Walmart’s ChatGPT Shopping Flop Leads to Pivot Towards Embedded Chatbots

Walmart’s pioneering venture into direct commerce within OpenAI’s ChatGPT interface, launched in November, has yielded disappointing sales results, prompting a strategic pivot. The initiative, which allowed a limited selection of products to be ordered without users ever leaving the chatbot, has revealed significant challenges for the nascent field of “agentic commerce,” suggesting that a future dominated by AI agents handling e-commerce transactions remains a distant prospect.

Last year, OpenAI made a notable bet on generating revenue through commissions on purchases facilitated by ChatGPT. This led to partnerships with major retailers like Walmart and Etsy, culminating in a feature called "Instant Checkout." The concept was designed to streamline the shopping experience, allowing consumers to interact with the chatbot, discover products, and complete purchases seamlessly within the AI environment.

Walmart made approximately 200,000 products available for direct purchase through ChatGPT. This meant users could provide their shipping and payment details directly to OpenAI and finalize an order without navigating to Walmart’s own website. However, for larger or more complex items, such as televisions, shoppers were still directed to Walmart’s website to complete their purchase through traditional online methods.

The performance metrics for Instant Checkout have been stark. Daniel Danker, Walmart’s executive vice president overseeing design and product, exclusively revealed to WIRED that conversion rates – the percentage of users who followed through with a purchase after being shown an item by ChatGPT – were three times lower for products sold directly inside the chatbot compared to those that required clicking out to Walmart’s website. In plain terms, Instant Checkout proved to be a significant commercial setback for the retail giant.

Danker acknowledged that fixing the "unsatisfying" consumer experience of Instant Checkout could have taken years of iterative development. He commended OpenAI for its agility in choosing to quickly pivot to a new system that Walmart had long advocated for. This new approach will see Walmart’s proprietary chatbot, "Sparky," begin operating within ChatGPT as early as next week, effectively creating a "chatbot inside a chatbot." A similar integration is slated for Google’s Gemini chatbot next month.

The primary flaw identified by Danker in the Instant Checkout model was its inherent limitation to single-item purchases. "They fear that when checkout happens automatically after every single item that they’re going to receive five boxes when they actually just want it all in one," Danker explained. He further elaborated on the consumer preference for consolidated transactions: "They generally don’t want to split the checkout experience, where it buys the one item, even though they had other items in their Walmart cart already." This fragmented purchasing experience, lacking the ability to bundle multiple items into a single order, proved to be a major deterrent for users.

Despite the overall disappointment, certain product categories did exhibit some traction within the Instant Checkout framework. Top sellers included vitamin and protein supplements. Danker noted that users exploring information about GLP-1 weight-loss drugs often received advice to increase nutrient intake from ChatGPT, leading to supplement purchases. Other items that performed relatively well tended to be those with a high enough price point to circumvent additional shipping or small-basket fees. Collectively, the automotive, beauty, home management, hardware, and tools categories accounted for over half of all Instant Checkout orders, suggesting that specific use cases or product types might be better suited for this direct-purchase model.

The new Sparky integration aims to resolve the critical single-item checkout issue and better align with actual consumer shopping habits. When users first encounter Sparky within ChatGPT, they will log into their Walmart account. This action will sync their existing shopping basket from Walmart’s website or app with their activity within ChatGPT. This synchronization is designed to reflect how people genuinely shop – adding peanut butter one day via the app, foil the next, and a last-minute birthday gift on the website before finally checking out a consolidated order. "When Sparky travels, it’s the Walmart store meeting you where you are, instead of a completely broken experience," Danker emphasized, highlighting the goal of a consistent and holistic shopping journey.

Walmart’s motivation for refining its presence within ChatGPT is substantial. According to Danker, the chatbot is currently bringing in new customers at roughly twice the rate of traditional search engines. He speculates that this trend is partly due to the demographic of ChatGPT’s "power users," who may not represent typical Walmart customers. Nevertheless, Walmart’s competitive pricing, extensive product selection, and vast geographic footprint mean its products frequently appear in ChatGPT responses, making it a critical channel for customer acquisition and engagement.

