When an AI agent goes shopping, how does it decide what to buy?

New York | The New York Times When an AI agent goes shopping, how does it decide what to buy?

That's the big question on every marketing executive's mind as consumers increasingly turn to artificial intelligence for recommendations and new AI agents from Meta and OpenAI offer to help anyone send a chatbot shopping instead.

A new category of startups is vying for space to help brands find the answer. Situated on the border between technology and marketing, their strategies radically move away from traditional advertising.

"When you advertise to humans, you target eyes, emotions, and dopamine; but when you advertise to agents, you appeal to logic — extremely mathematical logic," said Aviv Shamny, co-founder and CEO of Limy, an "AI search platform" backed by venture capital firm Andreessen Horowitz.

Until recently, most websites blocked robots. Some stores, like Amazon, are also blocking certain AI purchasing agents.

Many retailers and brands, however, want to cater to AI. Traffic from AI sources to US retail websites increased 393% in 2026, according to a report from Adobe.

Experts envision a future that is more focused on them. By 2030, "agent-influenced spending" could represent up to 20% of all e-commerce in the United States, or US$385 billion (R$2 trillion), according to research by Morgan Stanley.

“Increasingly, you’ll just click ‘buy’ in the AI ​​chat interface — you’ll never get to a brand’s website,” said Jeff Blackman, managing director of Barbarian Group, a marketing agency that’s betting on AI. "The risk of not being prepared is simply not showing up."

Agent trading

Tyler Ackerman, CEO of RTA Store, a family-run online cabinet retailer, decided last year to invest in his company's AI visibility after team members realized it was no longer appearing in recommendations from major language models (LLMs).

“If we don’t focus on ‘How do we make these agents read our products easily, quickly and securely?’ we will be left behind,” Ackerman said.

RTA Store paid US$15,000 (R$78,400) to New Generation, a San Francisco startup that promises to transform companies' "static product catalogs into structured data and generative interfaces that AI can understand."

New Generation made a series of changes to Ackerman's website — which, according to him, paid off — from changes to the basic code to the inclusion of an AI widget that answers questions about products, whether asked by humans or not.

The company has also added large volumes of content outlining the details of each product offering — something often recommended by AI marketing experts.

Agents "read" more than human consumers and prefer "very factual, straightforward, stat-filled content," said Caelean Barnes, whose two-year-old startup Gauge also focuses on companies that want to make their products more noticeable to AI agents.

“You need to make sure the facts about your business are readily available to agents across all the different platforms they might search on.”

In practice, this means producing more: more product descriptions, more blog posts and more social media activity.

Websites must also allow agents to navigate without difficulties. In partnership with Visa, New Generation assigns a score from 1 to 100 to the "degree of preparedness to interact with agents" on a brand's website, simulating navigation attempts made by these agents.

The score helps brands understand what agents can do and where they encounter obstacles.

"Let's say you're a bank and agents are asking you about your credit card interest rates," said Jonathan Arena, CEO of New Generation.

"So a consumer says, 'I want to apply for your premium credit card,' but the transaction fails because their agent couldn't access the application form. You've just lost a customer."

Technical tricks

Some strategies are extremely technical. Shamny said Limy focuses on the way AI systems represent language through a series of numbers called vectors.

A command like "best running shoes for a marathon" is converted to a numeric representation of its meaning. A website's content can also be represented numerically, and Limy helps "identify vector spaces of interest to agents." (The startup also identifies brand MVPs — the most valuable commands — that result in the most purchases.)

Not everyone is enchanted by this type of artifice. Models are changing “faster than our ability to try to manipulate them,” said Toby Coulthard, founder of Jacquard, an AI-powered marketing platform.

"Maybe these recommendations will be valid for the next two months, but not beyond that. We need to go back to the fundamental principles of marketing."

Akshat Trivedi, industry analyst at FTI Consulting, sees parallels with both the transition to mobile commerce and the race for search engine optimization.

Whoever arrived first made big gains. But in the end, when everyone joined in, the market was balanced.

The difference now, according to him, is the speed with which companies are rushing to get ahead of what appears to be a new marketing paradigm.

“Soon, we may see products launched AI-first or brands named so that AI identifies them more than others, boosting their ability to be found,” Trivedi said.

This likely means that a greater portion of marketing budgets will disappear in this new agentic frontier.

Ackerman, from RTA Store, imagines that he will soon be able to spend up to US$12,000 (R$62,700) per month to attract the attention of his robot clientele.

"As these systems become more intelligent, we will have no choice but to spend a little more," he said.