Sneaker Product Matching

Style codes as the primary key

The strongest matching signal by a wide margin is the manufacturer style code. When two listings carry the same code, they almost always refer to the same model-and-colourway, regardless of how differently their titles read. A matcher that finds a shared style code can skip most of the fuzzy work and group the offers with high confidence.

The catch is coverage. Not every merchant publishes the style code, and those that do sometimes bury it in free text, mistype it, or include a regional suffix that makes an exact-string comparison fail. So the code is the primary key when present and clean, but matching cannot depend on it alone — a large share of listings arrive without a usable code and must be matched on weaker signals.

Title normalization and token matching

When the code is absent, the title carries the load. Matching normalises each title — lowercasing, stripping punctuation, expanding known abbreviations, mapping model aliases and nicknames to canonical model names — and then compares the resulting token sets. Two listings that share a brand, a canonical model and a colour token are strong candidates even if their raw titles look nothing alike.

This is where the alias catalogue built during normalization pays off. A collector nickname that appears in one title and the official name in another will only match if the system already knows they are the same model. Title matching therefore improves over time: every resolved alias makes the next ambiguous listing easier to place.

Image similarity as a secondary check

Images are a useful corroborating signal but a poor primary one. A perceptual comparison can confirm that two listings show the same silhouette in the same colourway, which helps distinguish near-duplicates that titles alone cannot separate. It is especially valuable for catching the case where two offers share a plausible title but actually depict different colourways.

Image matching has real limits, though. Merchant images are inconsistent in angle, lighting and crop; the wrong image is attached to a listing often enough that a strong image mismatch is more informative than a strong image match. In practice image similarity is best used to raise or lower confidence in a match proposed by stronger signals, not to propose matches on its own.

Colourways, sizes and variant structure

Colour is a matching signal and a splitting signal at once. Two offers for the same model in different colourways are different products and must not be merged, so the matcher has to resolve colourways carefully — reconciling official names, retailer colour words and translations before deciding whether colour unites or separates two listings.

Size is a variant dimension, not a matching dimension: offers for the same shoe in different sizes belong to the same product, attached as separate size-level availability. The matcher must keep these straight so that size variance does not fragment a product while colour variance correctly separates one.

Duplicate offers and confidence scoring

Within a cluster, the platform still has to handle duplicate offers from the same merchant — the same shoe re-listed, or split across regional storefronts. These are collapsed so a shopper does not see one retailer three times, while genuinely distinct offers from different merchants are all preserved.

Every match carries a confidence score built from the signals above:

• A clean shared style code contributes high confidence.
• Strong title-token agreement on brand, model and colour raises it.
• Image agreement corroborates; image disagreement penalises heavily.
• Conflicts — matching model but mismatched colour, or matching title but divergent image — lower confidence or block the match.

Low-confidence matches are held back rather than shown, because a wrong merge is more damaging than a missed one. This threshold — automatically merge, hold for review, or keep separate — is one of the most consequential tuning decisions a discovery platform makes.