Opinion: ChatGPT Is a Bad Personal Shopper, But It’s Learning

As OpenAI seeks more ways to make money, ChatGPT’s potential as a shopping assistant is one promising idea: You tell the bot what you’re thinking about buying, and it will help you make comparisons, sort prices and answer intricate product questions. With any luck, those answers might even be right.

Crucially, the assumption is that many people will prefer to shop online this way rather than use their current methods, which would allow OpenAI to grab a meaningful slice of e-commerce spending, projected to reach $2.9 trillion in the US by 2030. At least one consumer is already convinced. “I mostly buy stuff from ChatGPT now,” OpenAI Chief Executive Officer Sam Altman said last month.

But a revealing study from two German academics suggests Large Language Models, and ChatGPT specifically, have a lot of catching up to do on the key performance metrics that retailers care about the most. They found that shoppers who arrive at an e-commerce site through ChatGPT were less likely to buy something than those who arrive by almost any other online journey, such as clicking a search result or visiting an online store directly. A shopper who arrived from an affiliate — i.e. a review recommending and linking to a product in return for a commission — had as much as an 86% better likelihood of a purchase than one who arrived from ChatGPT.

“These indicators make it seem unlikely that organic LLM will ‘kill Google’ or otherwise displace traditional channels in the near term,” wrote Maximilian Kaiser and Christian Schulze, from the University of Hamburg and Frankfurt School of Finance and Management, respectively.

But the study makes clear the trajectory is heading in AI’s favor. Conversion rates and revenue per session improved over the time of the study — August 2024 to July 2025. ChatGPT’s bounce rates — when visitors to a site leave without doing anything else — were better when compared with direct visits, email marketing, referral links, affiliate marketing and paid social ads. This suggests that what ChatGPT served up to users was at least in some way relevant and useful, even if they didn’t necessarily end up buying anything. And the study points out it did not capture the upstream effects of AI, such as a customer using AI to research a product, making a decision, but then looking it up on Google.

ChatGPT Trails Traditional Sales Channels but Is Improving

Most important, the study notes that ChatGPT’s shopping capabilities are still in their infancy. Traffic volume to e-commerce sites from organic chatbot conversations is around 200 times smaller than from organic Google searches, the study authors wrote, meaning e-commerce sites had not yet bothered to reorient their site designs and processes to better cater to chatbot-sent visitors.

That’s changing: Since the period of the research, OpenAI has entered partnerships with Etsy, Shopify and Walmart to streamline ChatGPT-induced purchases. OpenAI’s Instant Checkout, unveiled last month, has been offered to merchants that want to make it possible for ChatGPT users to buy items and arrange shipping “without ever leaving the chat.” OpenAI keeps a processing fee, it said. It will face competition from Amazon.com Inc., which is experimenting with its AI bot, Rufus, which has so far been underwhelming.

On a broader level, consumers still finding their feet with AI should, in time, become more comfortable with both how the technology works and where it sends them. “Effectively the value of the traffic from ChatGPT is improving as people learn to trust its recommendations,” observed e-commerce analyst Juozas Kaziukėnas.

So yes, as it stands, ChatGPT is something of a lousy shopping buddy. That’s reflected in this study. In time, though, we can expect ChatGPT-assisted shopping to be more personalized, with a deeper knowledge of your tastes, existing wardrobe, measurements … and who knows what else. As the AI adage goes, this is the worst it will ever be.