A product can sell well and still be the wrong choice for a retailer’s overall assortment. When shelf space is limited and inventory commitments must be balanced against customer demand, selecting the most popular items does not necessarily produce the strongest product mix.
Historical sales are an essential starting point, but they describe performance within assortments that have already been offered. They cannot, on their own, establish what shoppers would choose if the available alternatives changed. Removing one item may shift purchases toward another; introducing a new option may redistribute existing demand without expanding the category.
Retail assortment optimization therefore asks a different question from product ranking: which combination of products creates the greatest value under the retailer’s objectives and constraints?
Why Historical Bestsellers Are Not Enough
Consider a retailer deciding which three products should occupy limited shelf space. Products A and B both sell well but appeal to similar needs. Product C generates fewer sales, although it serves customers who are less interested in the first two options.
Keeping only the strongest sellers could reduce coverage of that third group. Conversely, retaining two similar products might be commercially worthwhile if shoppers genuinely value the choice, or less attractive if most purchases simply shift between them. Sales rankings alone cannot distinguish these scenarios.
Product substitution is central to the problem. If a preferred item disappears, some shoppers may choose an alternative, while others may leave the category without buying. The effect of introducing an item can be equally ambiguous: it might attract additional demand, or take sales away from products already stocked.
Profitability complicates the calculation. A lower-volume product can contribute more margin per transaction than a bestseller. Retailers also need to consider shelf productivity, supplier commitments and strategic items that serve specific customer needs.
An assortment that looks attractive at a national level may also perform differently in individual stores. A neighborhood location with limited space and repeat customers may require a different balance from a larger outlet serving occasional shoppers. Treating all locations as interchangeable can hide those differences, even when the underlying product sales data is accurate.
The data requires context as well. A product with low recorded sales might have been out of stock repeatedly, while a high-selling item may have benefited from temporary promotions. Connecting transactions to availability, pricing and campaign information helps avoid treating every observed purchase as an unqualified measure of customer preference.
Understanding the Customer’s Choice Set
Customers choose among the options available at a particular moment. A product’s probability of being selected can therefore change when competing items enter or leave the assortment, even if underlying customer preferences remain relatively stable.
Customer choice models attempt to represent this process. A widely used framework is the multinomial logit model (MNL), which assigns probabilities to the selection of different alternatives based on their relative attractiveness.
In assortment planning, the focus shifts from estimating whether an item is desirable in isolation to examining how choices may be distributed across a set of products. This matters when a category manager must decide between several plausible combinations rather than produce another list of individually ranked items.
The possibility of making no purchase deserves attention too. A customer who cannot find a preferred option may switch to another brand, but may also postpone the transaction or leave the category. Models and tests that consider only purchases among the remaining products risk overlooking this lost demand. The precise treatment of a no-purchase option depends on the model and data.
Yet model assumptions matter. Standard MNL specifications impose a particular structure on substitution between alternatives and may not capture every real-world relationship. Complementary products, distinct customer segments and changing purchasing contexts can require other approaches.
A mathematical model is therefore useful only insofar as its assumptions are suitable for the decision and its predictions are validated against relevant observations.
What the ICML 2026 Research Actually Shows
In Optimal Design for Multinomial Logit Model with Applications to Best Assortment Identification, presented at ICML 2026, Joongkyu Lee and Min-Hwan Oh examine how to learn efficiently from choices made between subsets of items.
Their work considers MNL bandits, a setting in which an agent repeatedly selects a subset from a larger set of available items and observes a single-choice response. Since the number of possible subsets can grow rapidly, evaluating every combination directly is computationally difficult.
The researchers propose a computationally efficient framework for optimal experimental design, including an approach based on mixed-integer linear programming and another using a tractable surrogate. They then apply the framework to best assortment identification under a model with linear utilities and non-uniform product revenues.
The central question is not simply which assortment appears most attractive according to current estimates. It is also which observations would be most informative for identifying a better assortment.
For instance, two combinations may seem similarly valuable commercially but differ in how much they reveal about customer preferences. Selecting a more informative subset can reduce uncertainty that would otherwise affect subsequent decisions.
This is a methodological contribution with theoretical guarantees under specified assumptions. It is not evidence that a retailer achieved a particular sales or margin improvement, nor does the paper solve every operational problem associated with stores, replenishment and product availability.
Its relevance lies in showing how experimental design can support assortment identification when a retailer, or a decision-maker working in a comparable modeled setting, cannot observe every possible choice combination.
Turning Learning into Retail Decisions
Retailers generally know which products customers purchased from assortments already offered. They know much less about combinations that have never been available, particularly when demand might move between similar items.
Thoughtfully designed tests can narrow that gap. A retailer considering the removal of a product, for example, might compare alternative assortments across suitable stores or periods, provided the test design makes the groups sufficiently comparable.
Price changes, promotions, seasonality, local preferences and stockouts can all affect results. If these factors are ignored, an observed improvement may be attributed to the assortment change when another explanation is equally plausible.
The choice of outcome measure is also important. Higher unit sales do not necessarily mean higher profit, and stronger aggregate revenue could conceal weaker service to a customer segment. Before running a test, the business should specify whether it wants to improve category revenue, gross margin, inventory productivity or another objective.
Constraints then determine which potential assortments are feasible. Shelf capacity, available inventory, store format and contractual obligations can rule out combinations that look attractive in a simplified model. Differences across locations may also mean that a single assortment is not optimal everywhere.
Experimental results should inform subsequent decisions rather than be treated as permanent rules. As products, prices and customer preferences evolve, new observations can challenge earlier assumptions and provide reasons to reassess the mix.
A useful evaluation also separates what is measured from what is inferred. Revenue, units sold and availability may be directly observed, while the sales that would have occurred under a different assortment must be estimated. Making that distinction explicit helps commercial teams interpret results without overstating their certainty.
The principle is consistent with the research’s emphasis on informative observations, but practical retail testing involves additional causal, statistical and operational considerations beyond the mathematical setting studied at ICML.
From Predictive Signals to Assortment Decisions
Predictive AI can contribute estimates of future demand, purchase probabilities and differences in customer behavior across locations, segments and product categories. These signals help retailers assess expected performance beyond the transactions already recorded.
But prediction and assortment optimization are different tasks. A strong demand forecast for Product A does not establish whether Products A and B should both occupy scarce shelf space. That depends on what customers would choose in the presence of alternatives, the economics of each option and the constraints of the store.
Personalized product recommendations address another decision again: which offer may be relevant to an individual customer. Assortment planning operates at the level of the products available in a category, store or channel.
This distinction is also useful when considering newer forms of product discovery. An AI Insights analysis of AI shopping agents examines how products become visible and are evaluated during AI-assisted selection. Visibility within a discovery process, individual recommendation and assortment availability are related questions, but none is interchangeable with the others.
In practice, retailers can combine predictive estimates with choice modeling and explicit commercial constraints to compare feasible combinations. The models supply evidence; category managers still define objectives, evaluate trade-offs and decide how much uncertainty they are prepared to accept.
Choosing the Product Mix, Not Just the Products
Popularity is easy to measure, which makes it tempting to turn sales rankings into assortment decisions. Yet a retailer needs to understand whether a product brings distinctive demand, substitutes for another item or contributes value that volume figures overlook.
Choice modeling can help represent competition between alternatives, experimental design can improve learning about unfamiliar combinations, and predictive AI can contribute estimates of future demand. None of these methods removes the need to define commercial objectives or examine operational feasibility.
The strongest assortment is the one that best aligns customer choice, product economics and real constraints with the retailer’s goals, not necessarily the one containing the most bestsellers.