Marketers have spent decades studying pricing, color psychology, social proof and other elements that influence people’s buying decisions. But so far we know little about what affects the decision-making processes of AI neural networks, said Kartik Hosanagar, Faculty Co-Director of the Wharton Human-AI Research Center, at the CommerceNext Growth Show in New York City last month.

He believes organizations need to quickly gain that knowledge, because AI isn’t just a new technology or tool; it’s a new customer.

In his keynote session, “How AI Search and Agentic Commerce Are Reshaping Retail,” Hosanagar predicted that AI will alter how people shop in two key ways. The first, AI-assisted mode, is already becoming commonplace: “You ask a question, AI takes over all your early-funnel research work and presents it to you, and you just go through the smaller, shorter consideration set and make decisions.”

In the second, shopping-agent mode, the consumer gives an AI agent parameters and specifications, and the agent not only does the research but also makes the decision — sometimes but not always with the human’s approval.

“There are going to be many things that the shopper has not specified, and the agent is going to make decisions around those unspecified criteria,” Hosanagar cautioned.

Three Dangerous Retail Misconceptions

To better understand how AI agents will be making purchasing decisions, Hosanagar said retailers need to know the truth behind three common misconceptions:

• A human will always be involved in the decision. While that may be true today, as shopping-agent mode, subscriptions and other types of automated purchases become more mainstream, people will take less and less of an active role. Brands need to determine how much of an effect this will have in their market category.
• So long as AI crawlers can read your content, you’ll be included in AI recommendations. Visibility is important, of course, but the content needs to satisfy the decision criteria set by both the human consumers and the AI agents. “That’s going to be a big area that your teams will have to think about: How do these AI systems make these recommendations?” Hosanagar reiterated. “Why do they make it? And how do we understand that better? How do we influence these systems?”
• AI is just another channel. Retailers should learn from the mistakes of movie studios, Hosanagar said. The studios had originally perceived Netflix and other streaming platforms as simply a new distribution channel. Instead these platforms restructured the entire industry, including how content was produced and who controlled it. For retailers, “the question is, how might agentic buying or even an AI-mediated commerce shift power dynamics in these industries as well?” Hosanagar explained.

Efficiency vs. Meaning

Brands can approach the AI-disrupted retail marketplace in one of two ways, Hosanagar said. The first is what he dubbed an efficiency play: optimizing your catalog and content with AI agents front of mind, ensuring that LLMs not only understand what you offer but also are persuaded to select it. This requires multiple teams, including marketing and tech, to effectively pull off.

The other option is to focus on human consumers by creating a unique, compelling experience for them, which he called the meaning play. “In some ways, the product is the shopping experience itself,” Hosanagar noted. “It requires a lot of craft, it requires patience and it requires a completely different way of approaching this.”

Hosanagar warned against taking a middle approach. Trying to have it both ways is difficult, if not impossible, to pull off, and it muddies what the brand stands for.

While deciding which approach to pursue, brands also need to assess how they and their competitors currently show up on AI search and what customers are searching for. Many companies are already tracking their presence and performance on various LLMs but don’t know how to use the data.

Hosanagar suggested creating, over the next 12 months, simulation sandboxes to test a variety of content strategies across multiple LLMs to measure the type, quantity and quality of traffic each draws, along with where the visitors are coming from and the pages they’re visiting.

To prove his point, Hosanagar cited wallet and luggage brand Ridge, which tested 20 content ideas in a simulation sandbox, then pushed out the winners. Within three weeks, the brand went from not being recommended by LLMs as a gift for men to appearing as a response to nearly half of all such queries.