For decades, shoppers learned that the way to find the products online you’re most interested in is to type two or three keywords into a search box, scan a grid of products, adjust the filters, and check the products or try another query.

Generative AI is changing that behavior. Consumers can increasingly describe a problem, ask a question and expect the ecommerce platform to deliver up the exact product and configurations they want. For online retailers, search is becoming less like a database lookup and more like a digital sales associate.

Ecommerce shops and retailers have reason to pay attention. A search query is one of the few moments when customers tell a store, in their own words, exactly what they want, something called purchase intent. While much of that information has historically not been put to good use, AI can connect those queries to recommendations, inventory, customer service and advertising.

The stakes rise further as shopping migrates into ChatGPT, Google AI Mode, Gemini and other AI interfaces. More importantly, these AI tools can do so without the interaction or visibility of the ecommerce shop.

The change is already showing up in traffic. Adobe reported that generative AI services drove almost a 700% year over year increase in traffic to U.S. retail sites during the 2025 holiday season. Shopify says AI driven traffic to its merchants grew eightfold year over year in the first quarter of 2026, with orders attributed to AI powered searches rising nearly thirteenfold.

Then comes a thornier problem: who owns the shopper? If consumers increasingly begin with ChatGPT, Gemini or another AI assistant, the retailer may no longer control the first interaction. The store could become a data source behind somebody else’s interface.

Who owns the customer relationship when search becomes a conversation? Does the retailer still control product discovery? Does an AI platform become the new storefront? Or does commerce split into layers, with AI agents controlling the conversation and retailers competing to make their catalogs intelligible to those agents?

Moving From Keywords To Intent

Paulius Nagys, co-founder and head of growth at Kaunas, Lithuania-based AI search and product discovery company LupaSearch, thinks the answer starts with the product data underneath it, much less glamorous than the chatbot.

We have spent decades of web experience learning how to communicate with machines to get what we want.

Want running shoes? Type “men’s running shoes.” Need a drill? Search “cordless drill.” To narrow down and get to what we want to buy, we’ve learned to strip away context and reduce an idea to the handful of words most likely to produce a useful result.

Nagys thinks that approach is now breaking.

“People are moving from typing keywords to just asking a full question and expecting an answer,” he said in a one-on-one interview. “Not a page of results to sort through themselves. So the search box has to actually answer now, not just index.”

Google is already changing behavior. The company says its AI Mode response box has surpassed one billion monthly users in 2026, and its queries have more than doubled every quarter since launch. Google’s redesigned Search experience gives people more room to describe complex requests and keep asking follow-up questions.

Salesforce has seen the change in shopping behavior too. Its 2025 Connected Shoppers research found that 39% of consumers had used AI for product discovery, rising above half among Gen Z respondents.

Caila Schwartz, Salesforce’s director of industry insights, framed the challenge this way: “The best-performing content anticipates what a shopper is trying to solve, not just what they’re searching for.”

The difference is pretty important in terms of connecting the shopper to what they want. In the past, retailers depended on the shoppers knowing what product will meet their needs. For example, a shopper typing “cordless drill” made life easy for the retailer. But what if the question is “What do I need to hang shelves on a concrete wall?”

The customer may need a drill, masonry bit, anchors and screws without mentioning any of them. Finding the right answer requires more than matching words in a query with words on a product page. The system has to infer the job.

That starts to look less like fulfilling orders and more like consultative selling.

The Search Box Is Starting To Look Like A Salesperson

Nagys expects ecommerce search, product discovery and customer service to begin bleeding into one another.

A shopper might ask about a product, then whether it is available nearby, then whether another alternative would work, then how returns work. Today those requests can send someone bouncing between search, product pages, product reviews, FAQs and customer support. In the past that has meant lots of open tabs, lots of searches and cutting and pasting while trying to figure it all out.

Nagys expects that behavior to change. In our interview, he described a future ecommerce experience built around questions and answers rather than isolated searches.

Google is heading down a related path. At the National Retail Federation conference in January, CEO Sundar Pichai described AI moving through product discovery, purchasing and customer experience. Google has been working with retailers on shopping agents that can answer detailed questions and make personalized recommendations.

Amazon is already experimenting in the wild. The company said Rufus, its AI shopping assistant, was used by more than 300 million customers in 2025 and contributed nearly $12 billion in incremental annualized sales. Rufus answers questions about products, and Amazon’s Buy For Me capability can purchase certain products from other online stores.

Not Every Search Needs An LLM

While AI is part of the equation and owning the conversational experience might be critically important, the catch is that someone has to pay the inference bill. Running conversational AI at retail scale can get expensive quickly.

A small merchant may process a manageable number of searches. A global retailer can handle millions of queries against catalogs containing hundreds of thousands or millions of SKUs. Nagys argues that sending every one of those requests directly through a large language model makes little economic sense.

