Semantic Search for Fashion: Style, Occasion, and Fit Queries

The queries keyword search can't touch

Fashion shoppers describe intent, not products. Semantic search for fashion ecommerce has to handle three query types that keyword matching fails on completely.

Occasion, style, and fit

• Occasion: "outfit for an outdoor wedding," "something for a job interview," "beach holiday capsule."
• Style: "minimalist," "cottagecore," "smart casual," "quiet luxury."
• Fit and body: "high-waisted jeans for tall women," "a dress that's not too clingy," "wide-leg trousers that aren't see-through."

The shopper is describing a vibe or a constraint; the catalogue describes products in attributes and marketing copy. Semantic retrieval bridges that gap — but only over content that exists.

The data has to be there

If the catalogue only holds "blue dress, sizes S–XL," the search can't answer "something for a garden party." Occasion tags, style descriptors, and fit attributes have to be in the product data. On our semantic search project search quality tracked the richness of the composed field per product — the same principle applies harder in fashion, where the useful signal is style and occasion metadata that many catalogues simply don't capture.

Fit is where the money is

Fit-related returns are the largest cost in fashion. "True to size," "runs small," size charts, inseam and rise measurements — these are structured data the search should use directly, so "runs small, size up" surfaces in the results rather than in a one-star review. Getting fit data into search has real, measurable ROI.

Where text-only semantic search hits its limit

"More like this but in green," "the same silhouette, longer" — that's visual similarity, and it's a genuinely different problem that needs image embeddings, not text. It's a further step, worth scoping as its own project rather than assuming a text pipeline covers it.

Seasonality and merchandising control

New-season ranges and trend-driven boosts change fast — faster than model retraining. The "push these for spring" that the model can't know is what the merchandising console is for.

Getting style and occasion data in

The data investment has three routes:

• Supplier-provided tags — cheapest, and usually inconsistent between suppliers, so they need normalising.
• A manual merchandising taxonomy — accurate, and labour-heavy to build and maintain across a large catalogue.
• An LLM tagging pass — a model reads the product copy (and, where you have them, the images) and proposes style, occasion, and fit tags for a human to approve. Fast to cover a big catalogue, and the human review is what keeps it honest.

Most fashion retailers end up with a mix: an LLM pass for coverage, a taxonomy for the categories that matter most, and supplier tags as a starting point.

The fit-data return on investment

Fit-related returns are the largest single cost line in fashion e-commerce. Putting "runs small / true to size / runs large," real measurements, and model-height context into the search and the results page — rather than leaving it buried in reviews — is one of the highest-return data projects a fashion retailer can run. It reduces returns directly, and it makes the search better at the same time.

Where this stops being right

• A catalogue without style and occasion tagging underperforms on these queries — the data investment comes first.
• Visual similarity scoped into a text project by accident leads to a capability gap and a disappointed team — call it out early.
• Trend-driven relevance belongs in the merchandising console, not in retraining the model every season.

FAQ

Can text semantic search handle "something for a wedding"? Yes, if occasion and style are in the product data. The model bridges the intent, but only over content that exists.

What about "more like this in another colour"? That's visual similarity — a separate capability needing image embeddings, not something a text pipeline does.

Why does fit matter so much? Fit-related returns are the largest cost in fashion. Structured fit data the search can use directly reduces them.

Do we need image search to compete? Not to start. Text semantic search over good style, occasion, and fit data covers most fashion queries. Image similarity is a strong add-on, not a prerequisite — scope it as a second phase.

ISTRALLEN builds fashion search over style, occasion, and fit metadata, with visual similarity scoped separately; see AI for Retail.