What Is Assortment Optimization in Retail? SKU Selection, Substitution, Constraints, and Measurement

Jun 08, 2026

Leave a message

Grace Lin
Grace Lin
Grace has spent the past seven years working directly with supermarket and convenience store buyers — mostly helping them figure out whether an ESL rollout actually makes sense for their operation, and then making it work when it does. She's covered

 

Assortment optimization is the process of deciding which products to offer from a larger set of possible products so that the resulting assortment performs best against a defined business objective.

In a retail environment, that objective might be category revenue, contribution margin, profit, space productivity, customer coverage, or another project-specific measure. The decision is usually subject to constraints such as shelf capacity, assortment size, category rules, operational complexity, and local demand.

The difficult part is that SKU performance cannot be evaluated independently.

If one product is removed, some customers may switch to another item in the assortment. Others may buy from another category, delay the purchase, visit a competitor, or make no purchase at all.

That means assortment optimization is not simply:

Retail assortment optimization using SKU selection, store demand, and merchandising data

rank every SKU by sales and remove the bottom performers.

It is a portfolio decision involving customer choice, substitution, cannibalization, constraints, and uncertainty.

A 2024 systematic review by Julia Heger and Robert Klein describes the classic assortment-optimization problem in essentially these terms: select a subset of available products to optimize an objective such as expected revenue while accounting for customer choice behavior. The review also shows that assortment optimization spans multiple choice models, constraints, and solution approaches rather than one universal algorithm. See the OR Spectrum systematic review of assortment optimization.

 

Assortment Planning and Assortment Optimization Are Closely Related

Assortment planning and assortment optimization should not be divided into a simplistic "static versus dynamic" distinction.

Assortment planning is usually the broader business process used to decide what product range should be offered by category, format, store group, season, or channel.

Assortment optimization is the analytical decision problem inside that process: given a set of candidate products, demand estimates, objectives, and constraints, which combination should be offered?

Terminology can overlap in both academic and industry use, so the more useful distinction is between the business decision being made and the analytical method used to support it.

Decision Main Question
Assortment optimization Which SKUs should be offered?
Shelf-space planning How much space or how many facings should each item receive?
Replenishment How much inventory should be reordered, and when?
Pricing optimization What price, promotion, or markdown should be used?
Planogram execution Was the approved assortment and shelf layout actually implemented?

These decisions interact, but they are not interchangeable. For example, dynamic pricing can change demand for products that are already in the assortment, but pricing does not answer the separate question of whether those SKUs should be carried in the first place.

A retailer can choose the right assortment and still lose sales because replenishment fails. It can also maintain excellent availability while carrying the wrong assortment.

 

Why SKU Sales Rankings Are Not Enough

A basic SKU productivity report is useful, but it is not an assortment-optimization model.

The missing question is:

What happens to customer demand when the assortment changes?

Substitution

Suppose a category contains five similar products and one is removed.

Not all of its historical sales necessarily disappear. Some shoppers may choose the closest remaining alternative.

If most demand transfers to another SKU already in the assortment, removing the first item may reduce complexity without losing the same amount of category demand. If shoppers consider the removed item unique, substantially more demand may be lost.

The value of an item therefore depends partly on the alternatives offered beside it. This consumer-choice principle is related to broader merchandising effects discussed in retail topics such as the framing effect in product presentation, although assortment optimization focuses specifically on which alternatives are available rather than only how they are presented.

Retail assortment research explicitly models substitution because purchase probabilities can change with the available offer set. Fisher and Vaidyanathan, for example, model substitution when a shopper's first choice is absent while choosing store assortments from a larger candidate SKU universe. See the Management Science assortment-optimization study.

Cannibalization

The opposite issue also matters.

Adding a new SKU may generate strong unit sales without creating the same amount of incremental category demand. Some of those sales may simply transfer from products the retailer already carries.

A stronger assortment question is therefore:

What incremental value does this SKU add to the portfolio?

not merely:

How much could this SKU sell?

 

Retail product substitution, cannibalization, and no-purchase behavior after a SKU is removed

No-Purchase Behavior

Customers do not always substitute when a preferred product is unavailable.

Some may:

leave the category without purchasing;

delay the purchase;

buy from another retailer;

move to a different category.

