What is a cold start?
A cold start occurs when a discovery or recommendation system has too little data to make confident, well-informed decisions. The system doesn’t know enough about a product, about a user, or about both, to surface the right thing at the right moment.
In e-commerce, this is more common than in services where users come back again and again, like Spotify or Netflix. The majority of shoppers have returned to online stores to browse and shop much less frequently, and e-commerce sites have less data to work with.
Cold starts fall into two broad categories:
Cold-start products are those that lack sufficient interaction history. This includes new arrivals that haven’t yet been clicked or purchased, long-tail products that rarely appear in search results, items in newly introduced categories, and products that have simply never gained enough visibility to accumulate meaningful data.
Cold-start users are visitors the system knows little or nothing about. New visitors arriving for the first time. Infrequent shoppers who return once a season. Anyone without a purchase history that the system can learn from.
Both create the same fundamental problem: without data, the system defaults to generic. And generic, in a world of personalized experiences, is increasingly not good enough.
The cold-start user: The majority, not the exception
It is tempting to frame cold-start users as an edge case, the occasional new visitor, the rare anonymous session. In reality, for most retailers, cold users are the majority.
Consider the typical distribution: a relatively small share of customers, often around 10%, accounts for the majority of purchases. The rest are infrequent, sporadic, or visiting for the first time. That means most sessions on most e-commerce sites are served by systems with very little to work with.
Retailers who focus their optimization efforts entirely on their most loyal customers, maximizing lifetime value among those who are already engaged, are solving the easier problem. The harder, more commercially significant opportunity is improving the experience for everyone else: the first-time visitor comparing options, the seasonal shopper returning after a year, the customer who came in through a Google search and has no history with the brand at all.
These visitors are not lost causes. They are customers in the making. But they need a discovery experience that doesn’t rely on a data history they don’t yet have.
The loop that keeps products invisible
The cold-start problem has a particularly cruel internal logic. A product that isn’t shown generates no clicks. No clicks means no data. No data means the system has no reason to show it. The product stays cold indefinitely, not because shoppers wouldn’t want it, but because it never got the chance to prove itself.
This is especially punishing for long-tail catalogs. Retailers often invest significantly in product depth, stocking a wide range of SKUs to serve diverse customer needs, only to find that the majority of that catalog is effectively invisible. It sits on the digital shelf, generating no revenue and attracting no attention.
The consequences extend beyond lost sales on individual products. Cold catalogs inflate inventory costs, complicate buying decisions, and make it harder to understand what customers actually want.
In different retail sectors
Cold starts are not a theoretical risk. They manifest differently, and with different degrees of severity, across the major verticals of e-commerce.
Fast fashion
Fast fashion is perhaps the sector most acutely exposed to the cold-start problem. Collections turn over rapidly, sometimes weekly, meaning that new products are constantly entering the catalog before the previous ones have had time to accumulate meaningful data. By the time a garment has enough interaction history for the system to recommend it confidently, it may already be heading toward markdown.
Speed is the entire commercial logic of fast fashion. A discovery system that needs weeks to “warm up” a product is fundamentally misaligned with the business model. Cold starts here don’t just mean missed sales, they can mean entire collections that never find their audience before the window closes.
Sports goods and seasonal sports clothing
In sporting goods, cold starts arrive on a schedule. Ski jackets are irrelevant in July and critical in October. Running gear surges before the spring race season. Winter cycling apparel has a narrow, well-defined window of relevance.
The challenge is that when a seasonal product re-enters the catalog, or when a new seasonal line launches, it arrives cold. Last year’s data, if it exists, may be partially applicable, but a new colorway or updated specification is still a new product from the system’s perspective. Getting these products visible quickly and ranking them intelligently against established items requires a discovery infrastructure that can work with thin data and compensate with contextual signals.
The stakes are high. Miss the window on a winter jacket, and no amount of spring discounting fully recovers the margin.
Furniture
Furniture presents a different dimension of the cold-start challenge. Purchase frequency is inherently low, most customers buy a sofa once every several years, if that. But when you are shopping for a new sofa, many shoppers check it out more than once before purchasing it. This means the vast majority of a furniture retailer’s website visitors are, from the system’s perspective, effectively cold users until they briefly become frequent. When they show up after several years, they likely have no purchase history, minimal browsing history, and few signals to work with
The long-tail problem is also acute. A retailer might carry thousands of SKUs across sofas, dining tables, storage, lighting, and accessories. Many of these, particularly in niche styles or configurations, will never accumulate enough interactions to be recommended with confidence through data alone.
