Walmart’s AI Shopping Experiment Flop Leads to Pivot for ChatGPT Integration
Since last November, Walmart has allowed a subset of ChatGPT users to purchase a limited range of products directly within OpenAI’s chat interface, no external redirect required. But sales from the experiment have underwhelmed significantly, a Walmart executive vice president revealed in an exclusive interview with WIRED.
These underwhelming results indicate that a future where chatbots and AI agents fully handle ecommerce shopping could still be years away—if it ever comes to pass at all. Last year, OpenAI laid a major bet on growing its revenue by taking a commission on every sale completed through ChatGPT. The company partnered with Walmart, Etsy, and other retailers to launch an “agentic commerce” tool called Instant Checkout.
Walmart made roughly 200,000 products available for direct purchase in ChatGPT’s chat responses, letting users save their shipping and payment details with OpenAI and check out entirely within the chatbot. For bigger items like televisions, however, shoppers still had to navigate to Walmart’s main website to complete their purchase the traditional online way. According to Daniel Danker, Walmart’s head of product and design, conversion rates—the share of users who actually buy an item ChatGPT shows them—are three times lower for products sold directly through in-chat checkout than for products that require users to click through to Walmart’s site. Put plainly, Instant Checkout has been a failure.
Rather than spend years trying to fix Instant Checkout’s “unsatisfying” user experience, Danker says OpenAI opted to pivot quickly to a new framework Walmart has long preferred. Starting next week, Walmart’s own in-house chatbot, Sparky, will run embedded directly inside ChatGPT—essentially operating as a chatbot within a chatbot. An identical integration will launch in Google’s Gemini chatbot next month.
This new setup fixes what Danker calls the biggest flaw of Instant Checkout: its requirement that users check out for each item individually. “Customers worry that if checkout happens automatically after every single item, they’ll end up getting five separate boxes when they wanted everything shipped together,” Danker explains. “Most shoppers don’t want a split checkout experience where one item is processed separately, even when they already have other items waiting in their Walmart cart.”
Of the products available through the original Instant Checkout program, the top sellers have been vitamins and protein supplements. Danker notes that users new to GLP-1 weight loss drugs often ask ChatGPT for guidance on nutritional needs, and the chatbot frequently recommends increasing nutrient intake—leading those users straight to Walmart’s supplements via Instant Checkout. Other high-performing items were typically expensive enough to avoid extra shipping charges or small-order fees. All together, automotive goods, beauty products, home goods, and hardware tools made up more than half of all Instant Checkout orders.
Under the new embedded Sparky model, users log into their Walmart account the first time they access Sparky through ChatGPT. Their existing shopping carts from Walmart’s website and app will sync automatically with the cart inside the chat integration, a design meant to align with how people actually shop. Most shoppers add items to their cart gradually: they might toss peanut butter on their Walmart app one day, add aluminum foil the next, and throw in a last-minute birthday gift via the website before checking out all at once. “When Sparky is embedded in other platforms, it brings the full Walmart shopping experience to where you already are, instead of forcing you into a completely disconnected broken process,” Danker says.
Walmart has strong incentive to get its ChatGPT integration right. Danker says ChatGPT already brings in roughly twice as many first-time customers to Walmart as traditional search engines. He suspects this is because ChatGPT’s most active users don’t overlap much with Walmart’s typical customer base, and Walmart’s low prices, wide selection, and broad national footprint mean its products get recommended in a huge number of ChatGPT queries.
Sparky was built in-house by Walmart, but it combines open-source generative AI models with proprietary retail-specific models trained on decades of Walmart customer and transaction data. “We can route different types of questions to different models because we’ve seen answer quality varies a lot by use case,” Danker says. “We never lock ourselves into just one model.”
The chatbot was also built to be intentionally flexible from the start, specifically with cross-platform integrations like this in mind. “We can make small tweaks to its look and feel so it feels like a natural part of whatever environment it’s embedded in,” Danker adds.
A Broader Shift in AI Commerce
Walmart’s pivot to Sparky is part of a larger strategic shift at OpenAI, which The Information reported earlier this month—though the outlet did not share the full reasoning behind the change. Danker discussed the shift at a recent Morgan Stanley investor conference, but had not previously released the performance data that prompted the change.
OpenAI spokesperson Taya Christianson says the company is now focusing on improvements that help users research products, while giving retail partners more control over the entire checkout experience. “We appreciate our partners for learning and iterating with us,” Christianson said.
Danker says Walmart already kept many product categories out of Instant Checkout because it knew the single-item checkout model would hurt the user experience in those cases. For example, shoppers buying a new TV almost always need accessories like HDMI cables. On Walmart’s own website, the company can prompt shoppers to add those accessories to their order to avoid a frustrating post-purchase installation experience. With Sparky embedded in chatbots, Walmart will be able to replicate that same helpful nudging.
Retailers were eager to test Instant Checkout because the only other option for serving ChatGPT users at the time was just linking out to their own websites. Walmart now believes the Sparky integration will be even “more seamless” for shoppers, because users can keep chatting and adjust their order without having to re-enter payment and delivery details they already saved to their Walmart account.
Sparky has drawn criticism from users claiming to be Walmart employees on Reddit, and independent third-party testimonials for the chatbot are hard to find on major social media platforms. But Walmart says half of its app users have already interacted with Sparky. While most users turn to the app’s regular search function for everyday staples like milk and bananas, they typically use Sparky to ask about harder-to-find items or get help solving more complex shopping problems. Walmart U.S. CEO David Guggina recently noted that Sparky users spend roughly 35 percent more per order than shoppers who don’t use the chatbot.
Danker acknowledges that Sparky still has flaws: it can be slow, and it generates unhelpful or inaccurate responses often enough that some consumers have written it off as unreliable. He says Walmart’s top priority this year is training Sparky to be more proactive, help it better understand individual shopper preferences, and expand its functionality across more of Walmart’s many business units, including its pharmacy division.
Even as Walmart expands Sparky to third-party chatbot platforms, the company has not blocked—and has no plans to block—other AI agents from shopping on its website, unlike some major competitors. Amazon, for example, recently secured a temporary court order barring AI search company Perplexity from using automated tools that impersonate humans to make purchases on its site. Danker says Walmart wants to support whatever tools customers choose to use, as long as the experience is high quality: meaning no incorrect orders, surprise charges, or unnecessary strain on customer support teams.
“We don’t want to dictate the exact shopping path every customer takes,” he says. “We don’t want to block tools based on speculative or hypothetical concerns.”
When asked how many consumers will ultimately trust AI to handle their full shopping experience, Danker says the hype around fully automated shopping may be overblown. “The idea that all shopping will become fully automated might be a little far-fetched,” he says. “People still get excited about shopping for clothes, items for their home, and gifts for their kids.” Walmart’s goal, he adds, is to keep control in users’ hands—just with Sparky available to help them wherever they choose to shop.
This piece is an edition of Will Knight’s AI Lab newsletter. Read previous editions here.
