Cold Start in E-commerce: How to Recommend Products Without Sales History?

A new shop just launched. The catalog is live, the first visitors are arriving… but the "Recommended products" block stays empty. This isn't necessarily a configuration issue: it's often what's known as cold start — the period during which an engine doesn't yet have enough behavioral data.

What is cold start?

Cold start refers to the difficulty of producing recommendations when clicks, carts, purchases, or ratings are too few to reveal reliable patterns.

It takes several forms:

• product cold start: a new item hasn't received any clicks or purchases yet;
• user cold start: a visitor is new or unidentified;
• shop cold start: the catalog exists, but the interaction history is too limited.

The problem isn't a total absence of information. The catalog already contains names, categories, descriptions, attributes, and sometimes images.

Why do classic engines wait?

Collaborative filtering learns primarily from observed relationships:

users → interactions → habits → recommendations

This approach is very effective when there's enough data volume. It can discover that one product is often bought alongside another, or that a group of visitors likes the same items.

But a new shop doesn't have these signals yet. A product launched yesterday has no co-purchase data to analyze. Showing only best-sellers or applying manual rules can be a stopgap, but it doesn't allow for a fine-grained understanding of the relationships between products.

Collaborative filtering isn't flawed, then — it simply answers a different question. It learns from actual behavior, while cold start requires reasoning from the catalog first.

Recommending from the catalog

Semantic recommendation flips the starting point:

products → meaning and attributes → relationships → recommendations

Each product is turned into a numerical representation called an embedding. This representation encodes part of the meaning of the name, category, and description. Two products with similar semantic content can end up in neighboring zones of the vector space, even if no customer has ever compared them.

For example, a listing for a "200m waterproof dive watch" can be matched with other dive watches with comparable characteristics from day one. No clicks are needed to establish this initial relationship. The merchant therefore doesn't need to wait several weeks or several hundred orders before new products can enter the recommendations.

This approach doesn't replace catalog quality, though. Missing information — a diameter, a voltage, a capacity — can't be reliably guessed. Known technical data must still be present on the product listings.

How RecoKit works from day zero

Understanding each product

During ingestion, RecoKit uses the name, category, description, and available attributes. Listings can be cleaned and enriched; when an image is available and the text is sparse, a visual description can complete the product signal.

Distinguishing similarity from complementarity

Two similar products generally meet the same need. They're useful for comparing or replacing an item.

Two complementary products play different roles. A drill might call for drill bits, safety glasses, or a measuring tool. A printer might call for cartridges — the reverse isn't necessarily true.

RecoKit distinguishes between these relationships. At startup, RecoKit relies on relationships already present in the catalog and on product proximity to build an initial recommendation base. These relationships are then refined as more specific signals become available.

Connecting purchase intents

A useful recommendation doesn't just link objects — it also links intents: installing, repairing, protecting, measuring, maintaining, or completing a project.

RecoKit uses a pre-built intent taxonomy covering several verticals. In home improvement, the intent "drill and fasten" can be linked to "protect yourself" or "measure." These relationships can be used even before any order history has been observed.

Verifying before publishing

Semantic proximity isn't always enough. An engine might find two products that are close in language but incompatible in reality.

Candidates are therefore subject to relevance checks and, when necessary, compatibility checks. For a pump and a filter, flow rate, capacity, or fittings can be decisive. The goal is to avoid turning a textual resemblance into an automatic recommendation without verification.

When does cold start affect you?

This matters especially if you're launching a new shop, regularly adding new products, releasing a seasonal collection, selling niche products, or have few transactions.

In these situations, waiting for each product to accumulate enough clicks can delay the engine's value right when recommendations would be most useful.

A hybrid approach over time

Catalog-based recommendation isn't meant to ignore behavior. It provides an initial foundation, which clicks, carts, and purchases can then help refine.

catalog from launch + behavioral signals progressively available
                              ↓
                    hybrid recommendations

The point is not having to choose between a semantic approach and a behavioral one. The first shortens the startup delay; the second progressively improves personalization as data becomes available.

Key takeaways

Cold start doesn't mean a catalog is empty of information. It means the behavioral history isn't sufficient yet.

An effective strategy is to immediately leverage product data, distinguish similar products from complementary ones, verify sensitive relationships — particularly technical compatibility — and then incorporate behavioral signals once they become reliable.

RecoKit is built around this catalog-first logic: produce an initial recommendation base without requiring a history of clicks or purchases, then enrich it as the shop's life unfolds.

FAQ

Can you recommend products without customer data?

Yes, provided the product information is descriptive enough. A semantic approach can leverage names, categories, descriptions, attributes, and images before the first clicks arrive.

Does cold start only affect new shops?

No. It also affects every new product, every new collection, and every visitor without usable history — as well as shop migrations or platform changes.

Does semantic recommendation replace collaborative filtering?

No. It addresses the startup problem. Behavioral signals remain valuable and can supplement the catalog base as interaction volume grows.

Can an image compensate for a poor product listing?

It can enrich the available signal, but it doesn't replace precise technical data. Dimensions, capacities, voltages, and compatibilities must be provided when they're necessary.

How many orders are needed before using a recommendation engine?

There's no universal threshold — it depends on the number of products, traffic, and the distribution of interactions. A catalog-first approach precisely avoids tying the startup to reaching an arbitrary volume.