AI search in an online store: cost and results

A customer of a bathroom fixtures store rarely knows the name of the category where what they are looking for sits. Instead, they type a sentence: “black bathtub faucet up to PLN 500” or “vanity with sink 80”. An ordinary search engine looks for those words in product names and either finds nothing or returns hundreds of random items. Filters work better, but you have to know where to click.

In Sklep Żukowscy from Giżycko I built a search engine that understands such sentences. It has been running since July 2026, and the store owner agreed to show data from its operation.

Starting point: customers don’t think in categories

In June 2026 the store had 3,109 products, and the search checked only the name, code and short description. It did not see the parameters (color, installation, material, dimensions), although 99% of products had them.

At the beginning of October 2026 the catalog already had 13,334 products, more than four times as many. With hundreds of faucets or cabinets in one category, scrolling through a list stops working.

Instead of a ready-made subscription service, I chose my own solution: the store had good data, a query to an AI model costs a fraction of a cent, and I decide what the customer sees.

How AI search works in an online store

The key rule: AI doesn’t choose products, it translates the customer’s sentence into filters. Products are chosen by the store’s ordinary code from the real catalog, so AI won’t show goods the store doesn’t have.

Step by step:

1. The customer types a sentence and presses Enter (suggestions while typing work without AI).
2. The AI model receives the sentence together with a list of categories and allowed filter values, and answers in a fixed format: category, filters, keywords, price from–to.
3. The store filters the catalog by these conditions and sorts the results from the most relevant.
4. Above the results the customer sees how AI understood the query, for example “AI zrozumiała: Baterie wannowe, Czarny, do 500 zł” (AI understood: bathtub faucets, black, up to PLN 500). When a result looks odd, it is immediately clear why.

Part of the work is deliberately done by ordinary code, because it is faster, free and always gives the same result:

• Brands I recognize using the list of brands from the store’s database. For “oltens vanity cabinet” the customer gets cabinets of that brand, not all the cabinets in the store.
• Dimensions (“shower tray 90x90”, “I have room for a 140 cm bathtub”) I compare with product dimensions, with a tolerance of 2 cm.
• Spelling without Polish diacritics is no obstacle: “wanna wolnostojaca” (freestanding bathtub, typed without Polish characters) once gave zero results, and after the fix, 23. Likewise “jednokomorowy” (single-bowl) and “1-komorowy” (1-bowl) are the same thing to the search engine.

What happens when AI isn’t sure

The AI model sometimes adds a feature the customer didn’t ask for, or understands the query too narrowly. So the search engine has several safeguards:

• Relaxing conditions one by one. When the full set of conditions gives zero results, the search engine drops them gradually: first unknown words, then additional features, finally color and price.
• A note instead of pretending. When a product isn’t available in a given color or size, the customer reads about it plainly, and below sees other colors or the nearest sizes.
• An empty result when nothing matches. Previously the search engine returned the whole category in that case: for “porcelain tile” it showed 88 random products, because the store had no tiles yet. Since 2 September 2026 the customer gets the message “Brak dokładnych dopasowań” (No exact matches) and a hint on how to change the phrase.
• The same answer to the same question. The AI model doesn’t always understand the same question identically, so the search engine remembers its interpretation for 30 minutes. A repeat takes a fraction of a second and costs nothing.
• Fallback mode. When the main model doesn’t respond, the query goes to a backup model from another provider. When that fails too, the store searches the ordinary way, by words.

Where does this safeguard come from? In August 2026, before there was a backup model, AI did not respond for three days and the store searched only by words: for “black washbasin faucet” it found nothing, and after AI came back it showed 956 products. That is why it’s worth checking AI’s operation regularly.

What the search report showed

Each search is saved anonymously in the store’s panel, without customer data. The owner sees the most frequent phrases and queries with no results, and can download the report as a CSV file, for example for Google Ads campaigns.

The search engine turned out to be a more important channel than the chat: in the week of 19–25 September 2026, the chat with the AI assistant had one real conversation, while the search engine had dozens of queries. From 13 September to 16 October the report recorded 663 searches, including about 200 technical repeats (the page sent the query twice; fixed on 16 October).

