Cold start is the problem that makes personalization harder than it looks in vendor pitch decks. The pitch goes: feed us your data, we build user profiles, we serve personalized recommendations. What the pitch skips is: what do you serve when there is no profile yet? When a shopper has never been to your site, when a product was just added to your catalog, when you have entered a new category with no prior behavioral data - the system has nothing to work with, and most personalization approaches degrade gracefully at best, and fall back to unacceptable defaults at worst.
Cold start is not one problem. It is three distinct problems that require different solutions, and understanding which version you are dealing with is the prerequisite for addressing it effectively.
New User Cold Start
The most common variant is the new visitor problem. A shopper arrives on your site for the first time. They have no purchase history with your brand, no browsing history in your system, no profile to draw on. What do you show them?
The naive fallback is popularity ranking - show the bestsellers. This is not wrong. Bestsellers became bestsellers partly because they appeal broadly. But it is also a missed opportunity, because even a new visitor is not a blank slate. They arrived from somewhere. The acquisition channel and landing page carry signals about why they came. A shopper who clicked a Pinterest image of a minimalist kitchen setup is different from a shopper who clicked a Facebook ad about a specific product on sale. The entry path is the first personalization signal you have, even before they interact with a single page.
Beyond the entry path, the first few interactions accumulate quickly. After a new visitor views two or three products, their session has established a preference signal. The category they went to first, the price tier of the items they clicked, whether they used the search bar or browsed - all of these are actionable within the current session, even without any prior history.
Session-based approaches handle new user cold start better than history-based approaches for exactly this reason: they do not require prior history to make predictions. They learn from what the current session is doing and compare it to patterns from similar sessions in the training data. A new visitor who has browsed three outdoor furniture items in a certain style register will get outdoor furniture recommendations calibrated to that style, even on their first visit, because the model has seen thousands of sessions that looked like that one and knows where they tend to go.
New Product Cold Start
New product cold start is often overlooked because it is less visible than new visitor cold start, but it is equally important for businesses with active product development pipelines. A new SKU has no purchase history, no view data, no click data. Collaborative filtering cannot make meaningful predictions about it because there are no "users who bought this also bought X" patterns to draw on. Without explicit merchandising intervention, the product is invisible to any recommendation surface.
The most reliable bridge for new product cold start is attribute-based similarity. A product has attributes even on day one: its category, its material, its price tier, its color family, its use case, its visual style characteristics. These attributes can be compared to the attributes of products the current shopper has already engaged with during their session or in their history. If a shopper has been consistently engaging with mid-century modern furniture in natural wood tones across a price range of $300 to $600, a new product with those attribute signatures is a reasonable recommendation even without behavioral data.
The practical implementation involves embedding both session signals and item attributes in a shared latent space, so that a query from a session can retrieve relevant items by attribute proximity even when those items have no interaction history. This is computationally tractable and produces recommendations that are noticeably better than the generic fallback.
Editorial curation can supplement attribute similarity for truly new items. Tagging new SKUs as "new arrivals" and surfacing them with explicit fresh-arrival signals lets shoppers self-select into them when they are browsing for novelty. The behavioral signal from those early interactions then feeds back into the model, gradually building the interaction data needed for recommendation-based surfacing.
New Category Cold Start
When a brand enters a new product category - typically through a line extension or a new seasonal range - the cold start problem compounds. Not only are the products new, but the behavioral patterns within that category are new too. The model has no historical data to learn from about how shoppers browse or purchase within this category, which shopper segments tend to respond to it, or how price sensitivity differs from your existing categories.
The most effective approach for new category cold start is transfer learning from adjacent categories. If you are launching a bed linen range and your existing business is home goods, the shopper profiles from the home goods categories carry genuine predictive signal about who will respond to bed linen and what product attributes they will prefer. The model can be warm-started with this transferred knowledge rather than beginning from scratch.
This works because shopper preference has transferable dimensions. Aesthetic orientation, price tier sensitivity, and purchase occasion type are characteristics that tend to be consistent across categories for a given shopper. Someone who consistently buys premium, minimalist home accessories is a reasonable candidate for premium, minimalist bed linen, even if they have never bought from that category before. That inference is available from cross-category transfer before you have a single bed linen transaction to learn from.
Practical Prioritization
Not all cold start situations require the same urgency of treatment. New user cold start affects every single session from new visitors - for most DTC brands, that is 60% to 75% of total traffic. It is the highest-volume cold start problem and deserves the most attention. Degrading to pure popularity ranking for that fraction of your sessions means you are leaving personalization-driven lift on the table for the majority of your visits.
New product cold start affects every item in your catalog during its first few weeks of availability. If you are launching seasonal collections, this affects a significant fraction of your active catalog simultaneously, particularly at the start of each season. Attribute-based bridging is essential during this window.
New category cold start is a lower-frequency event - it happens when you are deliberately expanding the business, not continuously. It warrants careful transfer learning setup when it does happen, but it is not a constant operational concern in the way that the first two variants are.
The connecting principle across all three is that cold start is a data sparsity problem, and the solution is always to find the most proximate available signal. Session context, acquisition channel, item attributes, and cross-category transfer are all ways of finding signal when the specific signal you would prefer - long-horizon user history, item interaction data - is not there yet. The system should be designed to use all of them, degrade to fewer signals gracefully when some are absent, and improve automatically as data accumulates without requiring manual reconfiguration.