Retailers risk losing control as AI turns search into conversation

From keywords to intent

Paulus Nagis, co-founder and chief growth officer of LupaSearch, a Kaunas, Lithuania-based company that builds AI-powered search and product discovery systems, argues that the starting point is not the chatbot. It is the product data behind it.

People have spent decades learning to compress their needs into phrases a search engine can match. “Men’s running shoes” or “cordless drill” leaves out context in exchange for a useful results page. That shortcut gets less useful when shoppers ask what they need to hang shelves on a concrete wall. They may need a drill, a masonry bit, wall plugs and screws without knowing the names of any of them.

The system has to infer the job, not just match words in a query to product descriptions. That makes search look less like a catalogue index and more like a sales consultation.

The behavior is already shifting. Google said its AI Mode answers feature had more than a billion monthly users in 2026, with queries doubling or more each quarter since launch. Salesforce’s 2025 Connected Shoppers study found that 39% of consumers had used AI to find products; among Gen Z, the share was more than half.

Salesforce director of industry insights Kayla Schwartz framed the content challenge this way: the most effective material answers the problem a shopper is trying to solve, rather than simply repeating their search terms.

One conversation, several jobs

Nagis expects product search, product selection and customer service to converge. A shopper might ask whether an item is available nearby, whether there is a suitable alternative, then how to return it. Today, answering those questions can mean jumping among search results, product pages, reviews, FAQs and support.

Google is moving toward the same model. At the National Retail Federation conference in January, CEO Sundar Pichai discussed AI’s role in product discovery, shopping and customer service. Google is also working with retailers on shopping agents that can answer detailed questions and offer personalized recommendations.

Amazon has put a version of this approach into use. The company said more than 300 million customers used its shopping assistant Rufus in 2025, generating nearly $12 billion in incremental annual sales. Rufus answers product questions; Buy For Me lets customers purchase some items from other online stores.

But sending every retail query to a large language model is not an obvious business case. A small store may handle a modest volume; a large international retailer may receive millions of queries against a catalogue of hundreds of thousands or millions of products. Nagis argues that the cost of inference makes using a large model for every request uneconomical.

LupaSearch takes a selective approach:

Retail search can combine several technologies: semantic vectors for synonyms, recommendation models to predict what else a shopper might need, image recognition to extract product attributes missing from descriptions, and a large language model when cheaper methods are not enough.

Algolia has also chosen a hybrid architecture. Its commerce technology combines keyword and semantic search with conversational systems that use retailer data. In Agent Studio, cost controls sit alongside agent capabilities—a reminder that the impressive demo has an operating budget behind it.

The catalogue becomes the storefront

Whichever approach a retailer chooses, the quality and structure of its product data set the limits. A query such as “waterproof commuter jacket under $200” sounds simple, but answering it depends on whether the catalogue distinguishes waterproof from water-resistant, normalizes sizes, tracks colors in stock and separates winter coats from light rain jackets.

Shopify’s Catalog API structures product information from millions of merchants so software can query it. Its Universal Commerce Protocol, developed with Google, standardizes how agents interact with merchants, from finding products to completing transactions. In 2026, Shopify opened more of its agentic commerce infrastructure to developers, including through a public Model Context Protocol endpoint.

In September, Algolia released its own MCP server, connecting systems such as ChatGPT, Claude and Gemini to current product information, including stock, prices and promotion rules.

Availability matters as much as visibility. A retailer may want products to appear in search even when they are sold out. That can make sense in a world where merchants control search optimization, but it is less useful to an agent trying to match a shopper with something they can buy now.

Imagine looking for a computer with a particular amount of memory and repeatedly seeing models that are unavailable. Today, I can check Amazon, Micro Center and Newegg, re-enter roughly the same query and compare results across tabs. An AI assistant could instead check participating retailers’ catalogues and identify who has the computer, at what price and on what delivery terms.

In that world, a retailer’s rank in search matters less than the quality and availability of the data the agent can use.

Who controls the answer?

The shift gives AI platforms influence over which brands appear, which sellers get considered and which products make the shortlist. Retailers know that customer intent is valuable, and advertisers want to pay search and AI companies to put offers in front of shoppers.

Google is testing ads in AI Mode, where Gemini selects sponsored products and explanations relevant to a shopper’s question. OpenAI has launched ads in ChatGPT and is developing a self-serve advertising platform that charges by the click.

That puts the answer and the commercial incentive in the same interface. An assistant can weigh more context than a conventional keyword ad, but the buyer still needs to know whether it is recommending the best fit or a sponsored product. I think that distinction will matter more as the assistant takes on the roles of search engine, salesperson and advertising channel at once.

Nagis does not expect retailer websites to disappear; he points to Google Shopping as an example of aggregators gaining strength while stores remain part of the transaction. His argument is that retailers need infrastructure to serve customers wherever the query starts.

LupaSearch is not trying to build the agent a shopper talks to. Nagis says its aim is to make a retailer’s catalogue understandable to whichever agent emerges—whether that is LupaSearch’s own search interface, Google, a ChatGPT chat or a conversation in AI Mode.

What I’d want to know is how retailers will retain a meaningful say in the recommendations and sponsored placements agents make from their catalogues. The first point of contact may keep changing; the product data those systems trust will determine which retailers remain in the conversation.