How to improve your online store’s site search with AI
A customer lands on your store, types “women’s running shoes” into the search box and gets an empty page, because in your catalog those shoes are called “female running footwear”. They don’t ask, they don’t try again, they leave. That customer already knew what they wanted and said it in their own words, which is exactly the kind of visit that’s hardest to get and the one that hurts most to lose.
An AI-powered site search tries to fix that. Instead of looking for exact text matches, it understands synonyms, forgives spelling mistakes and connects what people type with what’s on the product pages, even when the words don’t match. It’s worth separating it from two things that are often sold together but are different: the shopping chat, which talks with the customer, and the recommendation engine, which suggests products without anyone searching for anything. Here we’re only talking about the search box, which is where people who already intend to buy come in.
Much of the improvement doesn’t require hiring anything. The order we suggest goes from free to expensive, and at each step you have data to decide whether the next one is worth it.
Before changing anything: what your customers are really searching for
The first thing is to know what your customers search for. That data is tucked away and nobody shows it on the main dashboard, but without it you’re judging any tool by eye, and by eye they all look good in the demo.
Where to find search data in GA4, Shopify, WooCommerce and PrestaShop
In Google Analytics 4, enhanced measurement records site searches if your store puts the term in the URL with a recognizable parameter, such as “q” or “s”. You turn it on in the web data stream settings, under enhanced measurement, and from then on every search is logged as an event with its term. To see it in a table you need to build an exploration with the search term dimension, because the standard reports don’t show it by default. Watch out for one detail: GA4 only stores what happens after you turn it on, so if you don’t have it today, today is the day to turn it on.
Shopify has its own store search reports in Analytics, including one for searches with no results, and they export to CSV. In PrestaShop there’s a store search statistics module that lists the terms and how many results they returned. WooCommerce is the weak spot: out of the box it doesn’t keep a search report, so you either rely on GA4 or install a search plugin that logs them.
Whatever the source, what you want in the end is a spreadsheet with three columns: the term, how many times it was searched and how many results it returned. If you can also cross it with clicks or add-to-carts, even better, but those three are enough to work with.

The three lists to review
The first list, searches with no results, is the most profitable one to review. Three kinds of things show up mixed together: products you sell but under a different name, typos, and products you don’t sell. The first two get fixed in the next section; the third is sales intelligence worth its weight in gold, because it tells you what people are asking for that you don’t have.
The second, searches with results but no clicks, is trickier. The search returns products, but people don’t click any of them, usually because the results have nothing to do with what they asked for or because the good stuff is in the third row. You only have this list if you track clicks: GA4 gives it to you with some work, and most search services include it out of the box.
The third, the most frequent searches, helps you prioritize. Open the top searches one by one and look at what the customer sees, through their eyes, on their phone. If the most repeated ones already work well, your problem is in the long tail, and that changes which tool suits you.
Synonyms and typos: the cheapest improvement
With the list of no-result searches in front of them, most stores find the same thing: the catalog talks like the supplier and the customer talks the way they talk. The supplier writes “hooded sweatshirt”, the customer writes “hoodie”, “hooded sweater” or “hoodded sweatshirt”. None of the three finds anything if the search engine compares literal text.
The fix is a synonyms table, and almost every platform has one. In Shopify you do it with the free Search & Discovery app, which lets you create synonym groups. PrestaShop has search aliases in the search settings, designed precisely for typos and alternative names. In WooCommerce it depends on the plugin you use, because WordPress’s default search doesn’t support them. This is where AI comes in in the most practical way possible: you give ChatGPT, Claude or Gemini your list of failed searches along with your category names, and ask it to suggest, for each term, which product or category it matches, or to flag it as “we don’t sell this”. You review what it returns, because it knows the language but not your catalog, and you load what you approve as synonyms.
In a Spanish store there’s a nuance that English-language articles don’t cover. If you sell in Catalonia, the Valencian Community or the Balearic Islands, some of your customers search in Catalan even if the store is in Spanish, and type “samarreta”, “motxilla” or “sabates”. The same happens with Basque and Galician in their regions. Those terms show up in the no-result searches list and are handled the same way, as synonyms of the product in Spanish. It’s cheap and hardly anyone does it.
This isn’t a one-off task. The sensible approach is to repeat it every month or whenever a new collection comes in, and with the prompt template already set up, it comes down to a short review.
Three levels of AI site search for online stores
Once you’ve cleaned up synonyms and still see searches failing, it’s time to decide which search engine you have. Skipping one of these three levels usually ends up costly.