Examples of fixes from 25 September 2026:

The most important lesson concerned data, however, not AI. For “granite kitchen sink” the search engine showed 6 products, although the store had 369. AI understood the query flawlessly (sinks, material granite), but 364 of them had no material entered in the supplier’s data (status as of 2 September 2026).

The owner then asked whether another model, for example ChatGPT, would help. It would not: no model can find a feature that isn’t in the data.

How much one search costs

The first version, from July 2026, ran on Google’s Gemini 2.5 Flash-Lite model: fast (about 0.9 s) and very cheap. According to Google’s price list as of 25 October 2026, it costs $0.10 per million input tokens and $0.40 for output. In September the owner chose Anthropic’s Claude model: slower (2–3 s), but better at handling subcategories and typos. Gemini became the backup model.

I compare models on the same queries from the report, not on impressions:

• 25 September 2026: Claude Sonnet 5 gave 30 accurate answers out of 31, and the cheaper Claude Haiku 4.5 gave 27 to 28. By estimate, Sonnet would cost about $8 a month and Haiku about $3, but $5 of savings wasn’t worth the worse results.
• 29 September 2026: the newer Claude Sonnet 5.5 gave 85 accurate answers out of 86 (Sonnet 5: 84), and the median response time fell from 2.56 s to 1.84 s at the same cost, about $0.006 per query. Since 30 September the search has been running on Sonnet 5.5.

According to Anthropic’s pricing as of 25 October 2026, Sonnet 5.5 costs $2 per million input tokens and $10 for output. One search is about $0.006, that is, about 2.4 groszy (hundredths of a zloty) at the National Bank of Poland exchange rate of 23 October 2026 (PLN 3.9247), and the estimated $8 a month is about PLN 31.

So the AI bill is the smallest item. The most expensive part is the refinement: from July to October the search engine received over a dozen fixes, and since the end of September I check each one on several dozen queries from the report before it goes to the store.

Information for the buyer: AI is searching

Since 2 August 2026, Art. 50 of the AI Act applies. AI systems intended for direct interaction with people must inform that a person is dealing with AI, unless that is obvious (para. 1). The information must be clear and distinct, given at the latest at the first interaction and compliant with accessibility requirements (para. 5). The Omnibus package (in force since 27 July 2026) postponed, among other things, the deadlines for high-risk systems, but not this obligation; the European Commission still gives August 2026 as the start of the transparency rules. As of 27 October 2026; this is not legal advice.

Is a search engine that merely converts a sentence into filters such a system? The provision doesn’t say so explicitly. I choose the cautious version, because the label is a few lines of code:

• in the search field there is a permanent badge “Wyszukiwarka AI” (AI search; on a phone, just “AI”),
• above the results are “AI zrozumiała:” (AI understood:) and the filters that AI set.

On your own site, also add the search engine to your privacy policy: the content of the query goes to the AI model provider. I write more about Art. 50 in the guide AI chatbot and the AI Act, and I adapt the labels for the chat and the search engine as part of the AI Act package.

Does an AI search engine make sense in your store

It pays off when you have hundreds or thousands of products with parameters (color, dimension, material), and customers describe a need (“for a 140 cm bathtub”) instead of knowing the model name. It makes no sense when the entire range fits on one screen or customers buy by catalog number.

I can’t tell you today how many orders the search engine alone brought, because I haven’t separated it in the measurements yet.

Checklist before implementation:

1. Check how many products in the main categories have color, material and dimensions entered.
2. Save queries anonymously and review the ones with no results every week.
3. Plan a fallback mode: a backup model and ordinary word-based search.
4. Label the search engine as AI and show the customer how it understood the query.
5. Set a daily cost limit and measure how many searches end with an item added to the cart.

Summary

An AI search engine doesn’t replace filters or good data, it sets the filters for the customer. In Sklep Żukowscy one search costs about 2.4 groszy, and the biggest gains came from fixes based on real customer queries.

I describe the whole store on the Sklep Żukowscy project page. If you want such a search engine in your store, see the AI search and shopping assistant service or write to me: I will check your data and tell you whether it pays off.