Zalando — Pricing and Promotion Management Strategy
Most of what is written about “AI pricing” in fashion e-commerce is guesswork. Zalando is a rare exception. Its pricing team publishes research papers, speaks at conferences and describes its own teams in job postings; the company reports on margins and discounting; and its partner rules spell out what sellers may and may not do with prices.
I went through all of it. This case study uses only sources that can be opened and checked, and every claim links to one. Where something could not be verified, I say so.
In short: Zalando prices its own stock with algorithmic markdowns from a fixed starting price, optimised for long-term profit. People set the targets, override selected prices and are accountable for the result.
Who sets the price at Zalando?
It depends on the sales model. In the retail (wholesale) model, Zalando sets prices and merchandising and carries the inventory risk. In the Partner Program, the partner sets prices and carries the risk (Zalando Investor Factbook, Nov 2025).
Zalando retail: prices set by Zalando's algorithms, overseen by pricing managers.
Partner Program: partners set prices; Zalando charges a commission of 7–25%.
Lounge by Zalando: a closed shopping club with daily campaigns starting at 7am, running up to 72 hours, with discounts of up to 75%. It works on the wholesale model, so Zalando owns the stock (Partner University).
Physical outlets: up to 70% off the recommended retail price (Outlet by Zalando).
The algorithms described below price Zalando's own inventory.
A markdown machine, not a dynamic-pricing machine
Every article starts at its black price, the recommended retail price, which does not change over time. Zalando's system decides only how much to discount it: from 0% to 70%, in 5-point steps (Kunz et al., 2023).
Why markdowns rather than price rises? Because a large share of fashion inventory cannot be reordered mid-season, discounting is the main tool against end-of-season overstock (ibid.).
The scale is large. Zalando's pricing team reports about 600,000 products priced algorithmically at any time (Birr et al., 2026).
How the pricing loop works
The design is “forecast, then optimise”:
Forecast. A Transformer model, in production since 2019, predicts weekly demand for each article and country at a range of discount levels, 26 weeks ahead. Demand is constrained to rise as the discount rises (Kunz et al., 2023).
Optimise. An optimiser chooses the discounts that maximise long-term profit within business limits.
Guardrails. Limits cap week-on-week price changes and maximum discounts (Huelden et al., 2024). One example from a Zalando talk: the best articles may not get more than 10% off.
Human review. Category managers can override prices; pricing managers approve the overrides.
Learn. Prices go live, sales are collected and the models retrain.
The loop has a known weakness, which Zalando's Adèle Gouttes described openly: best-sellers are never heavily discounted, so the training data is biased by past decisions. Zalando counters this with decision rules and A/B tests (CO@Work 2024).
From 25 hard targets to one dial
The most instructive change is in how managers steer the system.
The old approach was one global optimisation with a financial target for each country, 25 in total. According to Gouttes, it was computationally expensive, conflicting targets needed arbitration, and the bi-weekly rhythm was too slow for urgent reactions.
The new approach splits the problem into independent article–country pairs. A pricing manager sets one parameter, alpha, which trades immediate revenue against long-term profit, and uses a map from alpha to the predicted financial result. Targets stopped being constraints and became outcomes.
The optimisation itself has also been sped up. A 2024 paper describes re-optimisation several times a week on a cluster of about 1,000 cores, with new heuristics improving revenue by 3–6% and profit by 2–5% (Streeck et al., 2024).
Sales events run on their own algorithm
Sales events last 6.2 days on average and take up about 35% of the year; discounts are about 5 points higher during them. For these periods Zalando built a separate daily demand model (LightGBM) and an optimiser that balances long-term profit against net merchandise value. Decision time fell “from hours to minutes” (Birr et al., 2026).
Zalando tested it in 23 A/B tests across 12 markets during 2023–2024 campaigns, comparing the algorithm with manually set discounts:
Profit: +6.0% (95% confidence interval 0.8–11.0%).
Net merchandise value: +2.2%, not statistically significant.
Germany: net merchandise value +17.9%.
The algorithm now handles most campaign pricing. The authors note what it does not yet cover: demand uncertainty, cross-price effects between products, and vouchers.
Humans stay in the loop, and get measured
Zalando's system is decision support, not full automation. Pricing managers own the financial targets and add context the data lacks, such as a competitor's campaign or a football tournament (CO@Work 2024). They can also intervene manually, as they did during COVID-19 (Birr et al., 2026).
Category managers can override algorithmic prices, usually weekly and only downwards, typically with one discount per product across all countries. Pricing managers approve or reject the overrides in batches per country. Overrides touched about 7% of daily price observations, about 15% of revenue and about 7% of profit.
Zalando then tested what those overrides were worth, in two field experiments (Huelden et al., 2024):
Blocking the overrides predicted to be weakest (about 32,900 products, 12 countries, 15 days): prices rose about 30%, sales fell about 65%, revenue fell about 60%, and profit rose about 237%.
Blocking all overrides (8 countries, 13 days): no significant change in revenue or profit across the sample, and blocking the better overrides reduced profit.
The lesson is not that “the algorithm always wins”. Human judgement varies in quality, and that quality can be predicted. The authors propose guardrails: let the algorithm set safe baseline prices and block only the overrides likely to destroy profit.
