AI Product Search in an Online Store: Helping Customers Find What They Need

A shopper types “jacket for rainy autumn” into the search box. In the catalog this model is described as a “membrane mid-season jacket with taped seams”. Not a single word in common — and the regular search shows an empty page, although the product is in stock. The person goes to a competitor, and the store never even learns that it lost a sale.

AI product search works differently: it tries to understand what the person means and match the intent with product characteristics. Rainy autumn means water resistance, mid-season use and possibly a hood. But the quality of such search depends less on the “intelligence” of the language model and more on how well the catalog is described.

Three types of shoppers in your search

The easiest way to design search is to start not from technology but from the people who use it. In most stores there are three different scenarios.

The one who knows exactly

They enter an article number, model or brand name: “Bosch GSR 12V-15”. They need a predictable result, not a “smart” one — an exact match in the first row. Classic lexical search works here, and it must not be spoiled by “understanding the meaning”.

The one who describes a problem

They write “quiet air conditioner for the bedroom” or “running shoes for winter”. These words are not in the catalog, but noise level, tread and insulation are. It is this shopper who benefits most from semantic search, which compares the meaning of the query and the product rather than individual words.

The one who is still choosing

They simply type “laptop” and get a thousand items. They need not results but narrowing down: for work or gaming, what budget, whether weight matters. Clarifying questions and filters matched to the intent are useful for such a query.

Why internal search affects sales so much

Search is used by people who have already formulated a need. This is the store’s hottest audience: visitors who used search usually buy noticeably more often than those who just browse the catalog. An empty results page is therefore a loss of a customer at the very moment they were ready to pay.

The problems are usually typical. Search does not understand typos and transliteration, so “aifon” and “iphone” give different results. Synonyms are not linked. Results ignore availability, and out-of-stock products come first. And zero-result queries are never analyzed.

The hybrid scheme and why it wins

In practice, hybrid search that combines both mechanisms turns out to be optimal for e-commerce. Exact matching handles article numbers, codes and models, semantics handles natural language, and the ranking system decides whose result to show higher.

The chain looks like this: typo correction and query normalization, intent detection, candidate selection by both mechanisms, filters, ranking, availability check and the final results page.

One principle matters more than technology here: commercial priority must not break relevance. If the system always pushes the highest-margin products, shoppers will quickly stop trusting search. A safer order is to filter out irrelevant items first, then check availability, and only within the relevant group slightly boost the items the business needs.

What data is required to make it work

A language model does not know why one jacket suits travel and another does not, if that is not in the data. The foundation of smart search in an online store is correct names and descriptions, categories, brands, filled-in attributes, variants, compatibility, up-to-date prices and availability, plus a synonym dictionary.

Behavioral signals are worth collecting separately: what people searched for, what they clicked, what they added to the cart. They help ranking more than any additional model. How to bring all of this into a single structure is covered in our article on preparing a product catalog, and the principles of working with queries in the material on online store semantics.

Zero results are a free source of insights

The “nothing found” page is a list of what shoppers want but do not get. Review such queries every month. Some of them mean the product really is missing from the assortment — a hint for purchasing. Others mean the product exists but a synonym is missing. Sometimes an attribute people search by is simply not filled in, or the query is about delivery or order status at all.

In some cases it is enough to extend the dictionary, in others to enrich product cards, in others to show a reference page instead of an empty screen. Even without any AI, this routine noticeably reduces the number of lost shoppers.

Example: how the results change

Imagine an appliance store and the query “quiet air conditioner for the bedroom”. Before the changes, search looks for the word “quiet” in names and finds nothing. After them, the system recognizes the intent, refers to the noise level attribute, selects models up to 24 dB, checks availability and shows five options with an explanation of the difference. It also offers to specify the room size, because the required power depends on it.

Technically not much changed: a normalized noise attribute, a synonym dictionary and a semantic layer appeared. But for the shopper this is the difference between an empty page and a ready solution.

How to measure the result

The main metrics are the share of search sessions with a click on a result, the frequency of zero-result queries, add-to-cart after search, the conversion of search sessions and the revenue per such session. A useful signal is reformulation: if a person changes the query two or three times, search did not understand them.

Numbers alone are not enough. Once a month it is worth manually checking relevance on a set of several dozen typical queries. A high conversion does not justify systematically odd results: shoppers may have found the product despite the search, not thanks to it.

Implementation order

Start not with choosing a platform but with analytics: collect real queries for several months and all zero results. Then clean up attributes and dictionaries and get exact search in order — this alone will bring an uplift without any model.

The third step is semantic search for natural-language queries, the fourth is hybrid ranking and an A/B test against the old version. Clarifying questions and a conversational mode come last: this is the logical bridge to a full AI assistant for an online store. Google Cloud describes such a narrowing scenario as conversational filtering.

Frequently asked questions

Does a small store need AI search?

If the catalog has a few hundred products, a well-configured regular search with a synonym dictionary is usually enough. AI gives the biggest effect in catalogs of several thousand items with many characteristics.

Can AI search show a product that does not exist?

A properly built system will not. Results are formed from the real catalog with an availability check, and the model only helps to understand the query.

How long does implementation take?

Basic setup of hybrid search on a prepared catalog takes a few weeks. The longest part is organizing the data, not connecting the technology.

Will AI search replace filters?

No. Filters remain the main tool for narrowing down, and search helps people reach them. On how not to break SEO in the process, see our article on faceted navigation.

AI search is a product that constantly learns from the store’s data. The GL.ua team can integrate the search engine with the catalog, stock, analytics and an AI assistant, so that the shopper moves from a query to the right product without extra steps. More on the online store development page.