Cold start

Cold start comes in three forms. New-user cold start: someone arrives with no history, so collaborative filtering has nothing to match. New-item cold start: a product launched yesterday has no interactions, so no engine will recommend it. System cold start: the whole platform is new and there is not yet enough data to learn from at all.

Practical fixes are layered. For users: use what is known at arrival, such as acquisition channel, device, location, the first search or the landing page, and update rankings within the session as each click arrives. Short onboarding questions help when the product warrants them. For items: rely on content attributes (category, price, brand, description embeddings) so a new item is treated like its nearest existing neighbours, and give it deliberate exposure to gather early signal. For systems: start with rules and popularity, and let learned models take over as volume grows.

It is not a rare edge case. On most consumer sites the majority of visitors in any given session are anonymous or first-time, and catalogues turn over constantly, so cold-start handling is the main event, not the exception.

• Product manager: your first-visit experience is a cold-start experience; decide what signal to capture and what to show before the model has anything.
• Founder / CEO: if you are told a recommender will "learn your customers", ask what it does for the new customer you paid to acquire today.
• Data lead: item metadata quality is your defence against new-item cold start, and it is usually worse than anyone admits.
• Operations head: for retail, new-season stock is cold by definition; exposure rules decide whether it sells or sits.

The cold-start problem exists because learned personalisation depends on history, and history takes time to accumulate. Ignoring it produces a system that serves regulars beautifully and treats every new visitor and new product as invisible, which is exactly backwards for growth. Handling it means accepting a trade-off: the early experience relies on weaker signals and deliberate exploration, which costs a little conversion now to gather the data that pays back later. Teams that plan for it from day one get a system that improves within a session, not just across months.

• A D2C store ranking products for an anonymous visitor from the ad campaign and landing category, then re-ranking after the first two clicks.
• A marketplace giving new sellers' listings guaranteed exposure slots so they gather interaction data.
• A SaaS onboarding flow asking two questions about role and goal to seed template recommendations.
• An ed-tech app placing a new learner with a short diagnostic before recommending a path.
• A streaming or news product using content embeddings so a freshly published item can be recommended immediately.

ConceptA concept every personalisation design must address rather than a single technique. Eazyware treats cold-start strategy as a required deliverable in personalisation engines work, with session-level fallbacks specified before model training starts.

Eazyware service that covers it: Personalization Engines. Starting prices are on the pricing page.