A useful assortment model therefore needs to recognize that changing the assortment can affect both product shares and total category conversion.

 

Out-of-Stock and Out-of-Assortment Are Different

Historical sales can be misleading if product availability is not understood.

An out-of-assortment SKU is intentionally not offered.

An out-of-stock SKU belongs in the approved assortment but is temporarily unavailable.

Those conditions have very different implications for demand estimation.

Consider a product with weak recorded sales. Possible explanations include:

  • customers do not want it;
  • it was frequently out of stock;
  • it was introduced only recently;
  • its shelf position was poorly executed;
  • its price or promotion changed;
  • nearby substitutes captured demand.

If persistent stockouts are interpreted as weak demand, an assortment model can recommend removing a product partly because the retailer failed to make it available.

Before estimating demand, separate true low demand from periods in which customers could not reasonably buy the product.

 

Define the Objective Before Optimizing

"Optimize the assortment" is incomplete unless the organization first agrees on what is being optimized.

Possible objectives include:

  • revenue;
  • gross or contribution margin;
  • category profit;
  • sales or margin per unit of space;
  • customer or attribute coverage.

The correct objective depends on category strategy.

A convenience retailer with severe space constraints may place more weight on contribution per unit of space. A specialty category may retain lower-volume products because they provide important brand, specification, pack-size, dietary, or price-point coverage.

This is why individual SKU velocity alone is insufficient.

A simplified conceptual formulation is:

Choose assortment A from the candidate product universe to maximize the expected business value of A, subject to the relevant constraints.

The important detail is that expected demand for each product can change when A changes.

 

Define the Constraints

A mathematically attractive assortment can still be operationally impossible.

Constraints should therefore be defined before candidate assortments are compared.

Shelf and Fixture Capacity

Physical retail cannot carry an unlimited range.

Useful inputs may include:

  • available shelf width;
  • fixture capacity;
  • product dimensions;
  • minimum facings;
  • storage and replenishment capacity.

Assortment selection and detailed shelf-space allocation are separate decisions, but physical capacity still determines what the assortment can realistically contain.

Merchandising tools such as digital endcap displays can influence how selected products are promoted or presented, but they do not remove the underlying assortment and fixture constraints.

Assortment-Size Limits

Some categories have a practical limit on SKU count because additional items can increase:

  • inventory complexity;
  • replenishment workload;
  • supplier management;
  • planogram complexity;
  • working capital.

The appropriate limit is category- and format-specific. There is no universal ideal assortment size.

Category and Merchandising Rules

A retailer may need to preserve:

  • opening, mid, and premium price points;
  • key brands;
  • private-label representation;
  • pack-size coverage;
  • dietary or technical attributes;
  • strategic supplier commitments.

These are legitimate business constraints. Their presence does not automatically make an optimization model less sophisticated.

Operational Constraints

A theoretical model that generates a unique assortment for every store may create more operational complexity than the expected benefit justifies.

Optimization can therefore include a constraint on the number of distinct assortment templates used across the network.

 

Store-Level Assortments vs. Store Clusters

Local demand matters, but a unique assortment for every store is not automatically the best answer.

The trade-off is:

local relevance versus operational complexity and data reliability.

Store-level assortment decisions may be appropriate when:

  • store demand differs meaningfully;
  • sufficient item/store data exist;
  • execution systems can manage the variation;
  • the expected benefit justifies the added complexity.

Clusters can be more practical when groups of stores have similar purchasing patterns or operating constraints.

Useful clustering variables can include:

  • category purchasing patterns;
  • basket composition;
  • store format;
  • available space;
  • local product preferences;
  • shopper mission;
  • channel mix.

There is no universal rule such as "every retailer should use four to eight clusters." Use as many distinct assortment groups as the data quality and operating model can justify.

 

What Data Does Assortment Optimization Need?

The required data depends on the category, decision scope, and modelling approach.

Transaction Data

  • SKU;
  • store or channel;
  • units;
  • revenue;
  • price;
  • promotion;
  • date or time where relevant.

Availability and Inventory Data

Availability data helps distinguish weak demand from unavailable product.

Useful signals can include:

  • in-stock periods;
  • inventory position;
  • stockout events;
  • replenishment history.

Product Attributes

Attributes can help identify alternatives, substitution relationships, and gaps in the range.

