The cold-start problem: recommending products with no history
Cold start is the gap between a product launching and the recommendation engine knowing anything about it. The practical answer is content-based similarity: recommend the new product alongside products it resembles in attributes, category, and price, while giving it deliberate exposure so it can earn its own behavioral data. Measure cold-start coverage, the share of catalog that ever gets recommended, and treat products the engine ignores as a defect, not a footnote.
Where cold start bites
Cold start hits in two places. New products arrive with no clicks, no purchases, and no co-view data, so collaborative filtering has nothing to say about them. New visitors arrive with no history, so personalization has nothing to personalize with. Most stores have one of these problems constantly and both of them during launches, seasons, and traffic spikes.
The business cost is quiet but real. Products that never get recommended effectively do not exist for the recommendation-driven share of revenue, which means launches underperform for algorithmic reasons rather than product reasons. Merchandisers blame the product; the engine was the problem.
Content-based similarity as the bridge
When behavioral data is missing, product attributes fill the gap. A new product is similar to existing products in category, price band, brand, material, style tags, and any structured attributes the catalog carries. Recommending it alongside its nearest neighbors gives it immediate placement that is usually relevant, because products that resemble each other tend to appeal to the same buyers.
The quality of this bridge depends on the catalog data, which is why attribute completeness is a recommendation problem, not just a merchandising nicety. Every missing attribute is a dimension along which the new product cannot be matched.
Deliberate exposure: the explore-exploit tradeoff
Similarity gets the new product onto the rail; deliberate exposure gets it data. Reserve a slice of recommendation slots for exploration: new and underexposed products shown to small, representative audiences so they can earn clicks. This costs some short-term relevance, since an explored slot could have held a proven product, but it buys the behavioral data that makes the engine smarter next week.
The common mistake is exploring without a plan: random new products in random slots, with results nobody reads. Exploration needs the same discipline as any test, defined audiences, minimum exposure, and a decision rule for what happens when the data comes in.
Measuring cold-start coverage
Track what share of the catalog ever appears in recommendations over a rolling window. If a fifth of your products get zero recommendation impressions, the engine is running a popularity contest, not a merchandising strategy. Break coverage down by product age: products under thirty days old should have a defined path to their first thousand impressions.
Pair coverage with performance. New products will convert worse than proven ones at first; that is the price of data. The question is whether the explored products graduate into the regular rotation, and how fast. An exploration program where nothing ever graduates is just a tax on relevance.
How long does cold start last for a new product?
Until it has enough interactions for the behavioral models to place it confidently, usually a few hundred views with normal engagement. Deliberate exposure shortens this from months to weeks.
Should new visitors see the same defaults as everyone else?
No. New visitors should see category bestsellers and trending products, which are the safest defaults, while the system learns. Showing them niche long-tail picks is personalization theater.
Can merchandisers override cold start manually?
They should be able to feature launches in curated slots, but manual featuring is not a substitute for the engine learning. Use it to buy time, not to replace the data.
Reviewed
Published Oct 2, 2026.
Reviewed
Published Oct 1, 2026.