In furniture, winning the cold-start problem is winning the discovery problem outright, because cold is the default state for most products and most visitors.
Consumer electronics
Electronics catalogs refresh with product cycles, but the cold-start problem here is often most acute at the moment of highest commercial opportunity: launch. A new smartphone, a newly released games console, or a flagship laptop arrives in the catalog with no interaction history, precisely when customer intent is at its highest.
There is also a meaningful long-tail in accessories and peripherals, where hundreds of compatible products compete for visibility, and only a fraction ever accumulate enough data to be recommended organically. A customer searching for a compatible charger or a specific cable is likely to find the top three options repeatedly surfaced, while dozens of alternatives, potentially better suited to their needs, remain effectively invisible
How good systems address cold starts
There is no single solution to cold starts, but the strongest discovery approaches share a common philosophy: don’t wait for data. Find alternative signals, apply intelligent defaults, and create the conditions for data to accumulate.
For cold-start products, this often means combining what the system knows about the product’s attributes, category, material, style, price point, with signals from structurally similar products that are already performing. And it means building ranking logic that doesn’t penalize products simply for being new.
For cold-start users, it means leaning on contextual signals in real-time: What is the user searching for right now? What page did they arrive on? Are there trending sales behaviors that are relevant, the session behavior, or the entry point? These signals won’t replace a rich purchase history or segment created based on known first-party data, but they can inform a more relevant experience than a purely generic default.
In session, personalized search plays a particularly important role here. When a cold user types a query, they are providing an explicit, high-quality signal of intent. A discovery system that interprets that query with precision, matching words accurately, applying intelligent classification, using structured product data to understand what the user is actually asking for, can deliver a sharp, relevant result even with no prior knowledge of who that user is. Search done well is one of the most powerful tools for serving cold users effectively.
Product listings are more complex because they must balance multiple objectives simultaneously: personalization where signals exist, commercial goals around visibility and margin, and the need to give new and long-tail products a fair chance to compete for attention.
The bigger picture
Cold starts are often treated as a technical problem, something to be solved in the model layer, managed through engineering by creating a ton of segments. Or worse, a retailer’s resort to seeing it as a pure merchandising problem, to be managed by rules and promotions when that product sits on the shelf too long. But the business implications are strategic.
A retailer that solves cold starts is one that can launch new products with confidence, knowing they won’t disappear into the catalog. It is a retailer that can serve first-time visitors with experiences that feel relevant rather than random. It is a retailer that gets more value from the full breadth of its catalog, not just the established bestsellers.
The long tail is only long if it stays cold. The right discovery infrastructure can change that.
How Voyado approaches the problem
At Voyado, we have built our product discovery capabilities with cold-start scenarios as a first-class concern, not an afterthought.
In search, our approach is grounded in precision. By focusing on accurate query interpretation and semantic search, structured product data classification powered by retail intelligence, and real-time optimization in ranking to surface the perfect match, we can deliver sharp, relevant results even when we know very little about the user. When someone knows what they are looking for, we make sure they find it, regardless of whether we have seen them before.
In product listings, the challenge is more nuanced. This is where our hybrid approach to personalization makes a meaningful difference. Voyado Product Discovery implicitly leverages behavioral signals from high-engagement products and shares them with similar products. A new product that shares attributes and style characteristics with something customers already love can inherit a degree of that behavioral context, entering the discovery layer with a head start rather than from zero.
What makes this work is the foundation underneath it. Voyado Elevate creates rich, retail-specific context around every product in the catalog, building a detailed picture of what each product actually is, going well beyond category labels. This enables the system to identify meaningful similarities and surface long-tail or newly introduced products in relevant recommendations much earlier than its own data history would otherwise allow.
On top of this, retailers can configure exactly how newness is treated, how much of a visibility boost new products receive, and for how long. A fast fashion retailer may want short, aggressive boosts matching their collection pace; a furniture retailer, a slower, longer ramp. There is no universal definition of “new” in retail, and the system reflects that. Once the boost period expires, ranking shifts to the signals that have accumulated, and the product is no longer cold.
The result is a discovery experience designed to work from the very first interaction, for every product in the catalog and every visitor to the site. Not just the familiar ones.