Examples include:

  • brand;
  • size;
  • flavor;
  • material;
  • pack type;
  • price tier;
  • technical specification.

Store and Channel Information

  • store format;
  • available space;
  • geography;
  • local demand pattern;
  • channel;
  • operational constraints.

Customer or Basket Data

Where available and appropriate, loyalty or basket-level data can improve understanding of customer preferences, cross-product substitution, and segment-specific demand.

A retailer does not universally need "at least 12 months" of sales history. The useful period depends on seasonality, lifecycle, promotional behavior, new-product frequency, data volume, and the modelling method.

More history is not automatically better if the market or candidate SKU universe has materially changed.

 

When Is Formal Assortment Optimization Worth It?

Not every assortment decision requires a complex choice model.

Simple SKU productivity rules combined with experienced category review may be adequate when:

  • the candidate SKU universe is small;
  • products have limited substitution;
  • the assortment changes infrequently;
  • constraints are simple;
  • the cost of a wrong decision is relatively low.

Formal choice modelling and optimization become more valuable when:

  • the candidate SKU universe is large;
  • many products are close substitutes;
  • cannibalization is material;
  • stores have meaningfully different demand;
  • multiple commercial and physical constraints interact;
  • the number of possible assortment combinations is too large for manual comparison;
  • the business needs to quantify trade-offs rather than rely only on rankings.

The objective is not to use the most complicated model available. It is to use enough analytical sophistication to represent the decision accurately.

 

A Practical Assortment Optimization Workflow

1. Define the Decision

Specify:

  • category;
  • stores or channels;
  • candidate SKU universe;
  • decision period;
  • business objective.

Define the decision before selecting software.

2. Define the Constraints

Document:

  • space;
  • assortment size;
  • required products or attributes;
  • commercial rules;
  • operational limitations;
  • store or template constraints.

3. Clean the Demand Data

Identify:

  • stockout periods;
  • product launches;
  • discontinuations;
  • unusual promotions;
  • major price changes;
  • data-quality problems.

Historical sales should not automatically be treated as unconstrained customer demand.

4. Estimate Choice and Substitution

Determine which products shoppers view as alternatives and what happens when a preferred item is absent.

Methods can range from relatively simple product-attribute and substitution analyses to discrete-choice, rank-based, machine-learning, or other formal demand models.

The appropriate complexity depends on the decision and the quality of the available data.

5. Generate and Compare Candidate Assortments

Evaluate product combinations against the chosen objective and constraints.

Do not evaluate each SKU independently.

A product that appears weak by itself may serve demand that other products do not cover. Another SKU may show strong sales but add limited incremental value because shoppers would otherwise choose close substitutes.

6. Review Business Exceptions

Category teams should review recommendations that involve:

  • strategic new products with limited history;
  • contractual requirements;
  • known availability distortion;
  • seasonal ranges;
  • products serving a small but strategically important segment.

Human review should challenge the model with evidence rather than override recommendations solely because they are unfamiliar.

7. Pilot the Change

Where the decision risk justifies it, test candidate assortments in a controlled group of stores or channels.

Depending on the retailer's analytical capability, comparison methods may include:

  • prior-period baselines;
  • matched stores;
  • holdout or control stores.

8. Verify Execution Before Judging the Model

If the approved assortment was never implemented correctly, the test does not provide a clean measurement of the assortment decision.

Check:

  • item authorization;
  • actual availability;
  • shelf or planogram implementation where relevant;
  • replenishment;
  • pricing and promotion consistency.

This is where execution technologies become relevant. For example, retailers using electronic shelf labels in grocery stores may be able to change shelf-edge information without reprinting paper labels, but the technology does not determine which SKU assortment is analytically optimal.

Price execution also matters when evaluating a pilot. If displayed prices do not match the intended commercial conditions, the observed demand may be difficult to interpret. The same principle is discussed in the guide to what happens when retail price displays are wrong.

Retail assortment optimization workflow from SKU data and constraints to pilot testing and measurement

 

How to Measure Whether the Assortment Improved

Use metrics that correspond to the objective defined before optimization.

Primary Business Outcomes

Depending on the project:

  • category revenue;
  • contribution margin;
  • category profit;
  • sales or margin per unit of space;
  • incremental performance versus a suitable baseline or control.

