AI-driven assortment optimization is the use of AI and machine learning to decide which products each store and channel should carry, at what depth, and in what breadth. It reads demand signals, product attributes, and store-level data instead of last season's range. It then recommends the product selection that best meets category goals within real constraints. Most retailers believe their range is already localized. Their sell-through numbers say otherwise, season after season. So what does an AI assortment plan actually change, and where does it pay? Answer that now and you protect margin before the buy. Answer it late, and you pay for it all season.
How AI-Driven Assortment Optimization Works
The process runs as a loop, not a once-a-season pass. The model learns what sold where, solves for the best mix inside real limits, and every cycle starts from what the last one taught it. Four steps repeat.
- Automated hindsight: Historical sales data, product attributes, promotion and price activity, gross margin performance, lost sales, and current inventory positions land in one layer. Automated hindsight analyzes how products performed across attributes and store clusters or customer segments, at store level rather than on chain averages, and plan-versus-actual comparisons run automatically.
- Intelligent store clustering: ML-driven clustering groups stores on the metrics that actually drive assortment—lost sales, revenue, sell-through, margin, unit turns—combined with key product attributes, store attributes, and trade area analytics. Clusters can be performance-based, attribute-based, or thematic and department-level, and they update dynamically. Localization starts here, because two stores of similar size in the same state routinely belong in different clusters. The same logic sizes the opening range for a new store, matching it to sister stores on demographic similarity, competitive intensity, and store attributes.
- Two-stage optimization: Stage one distributes budget at cluster and price band level. Stage two sets breadth and depth inside that budget, considering sell-through, sales, lost sales, and demand elasticity rather than SKU productivity alone. Depth and choice count resolve at cluster–subgroup level, the assortment wedge at cluster–style level, inside real constraints: min-max presentation minimums, choice constraints by store, MOQ, and pack size. Planners can override any algorithmic suggestion.
- Keep, remove, add: Optimized receipts are seeded by store, and the plan is validated at store level before it locks. Demand transference models the switching matrix between similar SKUs, estimating what demand moves to a substitute and what exits entirely if an item is delisted, so every removal carries a value attached. Keep/remove/add decisions resolve at department, category, brand, and SKU level, and forecast-driven carryover projections map carryover choices onto next season's placeholders.
How AI Technology Handles Products with No Sales History
New items break most models. A single engine covering long lifecycle, short lifecycle, and new product forecasting closes the gap, running three methods and picking the strongest per category. A hierarchical forecast borrows seasonality from a higher level of the product hierarchy and estimates the item's probable contribution to it. A similarity-based forecast matches the new item to comparable sellers on product description, price band, category, and thematic similarity. A sell-through heuristic projects forward from typical new-product sell-through rates. Automatic placeholder creation with ML-driven attribute recommendations speeds up product setup, and computer vision generates initial placeholder images before samples exist.
Every step produces data-driven insights instead of buyer memory, and turns range calls into data-driven decisions. Judgment still sets the direction. The model carries the scale across every store, every channel, and all product categories.
Why AI-Driven Assortment Optimization in Retail Matters
Assortment is the first bet a retailer places. It decides what to buy before a single unit sells. Pricing, allocation, and replenishment can only work with what that decision already committed to. Put the wrong product mix in a store, and no downstream move recovers the margin. Getting the right products into each store is a retail strategy question, not a reporting exercise.
Adoption is not the problem. Proof is. In a McKinsey survey of merchants published in January 2026, 71% said AI merchandising tools have had limited to no effect on their business so far. The gap sits in the foundations: sales history quality, attribute consistency, and financial reconciliation. The algorithms are rarely the weak link.
The business impact of a poor assortment
Weak product assortment choices cost money in four places at once.
- Excess inventory in stores that never wanted the item, then cleared through markdowns.
- Stockouts and lost sales in the stores that did want it.
- Non-productive SKUs, because duplicative and underperforming items proliferate and take capacity from items that would have sold.
- Falling customer satisfaction, since shoppers judge a store by what it carries.
Those four costs move together, and so do the gains. Inventory efficiency improves when the range matches local demand, because the same working capital sits where it sells. Gross margin follows, and so does assortment performance against plan. Retailers who get this right close the gap between planned and actual sell-through, cut markdowns, and clear stockouts across stores — the same range decision paying off in three places at once.
For a CPG supplier, the same logic runs through category management, where the retail partner wins only when the shelf reflects what that store's shopper actually buys. Ranges built around customer needs at each location lift sell-through, and they make retail operations simpler downstream.
AI-Driven Assortment Optimization vs Traditional Assortment Decisions
Legacy assortment planning starts with last year. Planners copy the prior range, adjust for known wins and losses, then apply a growth number. It is fast, familiar, and blind to anything that changed. AI-powered assortment strategies replace the copy with a model that learns.
The two approaches diverge on four dimensions.
One honest note belongs here. Neither approach wins alone. Most enterprise deployments run a hybrid model. The system proposes the optimized range, merchants approve or override any algorithmic suggestion, and automated hindsight with configurable guardrails carries the lessons into the next cycle. Buyers keep authority over brand direction and strategic assortment calls. AI changes decision-making processes rather than removing people from them, and that balance is deliberate. A range carries risk no model should carry alone.
Where to Go Deeper on Assortment Planning and Optimization
This primer covers the definition. Four deeper guides cover the work. Each one answers a different question, so start with the question you already have.
- You want the financial envelope the range has to fit. Our guide on how enterprise retailers use AI for merchandise financial planning explains how sales, margin, and open-to-buy targets get set.
- You want new-product handling inside an optimized range. Our guide to AI forecasting for new product launches covers the cold-start problem and the attribute-based approach behind it.
One sequence is worth holding on to. Merchandise financial planning sets the envelope. Assortment optimization decides what fills it. Teams that optimize product mix inside that envelope protect margin. Allocation and replenishment then execute against the decision, and store operations live with the result. Retailers who work in that order keep the range and the budget in agreement all season. Retailers who work out of order reconcile twice and lose the buying window. That reconciliation is cheapest when merchandise financial planning, assortment planning, and allocation and replenishment run on one platform rather than across three integrations.
Where This Leaves Merchandising Leaders
The advantage in the retail industry is moving from who owns AI tools to who trusts the output enough to act on it. Better assortments come from better inputs: clean sales history, a consistent attribute taxonomy, and a range that reconciles to the budget while it is still being built. Retailers with clean sales history, a consistent attribute taxonomy, and a range that reconciles to budget will know what each store should carry.
