AI-powered merchandising tools do not predict the future by guessing — they read the present more carefully than any spreadsheet can. By combining point-of-sale data, returns signals, browse behaviour and size-level stock positions, these platforms can tell you which SKUs are quietly dying weeks before you would normally notice. If you work in buying or planning, understanding what the dashboards are actually computing changes how you act on them.
Key takeaways
- Sell-through rate alone is a lagging indicator; AI platforms layer in returns velocity, browse-to-basket ratios and size-curve distortion to catch problems earlier.
- Social media and search data can meaningfully improve demand forecasts for colour and fit trends when combined with proprietary sales data, according to peer-reviewed research on multinational retailers.
- Markdown optimisation systems tested in live e-commerce environments have demonstrated profitability improvements well above manual strategies, based on controlled online tests published in academic literature.
- Size-curve analytics — tracking which sizes sell out first and which linger — is one of the fastest-growing capabilities in fashion merchandising platforms.
- Low growth across global fashion retail, flagged in the McKinsey State of Fashion 2026 report, makes precision in buying and planning more commercially urgent than ever.
What does a merchandising analytics platform actually measure?
Most buyers are familiar with sell-through rate: units sold divided by units received, expressed as a percentage. It is a useful number, but it tells you what already happened. By the time a SKU shows a weak sell-through, you have often already missed the window to reorder a faster-moving alternative or shift promotional spend.
AI merchandising platforms add several layers on top of that baseline figure.
Returns velocity tracks not just how many units come back, but how quickly. A product with a high return rate in the first two weeks of sale is signalling a fit or quality problem — not a trend problem. Treating it the same as a slow seller leads to the wrong response.
Browse-to-basket and basket-to-purchase ratios reveal demand that never converted. A product with strong page views but weak add-to-cart rates may have a photography, sizing information or price-point problem rather than a product problem. A platform that surfaces this distinction saves a buyer from marking down something that could have been fixed with a content update.
Size-curve distortion is perhaps the most underused signal in traditional planning. When size 8 and size 16 sell out in week one while size 12 sits at full stock, the aggregate sell-through looks mediocre — but the product is actually strong. AI platforms that track size-level sell-through separately from total-unit sell-through catch this pattern and can trigger a targeted reorder rather than a blanket markdown.
How does the data pipeline work?
The underlying process has three stages: ingestion, modelling and action.
Stage 1: Ingestion
A modern merchandising platform pulls from multiple live feeds simultaneously — POS transactions (in-store and online), warehouse management systems, e-commerce browse logs, and returns processing data. The richer the feed, the more granular the signal. Platforms that connect to a brand's own ERP can also factor in committed stock on order, which changes the urgency calculation for any given SKU.
Stage 2: Modelling
This is where the AI work happens. The platform is not running a single model — it is typically running several in parallel. A demand forecasting model estimates where a SKU will land by end of season given current trajectory. A price-sensitivity model estimates how much of a markdown would be needed to clear remaining stock. A size-curve model identifies whether the distortion pattern is a reorder opportunity or a genuine slow-seller.
Research published in peer-reviewed operations management literature confirms that combining proprietary sales data with external signals — including social media volume and search index data — produces meaningfully better forecasts for colour and fit trends than sales data alone, particularly for products with short selling windows.
Deep learning approaches are increasingly applied to this problem. A case study from the online fashion industry published in 2023 found that global, data-driven forecasting models are particularly well-suited to fashion's specific challenges: high catalogue turnover, irregular demand patterns and fixed inventory constraints. Critically, these models need to track the relationship between price and demand closely — a markdown changes the demand signal, and a model that does not account for that will overestimate how much stock a price cut will clear.
Stage 3: Action
The output is not just a report. The best platforms translate model outputs into recommended actions: reorder this size, move this SKU to a different store cluster, trigger a markdown at this depth on this date. The buyer or planner still makes the final call, but the cognitive load of monitoring hundreds of SKUs simultaneously shifts to the machine.
Which platforms are doing this well?
Two platforms worth understanding in detail:
Style Arcade
Style Arcade is a buying and planning workspace built specifically for fashion merchandising teams. Its visual range planning interface lets buyers see the full assortment in a single view, with sell-through, stock cover and size-curve data overlaid directly on product images rather than buried in a pivot table. The platform connects sales, stock and inventory data across channels and stores, and generates reorder recommendations and demand forecasts from that combined feed. Style Arcade is particularly focused on size accuracy — the platform is designed to help teams move toward tighter size-curve planning, which reduces both overstock and returns. It suits mid-market to premium fashion retailers who want their planning and buying workflows in one connected environment.
Vue.ai
Vue.ai approaches the merchandising problem from a broader enterprise AI orchestration angle. Its retail-focused capabilities include inventory demand forecasting, catalogue assortment visualisation, automated product tagging and personalised e-commerce journeys. For merchandising analytics specifically, Vue.ai's strength is connecting the demand signal from the front end — what shoppers are browsing and how they are responding to recommendations — back into inventory and assortment decisions. It is built for large enterprise retailers who need AI capabilities across multiple workflow areas simultaneously and want to go live quickly. Vue.ai positions its platform as deployable within 90 days, which matters for retailers who cannot afford a multi-year implementation.