Choice and Demand Outcomes

Where they can be measured credibly:

  • substitution capture;
  • lost-demand or no-purchase behavior;
  • customer or attribute coverage.

Operational Diagnostics

These help determine whether the test was executed fairly:

  • in-stock rate;
  • stockout frequency;
  • assortment or planogram compliance;
  • inventory turns;
  • markdown or shrink where relevant.

A higher planogram-compliance rate does not by itself prove that the assortment is better. It means the retailer has a cleaner test of the assortment it intended to offer.

 

Where AI Helps - and Where It Does Not

Assortment optimization does not require AI by definition.

Operations-research methods, discrete-choice models, heuristics, mathematical optimization, and statistical demand models have supported assortment decisions for years.

Machine learning can be useful for:

  • demand estimation;
  • product similarity;
  • substitution modelling;
  • store segmentation;
  • high-dimensional attribute analysis;
  • scenario prediction.

AI does not remove the need to define objectives, constraints, availability, decision rights, or validation.

A sophisticated model trained on distorted availability data can still produce a sophisticated wrong answer.

 

Where Electronic Shelf Labels and Shelf Technology Fit

Electronic shelf labels and other store technologies belong primarily to the execution and measurement layer around assortment decisions.

An electronic shelf label solution can help retailers manage shelf-edge information after products, prices, or promotions change. Retail-specific hardware such as retail electronic shelf labels can therefore reduce some of the manual work associated with assortment resets.

For readers evaluating the technology itself, the guide to how electronic shelf labels work explains the store-side system separately from assortment analytics.

An ESL should not be described as an assortment-optimization engine.

Likewise, automatic shelf-state or planogram detection requires the particular sensing, vision, location, inventory, or platform capabilities of the deployed system. It should not be assumed merely because electronic price labels are installed.

In supermarket applications, guides such as supermarket electronic price tags are therefore best viewed as execution-technology resources rather than substitutes for the assortment decision model.

A useful retail architecture keeps two layers distinct while allowing data to move between them:

assortment decision → shelf execution and measurement

 

Common Assortment Optimization Failure Modes

Failure Mode Why It Produces a Weak Decision
Ranking SKUs independently Ignores substitution and cannibalization across the portfolio.
Treating stockouts as low demand Can recommend removing products because of availability failure rather than customer preference.
Ignoring constraints Produces an assortment that may not fit the fixture, category strategy, or operating model.
Over-localizing Creates store-level complexity that may exceed the incremental value of localization.
Using one objective and judging another A model optimized for revenue cannot automatically be judged as though it was designed to maximize inventory turns.
Failing to verify execution Makes it difficult to distinguish a poor assortment from poor implementation.
 

 

Assortment Optimization Checklist

  • Define the business objective before selecting the model.
  • Define the candidate SKU universe.
  • Separate assortment selection from shelf-space, replenishment, and pricing decisions.
  • Identify periods when products were unavailable before interpreting weak sales.
  • Account for substitution, cannibalization, and no-purchase behavior where material.
  • Document physical, merchandising, and operational constraints.
  • Use store-level localization only when the data and expected benefit justify the complexity.
  • Do not use a universal rule for store-cluster count or historical-data length.
  • Compare candidate assortments as portfolios rather than isolated SKUs.
  • Use a formal optimization model when substitution, scale, heterogeneity, and interacting constraints make manual comparison unreliable.
  • Pilot material assortment changes when the decision risk justifies controlled testing.
  • Verify availability and execution before attributing results to the assortment.
  • Measure performance against the objective the optimization was designed to improve.

Assortment optimization is therefore not primarily about having more data, using more AI, or changing shelves more frequently.

It is about making a disciplined portfolio decision:

 

Which combination of products creates the strongest expected result under the real customer-choice and operating constraints of the retailer?

For retailers evaluating the store technology needed to execute assortment, pricing, and shelf-edge changes, the broader electronic shelf label product range provides one possible execution layer. Technology selection should follow the merchandising and operating requirements rather than define them.

For a project-specific retail display or electronic shelf label requirement, provide the store format, number of locations, product categories, integration requirements, label sizes, update workflow, and deployment environment when you request a quotation.

Send Inquiry