How does AI personalisation work in fashion e-commerce?
Recommendations, ranked listings, size advice and conversational assistants all rely on the same thing: clean customer and product data. A practical guide to what works, what it costs and where the limits are.
KEY TAKEAWAYS Summary by the editors
- AI personalisation in fashion e-commerce uses behavioural, transactional and product data to decide which products, sizes, content and messages each shopper sees.
- The most common applications are product recommendations, personalised ranking of category and search results, size and fit advice, personalised email and conversational shopping assistants.
- Zalando reports that moving its assistant to a newer language model lifted product clicks in recommendation carousels by 23 percent and wishlist additions by more than 40 percent compared with the previous version.
- Personalisation quality depends more on product attribute data and identity resolution than on the choice of algorithm.
- Online platforms in the EU that use recommender systems must explain the main parameters of those systems in plain language under Article 27 of the Digital Services Act.
AI personalisation in fashion e-commerce means using machine learning to decide, for each visitor, which products, sizes, images and messages to show. It works by combining what a shopper does (browsing, purchases, returns) with structured information about the products, and it delivers value only when both data sets are reliable.
What is AI personalisation in fashion e-commerce?
Personalisation is not a single feature. It is a set of decisions made on every page view: which products appear first in a category, which items fill a "you may also like" carousel, which size is suggested, which newsletter goes to whom and, increasingly, how a conversational assistant answers a styling question. Early rule-based systems used simple segments such as gender or last category visited. Current systems use machine learning models that learn patterns from millions of interactions and update rankings continuously.
Fashion is a demanding category for this. Assortments change every season, many items sell out before enough data has accumulated about them, taste is highly individual and size and fit drive a large share of returns. A recommender that works well for books or electronics does not automatically work for apparel.
How do fashion recommendation systems actually work?
Most production systems combine several techniques rather than relying on one model:
- Collaborative filtering: shoppers who behaved similarly in the past are assumed to like similar items. Strong for popular items, weak for new arrivals with no history.
- Content-based models: products are described by attributes (colour, silhouette, material, occasion, price band) and matched to the attributes a shopper has engaged with. This handles new items better but depends on rich, consistent product data.
- Visual similarity: computer vision models represent product images as vectors, so that items that look alike can be found and recommended, even when attribute data is thin.
- Large language models: used to interpret free-text requests ("something for a summer wedding in Rome") and to turn them into product queries, as in conversational assistants.
A ranking layer then blends these signals with business rules such as stock availability, margin, sustainability filters or brand agreements. That last step is where many personalisation projects succeed or fail, because a technically excellent recommendation is useless if the item is out of stock in the customer's size.
Which personalisation use cases deliver value in fashion?
| Use case | What it does | Key data needed | Typical maturity |
|---|---|---|---|
| Product recommendations | Suggests related or complementary items on product, basket and home pages | Clickstream, orders, product attributes, images | Mature, widely deployed |
| Personalised ranking | Reorders category and search results per visitor | Session behaviour, purchase history, stock by size | Mature at large retailers |
| Size and fit advice | Recommends a size based on past purchases, returns and garment measurements | Returns reasons, size charts, garment specs | Established, quality varies by brand data |
| Personalised email and push | Selects products and send times per customer | CRM profile, consent status, engagement history | Mature |
| Conversational assistants | Answers natural-language styling and product questions | Product catalogue, attributes, content, guardrails | Scaling since 2023 |
| Personalised discovery feeds | Shows inspirational content and products in a social-style feed | Content tagging, follows, engagement | Emerging |
Large platforms publish some of their results. Zalando's assistant launched in 2023 in four markets and expanded to all 25 of its markets in 2024, according to a case study published by OpenAI. After the assistant moved from GPT-3.5 to GPT-4o mini, Zalando reported a 23 percent increase in product clicks in recommendation carousels and more than 40 percent more wishlist additions compared with the previous version. These are platform-specific figures measured against Zalando's own earlier system, not industry benchmarks.
Zalando has also moved personalisation beyond the product grid. In October 2025 it extended an AI-driven discovery feed, with content tailored to user preferences, to 16 more countries, taking it to 22 markets, initially for selected iOS users, as reported by FashionUnited.
What data does fashion personalisation need?
The algorithm is rarely the bottleneck. Three data foundations matter more:
- Product attributes: consistent, granular descriptions of every item (fit, rise, neckline, fabric weight, occasion). Without them, content-based and conversational models cannot match intent to products.
- Identity resolution: linking app, web, store and email interactions to one customer profile, with consent recorded. Fragmented profiles make every model weaker.
- Inventory by size and location: recommendations must respect real availability, otherwise they generate frustration rather than sales.
Returns data is an underused signal. A model that knows a customer kept a size 38 in one brand but returned it in another can make better size suggestions than one that only sees orders.
Does personalisation pay off for fashion retailers?
Evidence is supportive but should be read carefully. McKinsey research published in 2021 found that 71 percent of consumers expect companies to deliver personalised interactions and 76 percent get frustrated when this does not happen, and that personalisation typically drives a revenue lift of 10 to 15 percent, with variation by sector and execution. These are cross-industry figures, not fashion-specific ones, and the uplift depends on starting point and data quality.
Costs are often underestimated. Beyond software, retailers need data engineering, product data enrichment, testing infrastructure and people who can interpret experiments. Smaller brands frequently get more from cleaning product data and using the personalisation features built into their commerce platform than from building custom models.
What are the risks and regulatory limits?
- Filter bubbles: over-personalised feeds can narrow what customers see and reduce discovery of new categories.
- Privacy and consent: tracking and profiling require a lawful basis and, for cookies and similar technologies, consent in many jurisdictions.
- Transparency: under Article 27 of the EU Digital Services Act, online platforms that use recommender systems must set out the main parameters of those systems in their terms and conditions in plain and intelligible language, and explain any options users have to modify them. This is directly relevant to fashion marketplaces.
- Bias: models trained on past sales can under-expose new designers, extended sizes or niche categories.
How should a fashion brand get started?
Start with one high-traffic surface, such as product page recommendations or category ranking, and define a clear success metric net of returns. Audit product attribute completeness before buying new tools, set up a holdout group from day one and review results with merchandising, not only with the e-commerce team. Personalisation works best when it reflects assortment strategy rather than overriding it.
Frequently asked questions
What is AI personalisation in online fashion retail?
It is the use of machine learning to tailor product rankings, recommendations, size advice and messages to each shopper. Models combine behavioural data such as browsing and purchases with structured product data. The aim is to show each customer the most relevant available items.
Which data is most important for fashion recommendations?
Rich and consistent product attributes, a unified customer profile with recorded consent, and accurate stock by size. Returns data is also valuable because it reveals fit problems. Without these foundations even advanced models perform poorly.
Does AI personalisation increase fashion sales?
Cross-industry research from McKinsey found a typical revenue lift of 10 to 15 percent from personalisation, with wide variation. Fashion retailers should verify gains with holdout tests and measure net revenue after returns rather than clicks.
Do EU rules apply to fashion recommender systems?
Yes, for online platforms such as marketplaces. Article 27 of the Digital Services Act requires them to explain the main parameters of their recommender systems in plain language and describe any options users have to change them. Data protection and consent rules apply to the underlying profiling as well.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
