Zalando makes 90% of its on-site content with AI, not its catalogue

Zalando cut campaign production from six to eight weeks to days and imagery costs by 90%. The figure covers its inspiration layer, not product pages.

Zalando says 90% of the marketing content on its website and apps was generated by AI in 2025, up from close to zero a year earlier, and that the change cut campaign production from six to eight weeks down to a few days. Its VP of content solutions, Matthias Haase, told Reuters in May 2025 that generative AI reduced the cost of producing imagery by 90%. What the 90% does not cover is the part of the site people actually buy from: product detail pages were excluded, and Zalando's own rules for the 7,000 brands selling on it still say a compliant model image must be photographed with a real person.

That distinction is the whole story for anyone running a retail business, because the 90% figure is now being used as shorthand for something much larger — that AI has started eating the software budget. At Zalando it did not come out of a software line at all. It came out of production: studios, shoots, retouching, and the yearly calendar that governed them.

What the 90% actually covers

Zalando put the figure in its full-year results on 12 March 2026, where it described scaling AI-generated product content from near zero to 90% in a single year, reducing campaign creation times from six weeks to days and boosting content output by 70%.

Two months later, in an interview published on its own corporate site, Haase narrowed it considerably. The 90%, he said, means the banners, trend highlights, campaign imagery and teasers you scroll past on the Zalando app and website — what he calls the inspiration layer. Not product detail pages. Not advertising off the platform. Hero campaigns shot with named talent still run as human-led productions; the 90% is what sits around them.

That is a much more ordinary claim, and a much more useful one. The work that was automated is the work that was being remade constantly and thrown away: 27 markets' worth of local, trend-led content that a six-to-eight-week production cycle could never keep up with.

The part that got automated is the part that expires. The part that sells is still photographed.

What did it cost, and what replaced it?

Zalando has not published a cost for the pipeline itself. What it has published is the effect: imagery that took six to eight weeks now takes about three to four days, at roughly 90% lower cost, according to Haase. Group capital expenditure was €222.5m in 2025, and the company guides to €240m–€300m for its technology platform and infrastructure in 2026.

What replaced the old process is a trend-detection loop rather than a calendar. Zalando says it reads search spikes, browsing shifts and sales patterns across its platform daily, validates them against a social-listening tool, and then generates a first visualisation from that signal. Its stated 2026 target is under 24 hours from spotting a trend to publishing.

The volume claim is the one worth noting: 170,000 AI-generated videos made from existing product photography, which now appear first on a product page when one exists. Producing video at that scale, the company says, was not feasible the old way.

On results, be careful with what Zalando has actually quantified. It reports around 10% higher engagement on this content and says customers who watch product videos are three times more likely to buy and return fewer items. For trend-reactive AI content specifically, it claims "a strong uplift" in add-to-wishlist and increases in add-to-cart — with no number attached. None of these figures is independently assured.

The quality layer nobody budgets for

The most transferable part of this deployment is the failure mode Zalando describes, because it is the one that costs money.

Generative models produce plausible images. When the source information is incomplete, Zalando says, they infer details that are not there — a seam, a texture, even a pocket — and those inventions are subtle enough to look convincing. In fashion retail, an image that flatters a garment into something it is not does not produce a sale. It produces a return.

So Zalando built a checking layer into the pipeline rather than at the end of it. A "product truth" engine compares each generated asset against the original product imagery, checking colour, construction and brand marks, and flags anything altered. The company's stated reason for building it is blunt: relying on the human eye alone to catch AI-generated deviations does not scale, and the models keep getting better at hiding the errors.

That is the line item most retailers leave out of the business case. The saving is on production. The new recurring cost is verification.

The rule Zalando pushed down to its 7,000 brands

If you sell through a marketplace, the more immediate consequence is not what the marketplace does with AI. It is what it now forbids you to do.

Zalando's partner image guidelines, updated 31 August 2026, say a model view only counts as compliant when it "must have been photographed with a real person". Images are rejected for poor-quality AI generation, and the guidelines name the tells: distorted logos, mispositioned buttons, unnatural teeth, facial features that shift across a set. Zalando reserves the right to reject any AI-generated asset that breaks believability or fails to represent the product accurately.

On disclosure, the rules split in two. From August 2026, no visible on-image labels are allowed at all — so partners are told simply not to submit content that would require one under law, which the guidelines define as deepfakes and AI-generated text on matters of public interest. By December 2026, partners must embed identifying data in metadata or as an invisible watermark, and must make sure it survives upload.

The legal backdrop is genuinely unsettled. The EU AI Act's Article 50 transparency obligations became applicable on 2 August 2026, requiring disclosure where AI-generated or manipulated image content constitutes a deep fake. On 18 June 2026, EuroCommerce — whose members include Amazon, H&M, Inditex and IKEA — wrote to the Commission's Henna Virkkunen asking that AI-generated ads not intended to mislead be excluded from that definition, arguing the alternative is labelling a very large share of all retail content. Reuters reported the letter on 22 June. The Commission had not responded at the time of that report.

Whichever way Brussels lands, the marketplace rule binds you now.

Does this apply to a 200-person retailer?

Partly, and only on one line. The production saving is real at any size, but it applies to content you remake and discard — seasonal banners, local campaigns, category headers — not to the catalogue images a customer buys from. If you sell on a marketplace, that marketplace's image policy already decides what you may submit, whatever your agency can now generate.

What it does not mean

It does not mean 90% of what a shopper sees on Zalando is synthetic. Product pages and off-platform advertising sit outside the number.

It does not mean this is a template. Zalando is a €12.3bn-revenue business with 62 million customers, around 3,000 tech specialists and its own foundational models, doing this at a scale where a week of production delay across 27 markets is a real cost. A 200-person retailer with four campaigns a year has a smaller problem and should buy a smaller answer.

And it does not mean a software vendor got rich here. Zalando names no vendor for the content pipeline; the partner it names is a photography studio, ORENDT STUDIOS, for the digital-twin model scans it began piloting in May 2025.

The next action is two lines of arithmetic, not a vendor call. Add up what you spent last year on photography, retouching and reshoots for content you replaced within twelve months — that, and only that, is the budget this touches. Then read your marketplace's image policy before you generate anything, because the rule about a real person in the shot was written by the same company that automated 90% of its own.