Store by store: what AI-tailored local assortments really deliver

When Zara's operations team and academic researchers tested a new way of deciding how much of each item to send to each store, they ran it as a controlled field experiment. The result, published in the journal Interfaces in 2010, was a sales increase of 3 to 4%, which the authors estimated was worth about USD 353 million in additional revenue for 2008. The model used no generative AI. It linked each store's stock in every size to its demand and chose shipments from two central warehouses to about 1,500 stores.

Fifteen years later, retailers and software vendors promise far more: ranges tailored to each neighbourhood, size curves tuned store by store, local trends caught before they spread. The evidence is thinner than the promise. The cases with reported numbers, from Guess to Hackett, concern allocation and replenishment: how many units of an existing range go to which store, and when. Genuinely different assortments per store, chosen by algorithm, remain rare and largely unmeasured in public.

That matters now because the economics favour precision. Retailers are holding back more stock for in-season decisions, cutting markdowns and running fewer, larger stores. Each unit sent to the wrong door is a markdown or a lost sale. The question for merchandisers is not whether to localise, but which decisions AI can reliably take at store level, and what data that requires.

What does localised assortment mean in fashion retail?

Localisation covers four different decisions that are often blurred together. Assortment decides which styles a store carries at all. Allocation decides the opening quantities of each style, colour and size per store. Replenishment tops up stores in season from central stock. Size and depth optimisation adjusts the size curve and the number of units per option to local demand. A store that receives the same styles as every other branch but different quantities and size splits is localised in the allocation sense, not the assortment sense.

Fashion makes all four harder than in grocery. Most styles sell for a single season, so there is little history per item and store. Demand is spread across many sizes and colours, so sales per SKU and store are small and noisy. And when an item sells out, the system records zero sales, not the demand that went unmet. Any model that learns from sales alone will underestimate demand in exactly the stores where a product did best.

What is the evidence that store-level optimisation lifts sales?

The Zara study remains the benchmark because it was tested against a control group. Its authors, academics working with Zara staff, described a model that took a size-assortment view of each store's inventory, linking the stock of every size to expected demand when choosing how many units to ship. The logic is intuitive to any merchandiser: a style missing its key sizes sells poorly even with units left on the rail. The estimated gain of 3 to 4% in sales came from distributing the same products more intelligently, not from changing the range.

H&M went a step further in 2018. As Retail Dive reported, citing The Wall Street Journal, the group used algorithms analysing receipts, returns and loyalty card data to tailor the merchandise mix of a store in Stockholm's affluent Östermalm district. The analysis showed most of the store's customers were women and that fashionable items such as pastel floral skirts, and higher-priced products, sold better than expected. Sales improved significantly, according to the report, though no figure was published. The context was pressure: H&M had about USD 4 billion of unsold inventory to clear and 4,288 stores worldwide, many stocked with similar merchandise. The Östermalm store was restocked with a wider price range, from USD 6 T-shirts to USD 107 cashmere sweaters and USD 118 leather bags, an early sign that local data can justify trading up as well as cutting back.

Which fashion retailers use AI to allocate stock store by store?

The most detailed recent case comes from Guess. According to a case study published in June 2025 by Nextail, its software provider, Guess allocates 50 to 60% of its stock for in-season needs and wanted to move away from rigid store clusters, high initial allocations and manual adjustments. It switched to fully automated replenishment driven by hyper-local demand forecasts, going live across EMEA in six months. Reported results: full-price sell-through up 5 percentage points, store coverage down 7.5% and 13% fewer stockouts. Daniel Botey, vice president of global inventory management, credited the change with better availability and data-driven, automated inventory decisions.

The scale makes the point. Guess reports more than EUR 2.8 billion in annual revenue and over 1,000 stores worldwide, plus partner-managed locations. At that size, a planner cannot review every SKU, store and day by hand, so the choice is between coarse rules applied to clusters of stores and a model that sets quantities for each location. "At Guess, we embrace disruptive technology and wanted to ensure solutions were purpose-built from the start," said Julien Jacotey, head of planning and allocation, explaining why the brand avoided a generic system.

Hackett London, the menswear brand owned by AWWG with more than 160 stores in over 30 countries, reported similar effects in a Nextail case study: 25% fewer lost sales in stores replenished by the system, 75% less time spent on replenishment tasks and 50% fewer weekly shipments from its Barcelona warehouse to the UK. "Stores were getting greater depth in the things that were selling, particularly best sellers," said Daniel Jenkins, merchandise manager. The case study gives no date, which limits comparison. Before the change, replenishment was manual, time-consuming and prone to error; afterwards, the brand reports a stockout rate below 1% and a daily report of the ten best sellers by location with their remaining stock cover.

Other chains have moved without publishing results. Jules, a French menswear retailer with 550 stores, said in February 2024 that it was using an AI platform for in-season allocation and replenishment to send stock where it is more likely to sell. In April 2026 AllSaints said it was adding AI-native tools across trading, forecasting, allocation and replenishment, markdown pricing and range planning, with weekly trading reports by department ready by 8am on Monday instead of later that day. "I want our teams to log on and be told what's happening across the globe in our brand," said Alfie Meekings, its chief transformation and technology officer.

What are the largest retailers doing with local demand data?

The biggest groups talk more about data infrastructure than about local ranges. Gap Inc. set up an Office of AI in 2024. Its former interim chief information officer said in 2022 that a predictive model processing sales data and product features informed buying, positioning and pricing, and could show which sizes of specific items sold best in individual stores. In its November 2025 results, Gap attributed a productivity gain of nearly 30% in its supply chain to automation and AI in its fulfilment network, and its chief financial officer, Katrina O'Connell, said disciplined inventory management left unit inventories slightly negative.

