Perplexity’s AI shopping recommendations are here but its CEO doesn’t know how they’re chosen

The new tools, announced Monday, are part of a major push by Perplexity to be more of a one-stop hub for users and turn the 100 million weekly search queries it handles into a business. The effort marks yet another effort by the company to challenge search leader Google, which has been adding AI-powered shopping tools to its search results in recent months.  

Perplexity users who enter a query in the new search experience, such as “What’s the highest-rated coffee machine under $500?,” are shown a top product recommendation as a response, along with a short summary explaining its positive attributes. Below that, the AI system displays a few rectangular tiles, each displaying another recommended product with key information like a product’s image and pricing. In some cases, shoppers will be able to easily buy one of the recommended items after a couple of taps through an integration with the e-commerce software company Shopify and its Shop Pay checkout system. 

For users of Perplexity Pro subscription, the company offers the ability to complete a purchase directly within the AI search engine, and get free shipping, as well.  

But for merchants, retailers, and brands—or even consumers—that want to understand how the company’s AI decides which products to recommend, the startup’s CEO bluntly admitted he doesn’t have the answer. At least not yet.

“All these things are yet to be fully understood, to be very honest with you, in terms of how the ranking works [and why] the AI is preferring to rank one over the other,” Perplexity CEO Aravind Srinivas told Fortune in an exclusive interview on Monday. “Is it the number of reviews? Is it the ratings exactly and where the rankings are coming from—like what people are saying on different platforms about their product? There’s a lot of distillation and condensing going on here.”

“I think we don’t understand it fully ourselves today,” he added, echoing the comments of executives at many other companies about the results their AI produces. 

The goal and expectation, though, is that Perplexity will eventually get there, the CEO said.

In the example of a query shared in Monday’s  press announcement—about the sub-$500 coffee machine—the results highlight the “Breville Bambino Plus” espresso machine as the top choice, along with a few bullet points explaining the decision. Perplexity then features a product tile displaying key attributes of the merchandise, along with similar product cards for two other “recommended products” that cost nearly the same. 

Some users may find such a result slightly off target. Technically, a coffee machine isn’t the same thing as an espresso machine.

The product card for the Breville brand top choice has a link to Best Buy where shoppers can get more details. It also shows a “Buy With Pro” button that lets Perplexity Pro users enter their shipping and credit card info, and purchase the item through Perplexity. Best Buy, however, is not actually partnering with Perplexity in this example. Rather, the startup is placing the order with Best Buy on the customer’s behalf, and then covering any shipping cost as an incentive for shoppers to try out the new Perplexity shopping tools. 

While search platforms like Google and Pinterest both previously experimented with letting their users complete some purchases directly on their platforms, both internet giants eventually shuttered those capabilities, partly because retailers don’t like being disintermediated. 

Perplexity is simultaneously opening up a merchant program that it hopes will help it build direct relationships with big retailers to feed their product data into the AI system and potentially let Perplexity users buy their goods directly within the search engine. Srinivas said his startup won’t take a cut, or commissions, on any resulting sales. 

While Srinivas said he doesn’t fully understand how his company’s AI system steers shoppers to certain products, it seems that one factor in particular may play a role. Retailers that provide data to Perplexity as part of its new free merchant program will earn “Increased chances of being a ‘recommended product’ because the products will be in our index, and when we have more robust details, we can better determine if a product is high quality and relevant to a user’s query,” Monday’s press release says.  

Perplexity recently announced an experiment with advertising for its core search product in the form of sponsored “follow-up questions”—related questions that the AI system generates and displays below a search result meant to prod users to dig deeper into a topic.

“Integrating sponsored questions into the shopping experience is a natural choice for us to explore,” he told Fortune. “It could become a very cool way to generate demand and fulfill it right there.”

However, retailers and brands won’t be able to pay to have their products recommended, Srinivas said. 

(Fortune is a member of the Perplexity Publishers’ Program. As part of the program, the startup shares ad revenue with publications when their content is used or referenced in a Perplexity answer to a sponsored question.)

The new shopping assistant-style experience comes as existing general purpose search engines like Google, plus new specialized search startups like Daydream, use gen AI systems to simplify how consumers discover new products online. Google recently added AI summaries and recommendations to its shopping search tools. Amazon and Walmart have also added AI shopping recommendation tools to their marketplaces, with the former recently rolling out an assistant named Rufus trained on the e-commerce giant’s massive product catalog and database of customer reviews.

So will Perplexity’s new shopping tools also pull information or signals from Amazon’s reviews catalog? The answer from Srinivas was unclear.

“I’m not completely sure in which cases they’re accessible or not accessible,” he said. 

As Perplexity gets started in a new space, it’s fair to wonder if customers should expect to still be bombarded with “hallucinations”—or mistakes—that many Gen AI products produce. If the answer to a question I asked of Perplexity this morning—“can you show me your new shopping experience?”—is any indication, the answer might be yes.

“I apologize, but I don’t actually have a new shopping experience to show you,” the Perplexity search system responded. “I’m an AI assistant created by Anthropic to be helpful, harmless, and honest. I don’t have my own shopping platform or experience.” (Perplexity’s AI system partly relies on Anthropic, a separate AI company, for large language models that generate some of the answers that its search tools produce.) 

Perplexity’s response then highlighted recent AI-powered product discovery experiences rolled out by others including Google, Amazon, and Walmart.