Walmart's OpenAI Checkout Converted 3x Worse Than Its Own Site

Walmart, the United States retail giant, is not walking away from AI shopping. It is walking away from the idea that OpenAI, the company behind ChatGPT, should own the last click. After a high-profile pilot of native checkout inside ChatGPT, Walmart is shutting down that flow and replacing it with Walmart-controlled experiences built around its Sparky assistant. That is less a referendum on "agentic commerce" as a category than a pretty direct verdict on one specific design choice: asking shoppers to complete single-item purchases inside a chatbot instead of moving through a retailer's own cart and checkout stack.

The strongest data point came from WIRED's original reporting, which said Walmart saw conversion rates for OpenAI's Instant Checkout run roughly three times lower than click-out transactions on Walmart.com, citing comments from Daniel Danker, Walmart's executive vice president of AI acceleration, product, and design. In a Morgan Stanley technology, media and telecom conference transcript, Danker explained why: native checkout inside the model was a "very temporary moment in time," and the experience broke down because customers were effectively pushed into fragmented, one-item purchases instead of building baskets the way they do on Walmart's own site.

That failure mode was visible in the product from the start, if you read the launch material with your eyebrows properly engaged. In Walmart's October 2025 announcement of its OpenAI partnership, the company sold the arrangement as an "AI-first" shopping future built around personalized recommendations and seamless checkout. But OpenAI's own post introducing "Buy It in ChatGPT" said Instant Checkout supported only single-item purchases at launch, with multi-item carts left for later. Retailers live on basket-building, merchandising, substitutions, promotions, repeat behavior, and all the messy context of real shopping. A one-item chat checkout was always going to look elegant in a demo and slightly ridiculous in a weekly household run.

What Walmart appears to have decided is that the model can keep the top of the funnel, but the retailer needs to control the transaction. In the same Morgan Stanley transcript, Danker said Sparky would travel into ChatGPT and Google's Gemini, the generative AI platform from Alphabet, as soon as next month. That lines up with OpenAI's recent post on Apps in ChatGPT, which lays out a platform model where merchants can run richer, controlled experiences inside ChatGPT rather than hand the entire purchase flow over to OpenAI's default interface.

That distinction matters. If discovery migrates to model platforms while checkout, customer identity, and merchandising logic stay with the merchant, the winners in AI commerce may look less like all-purpose shopping agents and more like retailers that learn how to meet customers inside those interfaces without surrendering the economics. CNBC reported on March 20 that OpenAI was already shifting away from native checkout toward retailer-controlled experiences, while merchants and analysts raised questions about onboarding friction and whether the system actually improved commerce enough to justify the extra plumbing.

That is also why this episode should cool some of the louder rhetoric around agentic shopping without killing the category. Walmart is still investing in AI commerce. OpenAI is still building commerce surfaces. The retreat is narrower and more revealing than that: the experiment suggests language models may be good at helping people find things, compare options, and narrow intent, but they are not automatically good at reproducing the dense operational logic of a mature retail checkout. The last mile turns out to be where all the hard stuff lives.

There is a broader platform story here too. The Information first reported that OpenAI was scaling back parts of its shopping push, though I could not independently access that report in this session because of the site's paywall. Still, taken together with WIRED's reporting, the Walmart transcript, and OpenAI's own product posts, the signal is pretty clear: retailers will happily use frontier AI for discovery, but they are much less eager to outsource the checkout relationship when the numbers get worse.

The thing to watch next is whether other merchants copy Walmart's conclusion. If they do, the first real shape of agentic commerce at scale may be disappointingly un-magical and therefore much more plausible: chat interfaces for search and intent capture, retailer-owned apps and assistants for execution, and checkout pulled back to the places where conversion, cart logic, and customer data already work. That is not the end of AI shopping. It is the end of pretending the cart was optional.