Sparky itself is an internal development by Walmart. Danker elaborated that the chatbot leverages a combination of open-source generative AI models alongside proprietary, retail-specific models. These internal models have been meticulously trained on decades of Walmart’s extensive sales and customer data. This hybrid architecture allows for sophisticated question routing: "We’re able to route certain questions to one model and certain questions to another because we find that the quality of answers differs," Danker explained. "It’s never stuck in any one," ensuring optimal response quality. Furthermore, Sparky has been designed with inherent flexibility, allowing for "slight tweaks to the look and feel" to ensure it integrates naturally and seamlessly into various external environments, such as ChatGPT and Gemini.

This strategic shift by Walmart is part of a broader re-evaluation within OpenAI regarding its e-commerce strategy. The Information previously reported earlier this month that OpenAI is increasingly focusing on having checkouts occur within embedded applications rather than directly within the chatbot interface itself. OpenAI spokesperson Taya Christianson affirmed this pivot, stating that the company intends to concentrate on enhancing user research capabilities for products while simultaneously granting merchants greater control over the checkout process. "We appreciate our partners for learning with us," she added, acknowledging the collaborative nature of this experimental phase.

Danker also noted that Walmart had proactively excluded certain products from the initial Instant Checkout experience because it understood that "the single-item checkout experience is detrimental" in specific scenarios. For example, when a customer purchases a television, they often require accompanying accessories like HDMI cables or wall mounts. On its own website, Walmart can effectively prompt shoppers to consider bundled items, thereby preventing a frustrating installation experience later on. Through the new Sparky integration, Walmart anticipates being able to replicate and enhance these bundled recommendations within the chatbot environment, providing a more complete and satisfying customer experience.

Initially, retailers were eager to participate in Instant Checkout because it offered a more integrated alternative to simply linking out to their websites, which was the prevalent method for serving ChatGPT users at the time. Walmart believes the new Sparky experience will be even "more seamless." Users will be able to maintain continuous dialogue with the chatbot, refine their orders, and proceed to checkout without the inconvenience of re-entering payment and delivery information, as this data will be securely saved and accessible via their Walmart account.

While Sparky has faced some criticism, particularly from individuals claiming to be Walmart employees on platforms like Reddit, and public testimonials on social media remain scarce, its performance within Walmart’s own ecosystem is noteworthy. According to the company, half of Walmart app users have engaged with Sparky. Interestingly, while the app is typically used for searching for common staples like milk and bananas, users often turn to Sparky for more unusual items or to find solutions to complex problems. Walmart US CEO David Guggina recently highlighted Sparky’s impact, stating that users who interact with the chatbot spend approximately 35 percent more per order than other shoppers.

Danker candidly admitted that Sparky, in its current iteration, can be slow and occasionally generates weak or unhelpful responses, which might lead some consumers to dismiss it as unreliable. However, he outlined Walmart’s strategic priorities for Sparky this year: enhancing its proactive capabilities, enabling it to learn more about individual shopper preferences, and expanding its helpfulness across a wider array of Walmart’s departments, including the pharmacy.

Despite its efforts to embed Sparky across various platforms, Walmart has no intentions of blocking other AI agents from shopping on its website, nor does it plan to do so in the future. This stance contrasts sharply with Amazon’s recent actions, which successfully secured a temporary court order preventing Perplexity’s automated technology from impersonating a human to make purchases. Danker articulated Walmart’s philosophy: the company aims to support whatever tools customers choose to use, provided the experience remains positive. This means avoiding erroneous orders, unexpected charges, or an excessive need for customer service intervention. "We don’t want to be prescriptive of the exact journey that every customer is going to take," he affirmed. "We don’t want to block things on a speculative or hypothetical concern."

Reflecting on the broader implications for AI in retail, Danker offered a tempered perspective on the extent to which consumers will ultimately trust AI with their shopping decisions. "This idea that it will all become automated might be a little bit far-fetched," he speculated, acknowledging the enduring human element of retail. "People do get excited about shopping for clothes, for their home, for their children." Walmart’s overarching goal is to maintain user control throughout the shopping process, now with Sparky serving as an accessible and integrated assistant in more places than ever before.

This report is an edition of Will Knight’s AI Lab newsletter.