Nagys says, “Not every AI problem needs an LLM solution.”

His company’s LupaSearch solution takes a more selective approach, he said. Conventional retrieval handles queries it can resolve efficiently. Heavier model calls can be reserved for cases where those methods struggle, including unusual natural language requests or searches that produce no useful results. Those interactions can then feed improvements back into the search system.

From this perspective, there isn’t one AI sitting behind modern product search. A retailer might use semantic vectors to understand synonyms, recommendation models to predict what someone could want next, image recognition to extract attributes missing from product descriptions and an LLM when cheaper methods fail.

And there really is no need to call it “AI search” if customers have questions and retailers want to provide answers. The technology obscures the questions merchants actually need answered. Does it work? Is it fast? What does each query cost?

Nagys has little patience for the industry’s loose vocabulary around AI. In our interview, he compared the current excitement to the dot-com era, when attaching the latest technology to a business could become more important than explaining what the technology actually did.

Algolia has landed on a hybrid architecture too. Its commerce technology combines keyword and semantic retrieval with conversational systems grounded in merchant data. Its Agent Studio product puts cost controls next to AI agent capabilities, a reminder that impressive demos eventually meet operating budgets.

Regardless of the technology approach, all of this is based on having good quality data. Bad data or bad connections make even the most robust catalog impenetrable for conversational assistants.

For example, if you ask an AI shopping assistant for a waterproof commuter jacket under $200, that might sound simple. But behind that request sits an ugly data problem.

Does the retailer reliably record waterproofing? Is “water resistant” treated differently? Are sizes normalized? Does the system know which colors remain in stock? Can it distinguish a winter shell from a light rain jacket? Can another AI system access any of that information?

“Is your catalog well prepared for this?” Nagys asked during our conversation. He described structured product data as the foundation on which more sophisticated AI functions can be built.

Shopify is attacking that problem at an enormous scale. Its Catalog API turns product information from millions of merchants into structured data that software can query. Shopify’s Universal Commerce Protocol, developed with Google, gives agents a standardized method for interacting with merchants through discovery and transactions. In 2026, Shopify opened more of that agentic commerce infrastructure to developers, including access through a public Model Context Protocol endpoint.

Algolia released its own MCP server in September, designed to connect systems such as ChatGPT, Claude and Gemini with live product information, including inventory, pricing and merchandising rules.

The issue with inventory is also important when companies are using their SKUs to get search visibility, even if the items are not in stock. This might work for an SEO-controlled world but it is much more frustrating when agents are trying to connect buyers with real intent to products they can purchase today.

Imagine that I want a computer with a particular amount of memory and it keeps appearing out of stock. Today I might search Amazon, then Micro Center, then Newegg, opening tabs and repeating roughly the same query.

A different model would let my AI assistant query participating catalogs for me. I can describe the machine once and let the agent find who actually has it, at what price and under what delivery terms. In that world, the retailer’s search visibilty matters less. The quality and accessibility of the data underneath it matters a lot more.

This is where retailers could lose control.

A customer can tell an AI assistant, such as a Google AI Mode conversation or ChatGPT chat to “Find me a waterproof commuter jacket under $200 that will arrive by Friday.”

The assistant will decide which brands appear, which retailers get queried and which products make the final cut and are surfaced as the best option that is available immediately within the customer’s budget. The customer never needs to visit ten websites or click through unavailable options.

Complicating this shift is advertising. Retailers know that search intent is highly valuable information. Advertisers want to tap into this by paying AI and search companies to steer transactions their way.

The biggest AI platforms are moving in that direction. Google is testing ads in AI Mode that use Gemini to pair sponsored products with explanations tailored to a shopper’s question. OpenAI has gone further, launching ads in ChatGPT and building a self service advertising platform with cost per click bidding.

This brings into question what should the customer see in responses? The best match or a sponsored product? AI can make that decision with far more context than traditional keyword advertising ever had. When an AI assistant becomes part search engine, part salesperson and part advertising channel, the line between answering a shopper’s question and monetizing the shopper’s intent becomes critical.

Nagys does not think that means retailer websites simply disappear. He points to Google Shopping as a useful precedent. Aggregators became powerful, yet merchants remained part of the transaction.

His bet is that retailers need infrastructure capable of serving the shopper wherever the request originates.

“We’re not trying to be the agent someone talks to,” Nagys said of LupaSearch. “We’re the layer that makes a retailer’s catalog actually readable by whatever agent shows up.”

That could be LupaSearch’s own search interface. It could be Google. It could be a customer’s ChatGPT session or an AI Mode interaction. The front door becomes less predictable, but regardless, ecommerce shops and retailers need to be ready for the new reality.