Walmart is localising in the classic sense. In April 2026 Bloomberg reported, as summarised by PYMNTS, that it was widening local and regional product ranges in stores, starting with categories such as coffee in Florida and local condiment brands elsewhere. A spokesperson said the company combines technology with merchant expertise to identify products local customers value and move them onto shelves quickly. Bloomberg noted the trade-off: local suppliers are usually smaller and more exposed to supply chain problems than national brands.

Fast Retailing shows that store-level performance does not always come from algorithms. Uniqlo's domestic sales rose 10% to about 1.03 trillion yen in the year to August 2025. Over five years, the group closed 30 Japanese stores while average sales floor per store grew 10% and average sales per store rose 13%, which Inside Retail, citing Nikkei, linked to larger floors displaying more products. The group uses customer reviews and support centre feedback to improve products and forecast demand, but its public reporting stresses bigger stores and fewer, stronger products, not store-specific ranges.

Online channels change the calculation too. Marks and Spencer wants to double annual online non-food sales to nearly £3 billion and raise the online share of its fashion, home and beauty sales from about 34% to 50% in the medium term, using more than 1,000 stores for click and collect and returns. When the full range is a click away and stores double as collection points, a store's physical range can be narrower and more local, as long as stock data across channels are reliable enough to promise what is actually available.

What data does store-level AI need to work?

Inventory accuracy comes first. Target's technology team reported in October 2023 that physical audits had revealed something uncomfortable: half of its out-of-stocks were unknown to its systems, because the records showed stock that was not on the shelf. Target built an ensemble of models that infer such phantom inventory from sales, movement and replenishment patterns, set the on-hand quantity to zero and trigger a refill, and reported a substantial sales lift on affected products without publishing a figure. The scale is instructive: nearly 2,000 stores and an assortment of more than 100,000 SKUs. Target also tested shelf-edge cameras and other sensors, then dropped them as detection methods because they were costly and hard to maintain at scale. In October 2025 Prat Vemana, its chief information and product officer, said Target was "using AI to spot when items are running low earlier than before."

For fashion, the lesson is direct. A model that tailors size curves or ranges per store from sales history will be wrong wherever recorded stock differs from reality, and it will read a sell-out as weak demand. Store-level AI needs reliable stock by SKU and store, ideally from item-level tagging, plus estimates of lost sales during stockouts.

• Sales and stock by SKU, store and day, cleaned of transfers, returns and promotions.
• Lost-sales estimates for periods when a size or colour was missing, so sell-outs are not read as low demand.
• Store attributes such as space, fixtures, climate, local events and customer mix from loyalty data where consent allows.
• Product attributes in consistent fields, so a new style can borrow history from similar ones.
• Operational constraints: minimum displays, pack sizes, delivery days and transfer costs.

What are the risks and limits of AI-driven localisation?

Most published results come from software vendors or the retailers that bought the software, without control groups, baselines or time frames. The Zara study is the exception, and its result, a gain of 3 to 4%, is a useful reality check against claims of double-digit improvements. Merchandisers evaluating results should ask what was compared with what, over which season and in how many stores.

Localisation also adds complexity. Different ranges per store mean more SKUs to buy, more packs to build, more visual merchandising variants and harder markdown decisions. Small stores produce little data per item, so models tend to fall back on clusters, which is close to the manual approach they were meant to replace. Walmart's caution about smaller local suppliers shows the supply side of the same problem.

There is also a trade-off between depth and availability. Guess reduced store coverage by 7.5% while cutting stockouts, which shows what good replenishment can achieve: less stock in stores, but in the right sizes. Pushed too far, lean stores leave little room for customers who want to browse, and a single delayed delivery empties a rail. Merchandisers need to set minimum display quantities that reflect the brand, not only the forecast.

Dependence on systems is a further risk. When a cyberattack hit Marks and Spencer in April 2025, it paralysed online sales, cost about £300 million in lost profit and forced the retailer back to manual processes, Reuters reported. M&S has since delivered a new merchandising capability that automates a previously manual task and is accelerating a new planning platform. The more allocation depends on automated decisions, the more a retailer needs fallback rules and people who still understand them.

Finally, there is the brand question. A retailer whose stores all look different risks diluting the identity customers expect, and an algorithm that only learns from what sold may narrow ranges towards safe bestsellers. H&M's Östermalm finding, that fashionable and higher-priced items sold better than expected, is a reminder that local data can also argue for more ambition, not less.

What should fashion merchandisers do now?

1. Fix stock accuracy first. Audit how often recorded store stock matches the shelf, and treat item-level tagging and cycle counts as a precondition for any local model.
2. Start with replenishment and size depth. Automate in-season replenishment and size curves before attempting different ranges per store, where results are better documented.
3. Hold back stock deliberately. Plan the share of buy reserved for in-season allocation, as Guess does with 50 to 60%, so local demand signals have something to act on.
4. Test against a control group. Run store pairs or regions with and without the new logic for at least a season and measure sell-through, markdowns and lost sales.
5. Define where local ranges add value. Limit true assortment differences to categories with clear local drivers, such as climate, tourism or local culture, and keep the brand core common.
6. Keep manual fallback rules. Document allocation rules that teams can run if systems fail, and keep planners trained to challenge the model's output.