A visitor arriving from a targeted ad, with no prior store activity, starts as a blank slate for personalization. There is no purchase history, browsing history, price response, or category preference to draw on. That is the cold start problem, a major ecommerce product discovery challenge because new visitors make up a substantial share of daily retailer traffic.
Cold start affects recommendation systems beyond ecommerce. Ecommerce faces extra urgency because a new visitor may arrive at peak purchase intent from a targeted ad, referral link, or external search query. Showing irrelevant products during that first session is an expensive failure, not a missed opportunity easily recovered on the next visit.
What Cold Start Means in Practice
Cold start is the condition where a recommendation system lacks enough behavioral data to personalize output for a person. "Insufficient" matters. Cold start has degrees, not a simple yes or no, and the response should match the information available.
A fully cold visitor has no previous store history and may have no referral context. At first page load, the available information is limited to the landing page, passed referral source, and URL parameters from the traffic campaign.
A partially cold visitor has limited history. They may have visited three months ago, viewed two products, and left without purchasing. The record exists, but it is thin, and preferences may have changed. This can be harder than a fully cold case because the data can bias ranking without reliably showing current interests.
Session Signals From Minute One
The key to cold start is recognizing how quickly a new visitor can stop being fully cold, provided the system listens. Each action adds information. A search query entered in the first thirty seconds may be the strongest intent signal available, often better than historical data gathered over time.
Typing "men's running shoes under $120" tells the ranking system the shopper's category, gender context, and price sensitivity at once. Ignoring that signal and continuing to show catalog-wide bestsellers is a substantial mistake. The session signal is already available.
First-session product page dwell time also helps. Forty seconds on a product detail page indicates substantive engagement, while a three-second bounce suggests irrelevance. Engagement across the first few product touches helps the model form a session-specific preference picture before a purchase signal appears.
Context Signals for Truly Cold Visitors
A fully cold visitor may have produced little useful behavior, but non-behavioral context can still provide signals at the session's start.
Referral context is useful. Someone arriving from a paid social post about a category is probably considering that category. A brand search visitor likely has higher purchase intent and lower price sensitivity than someone from a broad organic query. UTM parameters and referrer data are limited, but useful.
Device type and time of day offer some predictive information about category propensity, though the signal is weak. A mobile fashion browser at 9pm has a different context from a desktop visitor at 11am. Neither should lead ranking, but both can serve as weak priors when stronger signals are absent.
Geography is often overlooked. An Australian visitor browsing home goods may have different preferences from someone in another market, especially where categories have local relevance. Geographic context can therefore help with cold start.
Popularity Priors for Cold Start
When session and context signals are limited, use carefully built popularity priors instead of basic bestsellers. A naive bestseller list reflects aggregate demand, which may say little about what the visitor currently wants.
Stronger popularity priors reflect arrival context. For a social visitor, products most popular among social referrals in the past 14 days are more relevant than overall bestsellers. For a first-time visitor, products popular with first-time shoppers and weighted by recent purchase velocity are more useful than lifetime bestsellers shaped by one promotional campaign from eight months ago.
Narrow the reference group as far as context permits, so popularity reflects shoppers resembling the person being served. This is not full per-shopper personalization. It is contextual popularity ranking, and it is a meaningful improvement over catalog-wide bestsellers for cold start.
How Fast Cold Start Resolves
Cold start can resolve quickly within one session when real-time signals are available. In our early-access testing, once a new visitor viewed four or more product pages, the ranking model had enough signal to beat contextual popularity on revenue-per-session for the rest of that session.
A focused shopper may view four products in three to four minutes. That is brief, so cold start occupies only a small part of a deeper session. Handle the opening minutes well, then let session signals guide ranking for the remainder.
In a shallow session of two page views followed by a bounce, cold start dominates because the model never receives enough information to engage. Such visits are harder to recover, and some are exploratory sessions that would not have converted regardless of ranking.
Cross-Session Learning for Return Visits
Cold start shrinks with each return visit only when signals carry across sessions. Treating every visit as independent discards that history. A shopper returning for the fourth time after consistent browsing in one category can otherwise still appear nearly cold.
Cross-session identity stitching is a non-trivial engineering challenge, especially without login. The three most reliable continuity methods are probabilistic matching through device fingerprint, email from cart abandonment flows, and loyalty program login. Even incomplete identity capture reduces cold start over time, since most high-value repeat visitors have at least one anchoring identifier linking their sessions.