Level 1: your platform’s search, properly configured
Before paying, look at what you already have. In Shopify, besides synonyms, Search & Discovery lets you configure filters, boost specific products for certain searches and hide others, and that fixes a good share of “the wrong thing shows up first” cases. In PrestaShop you can adjust the weight each product field has in search, for example so the name counts more than the long description. In WooCommerce, level 1 is poor out of the box, and in practice it already starts with a plugin.
The ceiling of this level is clear. The search still compares words, just with more help. If your customer types “something to give my dad, who loves the mountains”, none of these search engines understands the sentence. For many small stores that’s not a real problem, because people type two or three words and that’s it.
Level 2: smart search services like Doofinder or Algolia
This is where dedicated services come in: they connect to your catalog, index it on their own and replace your store’s search box with theirs. Doofinder has connectors for Shopify, WooCommerce and PrestaShop. Algolia is more technical and more flexible, and it’s widely used when there’s a developer behind it. Klevu targets medium and large catalogs. For WooCommerce there are more modest options that live inside WordPress, like FiboSearch or Relevanssi, which greatly improve the default search without reaching what external services do.
What they have in common is typo tolerance with no setup, autocomplete with products, semantic search to a greater or lesser degree depending on the plan, and above all a dashboard where you see what’s searched for and what fails without going through GA4.
What differs between them is how much you can tweak, how they charge and how long it takes to get running. Each one’s features change often, so before deciding, check on their website what the plan you’d actually need includes, not the top one.
Level 3: your own semantic search, only with a technical team
The custom version means turning each product page into an embedding, which is a numerical representation of the text’s meaning, using a model from OpenAI, Google or an open one, storing them in a vector database like Qdrant, Pinecone or PostgreSQL with pgvector, and comparing what the customer types against them. Engines like Meilisearch or Typesense already combine classic and semantic search, which saves part of the work. It works, and it gives you full control, but someone has to build it, reindex when the catalog changes and tune the results when semantic search returns similar items instead of the exact one. If nobody at your company knows what an API is, this level isn’t for you. Our recommendation for an SME is clear: exhaust level 2 before considering level 3, since it covers the same ground for the vast majority of catalogs.
How to test an AI search engine with your real searches
Vendor demos are done with prepared catalogs and searches chosen to show off. Your catalog has half-filled product pages, SKUs with odd names and discontinued products that are still visible, and that’s where you see whether the search engine is any good.

You build the test bench with what you pulled together in the first section:
- Take about fifty searches spread across the three lists.
- Add a few with typos exactly as people write them, some in Catalan if you sell there, and a couple of long phrases along the lines of “waterproof boots for a six-year-old boy”.
- Write down in the spreadsheet, for each search, which product should come up at the top. Do it before testing anything, so the result doesn’t win you over after the fact.
- Ask for trial access to two or three services with your real catalog and run the same searches on each one and on your current search.
- Mark in each row whether the right product shows up in the top positions, further down or not at all.
When you’re done, you have a table that shows which search engine gets it right most often for your customers, and with that you can negotiate.
Two things worth checking in the test, besides the hits. How it looks on mobile, because that’s where most people search and where a badly built dropdown covers half the screen. And how long it takes to reflect a stock or price change, because a search engine that recommends sold-out items is worse than one that recommends nothing. If we had to pick a single criterion, we’d go with hits on the no-result searches: that’s where you’re losing sales today.
How much it costs, how much work it takes and who maintains it
Level 2 services almost always charge a monthly fee that depends on volume, some by number of searches per month, others by number of indexed products and others by both, and the most advanced features, like full semantic search or personalization, are usually in the higher plans. As a ballpark, not a price list, entry plans run in the tens of euros a month and plans for high-traffic stores go into the hundreds. You’ll need to check the real price on each vendor’s website with your own numbers in front of you, because it changes often and because a traffic spike during the sales can push you into a higher tier.
What isn’t on the price list is the hours. Connecting the catalog with an official connector takes a few hours; tuning synonyms, rules and design until the search works smoothly usually takes a few days spread over a couple of weeks, and if your catalog has thin product pages, with supplier names and no attributes, no search engine works miracles and that cleanup work comes on top. Level 3 is a different league: it’s measured in weeks of development and a person to maintain it afterward.
The question people forget most is who maintains it. An AI search engine isn’t something you install and forget: every month there are new searches that fail, collections coming in and products going out. Someone at the store has to open the search dashboard, review the failed searches and add synonyms or rules, and that’s an hour or two a month in a mid-sized store, more in peak season. If nobody’s going to do it, the search engine slowly degrades, just like the default one, and you’ll have paid a fee for the same thing you already had.
If you’d like someone to go through your store’s searches with you and set up the test bench before you decide, get in touch.