Cold-start recommendations: what to show when there is no history

Cold-start recommendations have to work with no history: new visitors, new products, or new stores with no behavioral data at all. The toolkit is cohort behavior for new visitors, content-based matching and deliberate exploration for new products, and editorial curation plus aggressive instrumentation for new stores. Every cold start is temporary if you collect the data that ends it.

The three cold starts

The cold-start problem has three distinct forms, and they need different solutions. New visitors have no history, but the catalog is rich with behavioral data from others. New products have no behavioral data, but visitors have histories. New stores have neither, which is the hardest case and the one most advice ignores.

Most recommendation engines are built for the warm middle: known users, established products, deep history. The cold edges are where first impressions form and where new products live or die. A store whose recommendations only work for returning customers is personalizing for the minority of its traffic.

Borrowing signal for new visitors

For a visitor with no history, the best available proxy is the behavior of similar visitors. Traffic source, entry page, device, and geography cluster visitors into cohorts whose collective behavior is a decent stand-in for individual taste. 'Visitors from this campaign who viewed this category went on to buy these' is a real recommendation with no personal data required.

Session behavior fills in the gaps within minutes. The first few product views define a taste vector; the engine should update its recommendations after every view, not after the session ends. By the fifth pageview, a cold visitor is measurably warmer, and the recommendations should reflect that.

Warming up new products

New products need impressions to earn behavioral data, but the engine will not give impressions to products with no data. Break the deadlock with deliberate exploration: reserve a fixed share of recommendation slots for new and low-data products, and measure their engagement separately from the main ranking.

Content-based matching accelerates the process. A new product's attributes, description, and imagery can be matched to the products it resembles, inheriting a provisional taste profile from its nearest neighbors. When the real behavioral data arrives, it overrides the borrowed profile. The goal is a sensible ranking on day one, not a perfect one.

The new-store playbook

A brand-new store has no behavioral data at all, which means the first recommendations must come from elsewhere: vendor bestseller data, category benchmarks, editorial curation, and content-based similarity built from the catalog alone. This is the phase where merchandisers earn their keep, because human judgment is the only signal available.

Instrument everything from day one. Every impression, click, and purchase should feed the future engine, even while the current recommendations are curated. Most stores under-invest in event tracking during the launch phase and then discover, months later, that their recommendation engine is starving for the data they failed to collect.

Knowing when the cold start ends

Cold start is not a permanent condition; it is a phase with an exit. Define the thresholds explicitly: a visitor graduates from cold-start recommendations after a set number of product views or a first purchase; a product graduates after a set number of impressions with engagement. The engine should blend from cold-start strategies to behavioral strategies gradually, not flip a switch.

Measure the blend. Track recommendation click-through separately for cold versus warm traffic, and watch the cold-start metrics improve as your strategies and your data mature. The stores that win at recommendations are not the ones with the best algorithm; they are the ones that made the cold edges work while the data was still thin.