Why beauty breaks a fashion platform's assumptions
The instinct when a marketplace adds a category is to treat it as more catalogue. Beauty punishes that instinct, and it does so in ways that reach right down into fulfilment and merchandising.
Start with the basket. Beauty orders run around four items; fashion runs closer to one and a half. That single difference changes the economics twice over — it demands far deeper assortment to cover every shade and size variant, and it makes split shipments a real cost, because four items sourced across a hybrid of owned inventory and marketplace sellers will fragment unless something actively prevents it.
Then the variant model. A lipstick is not one product; it is a shade family where the customer's decision is the shade, made on a screen, with the return risk sitting almost entirely on getting that decision wrong. Fashion size charts do not prepare a catalogue model for this.
Then the buying behaviour. Beauty is replenishment shopping with high repeat and high lifetime value, which rewards knowing a customer's skin, hair and preferences over time rather than treating each session as new. And it carries anxieties fashion does not — counterfeit fear, and a fast-rising demand for ingredient transparency.
Beauty is not fashion with different products. The basket, the variant model and the buying anxieties are all different.
Establishing the thesis before the build
The category was sized properly first: the overall beauty and personal care market, the share transacting online — low single digits at the time — and the growth rate that made the gap interesting rather than discouraging. A category that is 3% online in a market growing steadily is a different investment case from a mature one.
That was paired with a competitive read of the players who already owned online beauty, and with primary customer research — qualitative interviews and questionnaires — rather than assumptions inherited from the fashion business. The research ran on SurveyAnalytica, our own research and customer-intelligence platform, which meant instrument design, fielding and analysis stayed in one place and the findings came back structured enough to drive a data model rather than a slide.
The research produced insights that each had to earn a product decision. Beauty buying leans on peer recommendation and expert validation, so the design answered with a council of experts providing ratings, reviews and authority. It is an intensely visual category, which shaped the content strategy. Counterfeit anxiety demanded authenticity signals at specific touchpoints rather than a generic trust badge. Ingredient transparency demanded product data the existing catalogue model simply did not carry.
From insight to data model
The part that determines whether a category launch works is unglamorous: the taxonomy and the customer model.
We defined a three-level category hierarchy across makeup, skin, fragrance and wellness — down to the level of setting sprays, eyeshadow palettes, under-eye serums and lip stains — because filters, merchandising, recommendations and reporting all inherit from it, and retrofitting a taxonomy after launch is close to impossible on a live catalogue.
Alongside it, a customer attribute model built for this category specifically: skin type, tone and concern; hair type, colour and concern; fragrance note preferences; makeup usage frequency; and loyalty stage. Some attributes exist to power recommendations directly — city, for example, because weather drives what suits a customer's skin.
The personalisation roadmap was staged deliberately rather than promised all at once: most-suited recommendations first, hyper-personalisation next, bespoke solutions last. That order reflects how much data the model actually has at each stage.
Specifying the build against the platform that existed
The vertical had to run on the group's existing commerce platform, so the design was expressed in that platform's own terms: roughly fifty storefront components mapped to the content model — hero and split banners, flash sale, curated listing strips, smart filter widgets, automated brand and product carousels, recommendation slots — alongside catalogue templates, product image guidelines and filter behaviour.
Working in the platform's component vocabulary rather than in pure design terms is what makes a specification buildable. It also exposes early where the platform genuinely cannot do something, while there is still time to decide whether to extend it or drop the requirement.
A design system and wireframes were produced in parallel, so the visual language and the component contract stayed in step.
Building the launch to keep learning
A new category is a set of hypotheses, so the launch shipped with an experimentation framework rather than a set of opinions: segment users into cells, run variants, promote the winner to control.
The tests chosen were the ones where the team genuinely disagreed. What order should product-page components appear in? Should a recommendation carousel offer add-to-bag directly, or push the customer to the product page first — which is really a question about two different shopper mindsets? And the one specific to this category: how should a shade selector work?
That last test carried the most interesting success metric. Alongside the usual conversion and browsing measures, it tracked returns caused by incorrect shade selection — tying an interface decision directly to a cost line. Most A/B tests measure whether a change lifts conversion. This one measured whether it lifted conversion at the expense of returns, which for beauty is the question that matters.
What I would tell someone facing this
- Check whether the new category breaks your platform's assumptions before you scope the build. Basket size, variant model and return drivers are where the surprises live.
- Fix the taxonomy first. Filters, merchandising, recommendations and reporting all inherit from it, and it is the hardest thing to change once a catalogue is live.
- Build the customer attribute model for the category, not the company. Skin concern and fragrance notes have no equivalent in fashion, and generic profile fields will not substitute.
- Stage the personalisation promise against the data you will actually have. Promising bespoke recommendations at launch guarantees an underwhelming launch.
- Specify in your platform's component vocabulary. A design that cannot be expressed that way is a design that will be reinterpreted during build.
- Instrument the metric the category actually cares about. In beauty, a conversion lift that increases shade-related returns is not a win.
Research on our own platform
The customer research for this engagement ran on SurveyAnalytica, the research and customer-intelligence product we build. That is worth stating plainly because it changes what the research can do.
Most consulting research ends as a deck: findings summarised, then manually translated into requirements, losing structure at every step. Running it on a platform we control meant the outputs stayed structured — attribute by attribute, segment by segment — and could be carried directly into the customer attribute model and the personalisation design without a translation layer in between.
It also meant the client kept an instrument they could re-run. Category assumptions age; a research capability that outlives the engagement is worth more than a point-in-time answer.
Technology
- SurveyAnalytica (own research platform)
- SAP Commerce (hybris)
- WCMS component model
- Adobe Target
- Product content management (PCM/PIM)
- Recommendation and personalisation services
- Figma design system