Why Fashion Product Data Is Hard

Brand names are not standardised

The first field you would expect to be clean — the brand — is often the messiest. Retailers spell the same brand differently, abbreviate it, translate it, or fold sub-brands and collaboration partners into a single string. One shop lists a shoe under adidas Originals, another under Adidas, a third under adidas Consortium, and a fourth buries the brand in the product title with no dedicated field at all.

Collaborations compound the problem. A shoe co-designed by a brand and an external partner may be filed under either name, both, or a joint label, and the ordering is inconsistent. Any system that treats the brand string as authoritative will fracture the same product into several apparent brands.

Model names, aliases and nicknames

Below the brand sits the model, and this is where the data becomes genuinely adversarial. A single model line accumulates official names, marketing names, internal names and community nicknames, all of which appear in live listings:

• The official model name as the brand markets it.
• Silhouette shorthand used by retailers to save space.
• Collector nicknames that never appear in any official material but dominate how buyers search.
• Generation or version numbers that some shops include and others omit.

None of these is wrong, exactly, but they cannot be matched by string comparison. Two listings for the same shoe may share not a single word in their titles. The reliable anchor is the manufacturer style code — when it is present, which is not always.

Duplicate and near-duplicate products

Even within a single retailer, the same physical product can appear more than once: re-listed after a catalogue migration, split across regional storefronts, or duplicated because a size variant was created as a separate product. Across retailers, duplication is the default state — that is the entire point of a comparison platform — but it means the platform must decide, for every incoming listing, whether it is a new product or another offer on an existing one.

Near-duplicates are worse than exact duplicates. A shoe and its slightly different regional release, or a general-release version and a special-edition version, may differ in ways that matter to a buyer but are invisible in a truncated title. Collapsing them loses real distinctions; keeping them apart fragments the catalogue.

Multilingual titles and colour naming

Cross-border retail multiplies every problem by the number of languages involved. The same shoe carries a title in Dutch on one storefront, German on another and English on a third, with word order, gender terms and descriptive adjectives all shifting. Colour is the sharpest example of this. Brands assign official colourway names that are often poetic rather than descriptive, retailers translate or replace them with plain colour words, and the two rarely line up.

A listing described as Sail / Light Bone by one shop and simply white by another may be the identical shoe. The colourway problem deserves its own treatment, but note here that colour is simultaneously one of the most important matching signals and one of the least standardised fields in the entire dataset.

Size systems and per-size stock

Footwear sizing is not one system but several — men's, women's, EU, UK, US, and Japanese conventions — and retailers label the same physical size differently depending on their market and their audience. A discovery platform that wants to answer "is this available in my size" has to normalise all of these to a common scale, then reconcile them against stock that is tracked at the size level, not the product level.

Product-level stock is close to meaningless for footwear. A shoe listed as "in stock" may be available only in sizes no shopper wants. Genuine size-level availability is what buyers care about, and it is precisely the data that merchant feeds report least reliably. This is developed further under size-level availability.

Stale prices, feed errors and images

Even once a product is identified and sized, the offer data attached to it decays. Prices change faster than feeds refresh, so a platform routinely shows a price that no longer holds. Sale prices are especially fragile: a discount may have expired, or the "was" price may be inflated to make the discount look larger. Distinguishing a reliable current price from a stale or misleading one is a core trust problem, covered in price tracking.

Merchant feeds also simply contain errors — mislabelled fields, mismatched images showing a different colourway or a different shoe entirely, broken product URLs, and truncated titles that drop the exact word needed to identify the model. Images in particular cannot be trusted as ground truth: an image is a useful matching signal but a poor authority, because the wrong image is attached to a listing often enough to matter.