WooCommerce

AI search in an online store: semantic search instead of keyword matching

· Krzysztof Balicki

Glass and ceramic objects on a white table

AI search in an online store understands what the customer means instead of just comparing letters. That means it finds a product even when someone describes it in their own words. Standard WooCommerce and PrestaShop search, on the other hand, can return an empty page when a customer types a synonym, a word with a typo or a description of what they need instead of the product name. Frustrating? Very. Below we explain how semantic search works, what a good implementation looks like, what to ask the developer and how to check whether the change actually paid off.

How is AI store search different from plain keyword matching?

Classic search looks for exactly the words the customer typed in the product title and description. Nothing more. Semantic search works differently: it compares the meaning of the query with the meaning of the product. It does this with embeddings, which turn every product description and every query into a set of numbers that describe their meaning. Texts with similar meaning end up close to each other. According to Google’s embedding model documentation, these representations are used precisely for semantic search and give more accurate, context-aware results than keyword-based methods. In practice it looks like this: the query „rain shoes for kids” finds wellies, even though the word „rain” is not in their name.

Synonyms, typos and everyday language: what the customer really gains

A smart product search handles synonyms, Polish word forms and typos without dictionaries that someone would have to maintain by hand (and who actually does that regularly? exactly). Polish inflection is a real pain for simple keyword matching. „Kurtkę”, „kurtki” and „kurtkami” (forms of „jacket”) are three completely different strings for such an algorithm. A semantic model sees the same concept in all of them. It also copes with mixed languages, for example when a Polish customer types an English category name. The gemini-embedding-2 model described in Google’s documentation places content in many languages in a shared meaning space.

Where does classic search fail most often? On queries like these:

  • synonym: „pullover” instead of „sweater” used in the product name,
  • typo: „słuchawki bezprzewodowe” (wireless headphones) typed without Polish characters or with two letters swapped,
  • description of a need: „something warm for a winter walk with the dog”,
  • use case: „outdoor wood glue” for a product described only by technical parameters.

Semantic search and filters: how to combine AI with product attributes

Semantics is good at picking relevant products. But price, size, availability and brand filters still have to work strictly, on catalog data. That is why sensible solutions use hybrid search: semantic results are combined with exact matching of SKU codes, EAN numbers and model names, because every character counts there. In my opinion it also pays to catch filters typed straight into the search box, e.g. „black jacket XL”, and use them to narrow the results right away. The customer then does not have to click through the filter panel. One more thing. Embeddings run on product data: full descriptions, complete attributes, consistent categories. A thin catalog will give weak results even with the best model. Don’t count on a miracle.

What does a good semantic search implementation in WooCommerce and PrestaShop look like?

A good implementation starts with analyzing the queries customers already type. It ends with measuring results, not just with plugging in a model. A typical process looks roughly like this:

  1. Export search logs from recent months.
  2. Clean up product data: descriptions, attributes and categories.
  3. Index the catalog, which means calculating embeddings for all products.
  4. Configure hybrid search and filters.
  5. Test on real customer queries, not on examples from a sales presentation.
  6. Launch and monitor continuously.

The index has to update with every change to a product, price or stock level. Otherwise the search will happily show items that sold out long ago. And with an unusual catalog structure, a custom-built plugin or module works better than forcing an off-the-shelf solution to fit.

What questions should you ask an AI search developer?

The point is to check whether the developer thinks about data, running costs and measuring results. An impressive demo is not an implementation yet. I recommend asking:

  • how you will handle SKUs, EANs and product codes,
  • how long it takes for a new product to appear in the results,
  • where store data and customer queries will be processed,
  • how much model queries will cost under heavy traffic,
  • how the fallback search will work when the model API stops responding,
  • which metrics you will show after launch.

Also agree on who owns the index and configuration once the cooperation ends. This usually only comes up when you part ways. Finally, check whether the solution works well with your store theme and the filter plugins you already use.

How to measure the impact: zero-result searches and other metrics

The simplest and clearest metric? The share of searches with no results before and after the launch. Add clicks on search results, add-to-cart actions from search and repeat searches by the same customer. The last ones usually mean one thing: the first attempt missed. The list of the most common zero-result searches is basically a ready-made list of fixes, because it shows gaps in descriptions, attributes or configuration. Compare periods of the same length and with similar traffic. And promises of specific growth made before anyone has looked at the data? Treat them with a lot of caution.

Where should you start with AI search in your online store?

Pull customer queries from recent months out of your admin panel and check how many of them end on an empty page. That one number shows how many sales may be leaking through your current solution and which categories to start with. A high number? Then AI search in your store usually brings a quick, measurable effect. At k4.pl we build custom WooCommerce plugins and PrestaShop modules, so we can prepare a custom store search implementation tailored to your catalog and the filters you already have in place.

Frequently asked questions

Will AI search replace the standard WooCommerce search?

Yes, as the main search layer. Exact matching still stays for SKU codes, EAN numbers and filters, because there every character has to match. Combining both methods into one hybrid result works best.

Can semantic search handle Polish word forms and typos?

Yes, because it compares the meaning of the query, not the form of the word. Grammatical cases or missing Polish characters do not stop it from finding the product. But the quality of results still depends on product descriptions: the more complete the data, the more accurate the matches.

How long does it take to implement a smart product search?

It depends on the size of the catalog, the quality of the data and the number of integrations, for example with filters or a warehouse system. A reliable developer will give you a timeline after analyzing your queries and product data. Not before.