The choice between session-based and long-horizon recommendation models is not a theoretical one for DTC operators - it is a practical question about which approach will produce better recommendations given the actual data you have available and the actual behavior patterns of your shoppers. The right answer depends on who your customers are and what your catalog looks like, but there are structural reasons why the answer for most mid-market DTC brands leans toward session-based approaches, at least as the primary signal layer.
What Long-Horizon Models Need to Work
Long-horizon collaborative filtering and matrix factorization approaches are built on the premise that you have stable, dense interaction histories for a meaningful portion of your users. "Long-horizon" in this context means using a user's full transaction and browsing history - often 12 to 36 months of data - to build a profile of their preferences and then recommend items that fit that profile.
These models work well when three conditions hold. First, users transact repeatedly with high frequency, giving you enough interaction events per user to triangulate their preference structure. Second, the catalog is relatively stable over time, so that historical interactions with specific items remain predictive of future preferences. Third, a significant fraction of your user base has enough history to generate meaningful profiles - not just 5% or 10% of your most loyal buyers.
Subscription services, marketplaces with broad product ranges, and retailers with strong loyalty program participation tend to meet these conditions. A DTC brand with a typical mid-market purchase frequency of one to three orders per year per customer, a catalog that turns over substantially each season, and a high fraction of new-to-brand traffic does not.
The DTC Data Reality
Mid-market DTC brands typically see the following in their session data: 60% to 75% of sessions are from visitors with no purchase history at all. Of the remaining sessions from known customers, a substantial fraction are from people who have purchased only once - enough to identify them but not enough to characterize their preferences reliably. The cohort of customers with three or more purchases and enough data to build a meaningful long-horizon profile is often 15% to 25% of sessions.
So a long-horizon model that performs well on the cohort with rich history will be cold-starting for 75% to 85% of sessions. "Cold-starting" in this context does not mean the system fails; it means it falls back to some default behavior - usually popularity ranking, category-based recommendations, or a generic collaborative filter trained on aggregate patterns rather than individual history. Those fallbacks are not terrible, but they are not personalized.
The question is whether the 15% to 25% of sessions where long-horizon modeling works well justifies the infrastructure cost, or whether a different architecture better serves the full distribution of sessions.
What Session-Based Models Actually Do
Session-based recommendation models treat each visit as the primary evidence source. They look at the sequence of interactions within the current session - which products were viewed, in what order, how long the shopper paused, which categories were explored, which price tiers were engaged with - and use this to estimate what the shopper is looking for right now.
The "right now" framing is important. Intent in e-commerce is often session-specific. A shopper who bought running shoes from you six months ago may be visiting today specifically to look for a gift for someone else. Their historical purchase data says "running gear"; their current session says "home goods, gift-appropriate, under $75." A long-horizon model that leans on purchase history will pull them toward running gear. A session model that reads the current signals will pull them toward the gift they are actually shopping for.
Session-based models in their modern form use sequential neural architectures - often transformer-based models adapted from natural language processing - that learn patterns in interaction sequences. The model learns that shoppers who look at items A, B, and C in sequence tend to purchase items in category D, or tend to respond to items at price tier E. These patterns are learned from aggregate session data across all users, which means new visitors benefit from what the model learned from everyone who browsed similarly.
Why Session Models Punch Above Their Weight for Mid-Market DTC
The key advantage is coverage. A session-based model produces useful recommendations for every session that has any interaction history, even a single page view. A new visitor who has browsed three products in the linen bedding category has enough signal for the model to make informed next-item predictions - not as precise as what a long-horizon model would produce for a loyal repeat buyer, but far better than a generic bestseller list. The model has seen thousands of sessions that started with similar linen bedding browse patterns and knows where they typically go.
For catalogs with significant item churn - seasonal drops, limited runs, new SKUs coming in regularly - session models also handle novelty better. A new product has no historical interaction data that a long-horizon model can learn from, but it has item attributes. A session model trained with attribute-aware inputs can recommend new products to shoppers whose session behavior indicates alignment with those attributes, even on the day the product launches.
Where Long-Horizon Models Still Add Value
The case for not abandoning long-horizon signals entirely is that repeat buyer behavior contains real, durable preference information that session signals alone cannot capture. A shopper who has purchased from your brand four times over two years has established a clear pattern that is worth incorporating. Their session today may look atypical - they are browsing categories they have not bought from before - but their purchase history suggests they have a certain quality threshold, a certain aesthetic orientation, and likely fall within a certain demographic that responds to specific product attributes.
The practical resolution for most mid-market DTC brands is a hybrid: session-based modeling as the primary signal layer, with long-horizon features incorporated as supplementary context when they are available. For sessions with rich purchase history, the model weights the historical signals more. For new visitors, the model relies primarily on the current session. The relative weighting can itself be learned from data - models trained on this kind of hybrid input will naturally learn to rely on historical signals when they are predictive and discount them when they are not.
The Infrastructure Question
Session-based models require a real-time signal pipeline that captures and processes session events with low latency - typically sub-second ingestion so that the model's candidate set can be refreshed as the shopper continues browsing. Long-horizon models are typically computed offline on a batch schedule, which is simpler to operate but means the recommendations may be stale by 12 to 24 hours.
For most DTC operators, the real-time session pipeline is a higher-value infrastructure investment than the offline batch computation, not because batch is bad but because 60-75% of your sessions are from people who have no offline profile to serve. The infrastructure that helps new visitors is often the more important one to build first.