Price experiments, done carefully
Zalando randomises price tests by article (or article cluster), not by customer. Its researchers cite legal and PR concerns about customer-level price discrimination (Roemheld and Rao).
Article-level tests have a catch: shoppers substitute between similar products, so a discount on one item takes sales from its “control” neighbours, which inflates the measured effect. In two real tests, article-level randomisation estimated effects of 61% and 62%, while cluster-level randomisation estimated 41% and 35%.
A second trap is time. A short test can reward selling stock early, even if it would later have sold at a higher margin. Zalando's answer is Stock Lifetime Value: the expected profit from current stock over its remaining life. It is used to evaluate pricing initiatives and feeds quarterly business reviews (Decrouez et al., 2026).
Promotions, loyalty and the full price
Zalando's promotional calendar centres on sales events. In the 2018 Cyber Week, for example, it discounted a different category each day, with up to 70% off in-season fashion (Zalando, 2018).
Around the events sit loyalty and voucher mechanics:
Zalando Plus changed in September 2024 from a paid membership to a free, points-based programme with four tiers and perks such as birthday vouchers (Zalando, 2024). It covered 18.5 million customers in Q1 2026.
Plus members in Germany, France and Italy get early access to Lounge campaigns.
Promotional vouchers cannot be combined unless agreed otherwise, and some brands may be excluded (Zalando terms).
At the same time, Zalando works to protect the full price. After overstock hit margins in 2022, when gross margin fell from 41.8% to 39.2%, it named three levers for 2023: more cautious buying, “drive customer engagement towards black prices”, and an updated partner commission table (FY 2022 Financial Update). Lounge served as the channel for clearing the excess. In later quarters Zalando credited margin gains to better sell-through.
The rules for marketplace partners
Partners set their own prices, but within firm rules (Platform Rules v13.0; Zalando API docs):
A promotional price must be at least 10% below the regular price (5% for consumer electronics).
The black price must not significantly exceed the article's reference price, or the article may lose visibility or be deactivated.
A regular price cut by more than 60% or raised by more than 330% is flagged for review.
A regular-price increase cannot be combined with a promotion, and a promotion set too soon after an increase is rejected.
When several offers compete for the same product, the absolute lowest price wins; Zalando may require a partner offer to be 5% cheaper than its own wholesale offer.
Where the regulatory risk sits
The biggest external constraint on Zalando's promotions is the EU rule on announcing price reductions.
In January 2026 Poland's consumer authority UOKiK fined Zalando PLN 30,945,000. It found that the lowest price of the previous 30 days was not consistently shown, and that reductions were calculated from an “initial” or “regular” price instead of the 30-day low (UOKiK). The decision is not final, and Zalando said it would appeal.
Belgium's Economic Inspectorate raised similar issues about the 2022 summer sale, including “DEAL” banners without a real reduction, and referred the case to German authorities (ICPEN).
According to its partner documentation, Zalando now shows three prices when a reduction is announced: the original price, the lowest price in the last 30 days and the promotional price. In Poland, the sale section only includes articles reduced against the 30-day low (Partner University).
What we still don't know
Some questions cannot be answered from public sources:
Zalando's current discount rate. The company has not published it, or the value of its “price investments”, since 2022.
How competitive pricing works. Job postings mention a competitive pricing system with an “investment logic” and pricing intelligence, but give no details.
Personalised prices or vouchers. Stock Lifetime Value could be applied to personalised vouchers, but no results have been published.
ABOUT YOU after the acquisition. Zalando says ABOUT YOU's lower margin is temporarily diluting group gross margin, but has said nothing about changing its pricing.
Campaign rules for partners. Minimum markdowns for events like Cyber Week, and any co-funding of partner discounts, sit behind a partner login.
Five takeaways for e-commerce teams
Treat price as an inventory tool. Optimise profit over the stock's whole life, not this week's sales.
Give managers one interpretable dial. A single trade-off parameter is easier to steer than dozens of hard targets.
Measure human overrides like a model. Their quality varies and can be predicted, so block only the harmful ones.
Design price tests for substitution. Randomise by product cluster where you can, and value the stock you sell, not just the revenue.
Get discount presentation right first. In the EU, the 30-day lowest-price rule is where regulators look.
Sources
All sources were opened and checked on 28 October 2026.
Research and conference talks by Zalando employees
Birr, S., Huelden, T., Raslan, M., Gouttes, A. et al. (2026). High-Frequency Pricing at Scale for E-Commerce. arXiv.
Decrouez, G., Huelden, T., Nakhe, P., Prugger, D. (2026). Measuring Opportunity Cost with Stock Lifetime Value. arXiv.
Gouttes, A. (2024). Price optimization at Zalando. CO@Work, Berlin.
Huelden, T., Jascisens, V., Roemheld, L., Werner, T. (2024). Human-Machine Interactions in Pricing. DICE Discussion Paper 412.
Streeck, R., Gellert, T., Schmitt, A. et al. (2024). Tricks from the Trade for Large-Scale Markdown Pricing. arXiv.
Kunz, M., Birr, S. et al. (2023). Deep Learning based Forecasting: a case study from the online fashion industry. arXiv.
Roemheld, L., Rao, J. Interference Produces False-Positive Pricing Experiments. Working paper.
Zalando reports and documentation
Regulators
Komentarze