap preference prompt, style A or style B, or a quiz that returns value in under thirty seconds can warm a cold session instantly. The same quiz stretched to twelve questions becomes a gate that shoppers abandon, taking their data with them.

The key is that asking must visibly pay off. If a shopper picks three preferences and the recommendations do not change, the store has taught them that input is theatre. Immediate, obvious use of the answer is what makes the next shopper bother.

Measure cold start separately

Blended recommendation metrics hide the problem, because returning customers with rich histories perform well enough to cover weak first visit performance. Segment recommendation click through and revenue by session number and by new versus returning, and the cold start shelf gets the attention its traffic share deserves.

The first purchase matters beyond its value. A well matched first order creates the history that makes every later recommendation smarter, plus the account, the email relationship, and the return visit. Cold start is not a corner case; it is the top of the entire personalisation funnel.

Keep reading: How should recommendation ranking weigh margin and stock levels?, What belongs in checkout page recommendations, and what should stay off them?, and How does visual search change how shoppers find products? cover the nearby cases in this series.