In short
Keep keyword search for exact matches, add semantic search and query understanding for meaning, and let merchandisers stay in control.
Shoppers who use the search box usually already know what they want, so a bad result costs real sales. Yet most store search still matches words, not meaning. Try these on your own site:
- "black kurta for men size L under 5000"
- "shoes for standing all day"
- "gift for a 10 year old who likes science"
- a common misspelling of your best-selling product
If any of them returns nothing, or something odd, this post is for you.
Why keyword search breaks
Keyword search looks for the words in the query inside your product data. It cannot know that "standing all day" relates to "cushioned sole", that "L" is a size rather than part of a name, or that a shopper typing in Roman Urdu wants the same product as one typing in English. Teams patch this with synonym lists that nobody maintains.
What hybrid search does differently
We combine three layers:
- Keyword search stays. It is still the best tool for exact matches such as SKUs, brand names and model numbers.
- Semantic (vector) search turns product text and the query into embeddings that capture meaning, so "warm jacket for mountains" finds insulated parkas even if the word "warm" never appears.
- Query understanding. A language model reads long queries and pulls out structure: category, size, colour, price range. These become visible filters the shopper can remove, rather than hidden guesses.
The results from both searches are merged and re-ranked using rules you control: in stock first, promote margin where relevance is equal, never show discontinued items.
Recommendations come almost free
Once products have embeddings, "similar items" is a nearest-neighbour lookup. Blend that with what people actually viewed and bought together, and fall back to similarity for brand-new products that have no history yet. That fixes the classic cold-start problem.
Keep merchandisers in control
AI search should not become a black box the commercial team can't influence. Give them a console to:
- pin or boost products for a query ("eid collection")
- block results that are technically relevant but wrong for the brand
- preview changes before they go live
Measure it before you switch
Don't trust a demo. Replay a sample of real, anonymised search queries against the old and new engines and have your team judge the top results side by side. Then launch to a share of traffic behind a feature flag and watch zero-result searches, click-through from search and conversion from search. If the numbers don't improve, you will know before everyone sees it.
Where to start
Export your top few hundred search queries and the ones that returned zero results. That list tells you whether you need better synonyms, semantic search, query understanding, or all three, and it becomes your test set for everything after.