What does the research say about markdown optimisation?
Markdown timing and depth are where poor merchandising analytics cost retailers the most money. Mark down too early and you sacrifice margin on product that would have sold through at full price. Mark down too late and you are clearing at deeper discounts with less time left in the season.
An end-to-end machine learning framework for markdown management, tested in a live e-commerce environment and published at ACM SIGKDD in 2022, demonstrated profitability improvements of 86% and 79% relative to manual strategies in controlled online tests. The systems modelled price elasticity at the SKU level, adjusted recommendations dynamically as sell-through data accumulated, and outperformed experienced operations teams on the same decisions. These are not theoretical results — they came from real trading environments.
The implication for buyers and planners is straightforward: the question is not whether AI can improve markdown decisions, but whether the platform you are using is actually modelling price elasticity at a granular enough level to capture those gains.
What are the limits of current AI merchandising tools?
Honesty about what these platforms cannot yet do is as important as understanding what they can.
Cold-start problem. A new SKU with no sales history is difficult to forecast. Platforms handle this by using analogous products — similar silhouette, price point, colour family — as proxies, but the accuracy is lower than for established lines. Buyers should treat AI forecasts for genuinely new product categories with more scepticism than for repeat or evolved SKUs.
Causal confusion. A sudden spike in returns might reflect a fit problem, a fulfilment error, a viral negative review, or a competitor promotion that attracted the wrong customer. AI models can flag the spike; they cannot always diagnose the cause. Human interpretation remains essential.
Data quality dependency. Every model is only as good as its inputs. Retailers with inconsistent product attribution — where the same silhouette is described differently across seasons, or where size labelling is not standardised — will see degraded forecast accuracy. Cleaning the data taxonomy is unglamorous work, but it is a precondition for reliable AI output.
Channel interaction effects. When a product is promoted in-store and online simultaneously, isolating which channel drove the demand spike is genuinely hard. Platforms are improving at multi-touch attribution, but it remains an open problem in retail analytics.
What should buyers and planners actually do with this?
Three practical steps for teams evaluating or already using AI merchandising tools:
-
Audit your size-curve data first. If your platform is not tracking size-level sell-through separately from total-unit sell-through, you are missing one of the highest-value signals available. Ask your vendor specifically how size-curve distortion is surfaced in the dashboard.
-
Set action thresholds, not just alerts. An alert that tells you a SKU is underperforming is only useful if you know what action it should trigger. Work with your planning team to define in advance: at what sell-through rate and stock-cover combination does a reorder get triggered? At what point does a markdown recommendation go to a buyer for approval?
-
Track the model, not just the outcome. If a platform recommended a markdown and the product cleared, that is a good outcome — but it does not tell you whether the markdown was the right depth. Build a review cadence where you compare the recommended action to what actually happened, so you can calibrate how much you trust the model in different product categories.
FAQ
What is sell-through rate and why does it matter in fashion?
Sell-through rate is the percentage of received stock that sells within a given period. In fashion, where seasons are fixed and unsold inventory loses value quickly, a low sell-through rate signals that margin is at risk. AI platforms use it as one input among several rather than the sole performance metric.
How does AI improve demand forecasting for fashion specifically?
AI models can process more signals simultaneously than traditional statistical methods — including size-level sales, returns patterns, browse behaviour and external data like search volume. Research on multinational retailers found that adding social media and search data to proprietary sales data improved colour and fit forecasts meaningfully.
Can AI predict which trends will sell before the season starts?
AI can improve pre-season forecasts by analysing analogous historical patterns and external signals, but fashion involves genuine novelty that limits predictive accuracy for entirely new product. It is more reliable for evolved or repeat product than for category innovations.
What data does a merchandising AI platform need to work?
At minimum: POS transaction data, stock positions by size and location, and returns data. Better results come from adding browse and basket data, committed purchase orders, and historical sell-through by product attribute. Data quality and consistency matter as much as volume.
How do AI markdown tools decide the right discount depth?
They model price elasticity at the SKU level — estimating how demand changes at different price points — and combine that with remaining stock, days left in the season, and target sell-through. The goal is the minimum markdown depth that achieves the target clearance, preserving as much margin as possible.
What is the biggest mistake buyers make when using these platforms?
Treating the platform output as a decision rather than a recommendation. AI tools surface patterns and suggest actions; they cannot account for qualitative context a buyer holds — a supplier relationship issue, an upcoming editorial feature, a competitor move. The best outcomes come from teams that engage critically with the recommendations rather than accepting them automatically.
Further reading
- The Value of Social Media Data in Fashion Forecasting — Manufacturing and Service Operations Management
- Deep Learning based Forecasting: a case study from the online fashion industry — arXiv, 2023
- Promotheus: An End-to-End Machine Learning Framework for Optimizing Markdown — ACM SIGKDD, 2022
- The State of Fashion 2026 — McKinsey and Business of Fashion