Almost every store search starts the same way: a box, a few words, a list of products. Behind that box, though, there are two very different ways to decide what “matches”. Understanding the difference explains most of the frustrating results shoppers see, and what it takes to fix them.
Key takeaways
- Keyword search matches words through an inverted index. It’s fast, exact and great for codes and brands.
- Semantic search matches meaning through embeddings. It understands descriptions, synonyms and questions.
- Each one is strong exactly where the other is weak.
- The best store search is hybrid: instant keyword results while typing, meaning when it’s needed.
How keyword search works
Keyword search, also called lexical search, looks for the words in the query inside the text of your products: titles, descriptions, tags. The classic setup builds an inverted index, a map from every word to the products that contain it, much like the index at the back of a book.
When a query comes in, the engine looks up each word and ranks the products that contain them. Common ranking formulas such as BM25 give more weight to words that are rare in the catalog and appear prominently in a product, so “linen” counts for more than “the”.
Where keyword search is strong
- Exact things. Product codes, SKUs, model numbers and brand names. If someone types “40211”, they want that product, not something similar.
- Speed and predictability. Lookups in an inverted index are very fast, and it’s easy to see why a product matched.
- Short, specific queries. “white sneakers” works well when your products are called white sneakers.
Where it breaks
- Typos. “snekers” is a different word, so exact matching finds nothing.
- Different vocabulary. The shopper says “trainers”, the catalog says “sneakers”.
- Descriptions and questions. “something light for a summer wedding” shares almost no words with the dress that fits it.
- Language details. Plurals, prefixes and word forms need extra handling, which matters a lot in languages like Hebrew.
Making keyword search more forgiving
Before reaching for AI, there’s a lot keyword search can do with the right techniques:
- N-grams. Breaking words into overlapping pieces of three letters (“sne”, “nek”, “eke”…) means a typo still shares most of its pieces with the right word.
- Phonetic matching. Comparing how words sound catches spelling variants.
- Stemming and normalization. Treating “dresses” and “dress” as the same word, and handling prefixes and final letters in languages that need it.
- Synonyms. Telling the engine that “couch” and “sofa” belong together.
These fix typos and vocabulary. They still can’t understand a description.
Three queries, and what each approach does with them
snekersSneakersn-grams and sound
light, for a weddingLinen dressmeaning (embeddings)
40211That product, firstexact keyword match
How semantic search works
Semantic search compares meaning instead of words. A machine-learning model turns text into an embedding: a list of numbers that places the text as a point in a space where similar meanings sit close together. Every product gets an embedding from its title, description and attributes. When a query arrives, it gets an embedding too, and the engine returns the products whose points are nearest.
Because the comparison happens in meaning space, “something light for a summer wedding” lands near a white linen dress even though the words barely overlap. To keep this fast across thousands of products, engines use approximate nearest-neighbour indexes that find close points without comparing every single one.
Where semantic search is strong
- Natural language and descriptive queries.
- Different vocabulary, without maintaining synonym lists by hand.
- Questions, when combined with a language model that writes an answer from the matching products and pages.
Where it struggles
- Exact identifiers. A product code has no “meaning”, so the nearest point might be the wrong product.
- Precision on short queries. “black” is close to many things; keyword search is often sharper here.
- Cost and latency. Turning every query into an embedding takes compute, which is why it shouldn’t run on every keystroke.
Why the best store search is hybrid
Each approach is strong exactly where the other is weak, so good store search combines them. A typical hybrid design looks like this:
- Instant keyword results while typing. Fast, typo-tolerant lexical search runs as the shopper types, so suggestions appear from the first letter.
- Exact matches first. If the query is a product code, that product goes straight to the top.
- Meaning when it’s needed. When the shopper presses enter, or the query reads like a description or a question, semantic search joins in.
- Blended ranking. Results from both are merged, so a product that matches both the words and the meaning rises to the top.
What this means for your store
If your search fails on typos and synonyms, you need better keyword search. If it fails on descriptions and questions, you need semantic search. Most stores see both in their failed-search reports, which is why the question isn’t “keyword or semantic” but “how well are they combined”.
Frequently asked questions
What is the difference between keyword search and semantic search?
Keyword search matches the words in a query to the words in your products. Semantic search turns both into embeddings and matches by meaning, so it can find relevant products even when the words are different.
Is semantic search always better?
No. It’s better for descriptive and natural-language queries, but keyword search is more precise for product codes, brand names and very short queries. The strongest setups use both.
What are embeddings?
Embeddings are lists of numbers produced by a machine-learning model that represent the meaning of a piece of text. Texts with similar meanings get similar embeddings, which lets a search engine compare meaning mathematically.
Does semantic search slow down a store?
It can if it runs on every keystroke. A common approach is to show instant keyword suggestions while typing and add semantic search only when the shopper submits the query.
Sources
- Inverted index, Wikipedia
- Okapi BM25, Wikipedia
- Word embedding, Wikipedia
- Nearest neighbor search, Wikipedia
Back-end engineer and AI specialist who thinks in vectors and dreams in latency budgets. Aviad builds the engine under Lets: the indexes, the embeddings, the AI answers and the systems that keep them fast and secure. If a query takes longer than a blink, he takes it personally.
See it on your store.
Leave your details and we'll set Lets up on your catalog for a free trial. Typos, questions and photos, answered from your own products. No credit card needed.
You're in. We'll email you shortly with how to try Lets on your store.
