When the world’s largest retailer hands you the numbers on a failed experiment, you pay attention.

In an exclusive with WIRED, Walmart executive vice president Daniel Danker revealed that conversion rates for products sold through OpenAI’s Instant Checkout inside ChatGPT were three times lower than products that required users to click out to Walmart’s own website. Around 200,000 products were made available directly in chat responses. Users browsed. They asked questions. They researched. They did not buy.

Danker called the experience “unsatisfying.” OpenAI could have spent years trying to fix it. Instead, both companies chose to scrap the model entirely. Starting next week, Walmart’s own chatbot, Sparky, will operate directly inside ChatGPT, a chatbot embedded within a chatbot. A similar integration with Google’s Gemini follows next month.

The core problem was deceptively simple. Instant Checkout forced single-item purchases. Every product recommended by ChatGPT triggered its own individual transaction, its own shipping calculation, its own delivery.

Consumers rejected it intuitively. As Danker told WIRED, shoppers feared that interacting with five product recommendations would result in five separate boxes arriving at their door when they wanted everything in one order. They did not want a fragmented checkout experience, especially when they already had items sitting in their Walmart cart from earlier browsing sessions.

The data on what did sell is revealing. Top-performing categories through Instant Checkout included vitamin and protein supplements, automotive parts, beauty products, hardware, and tools. These five categories accounted for over half of all Instant Checkout orders. The pattern is telling: these are items expensive enough individually to clear shipping thresholds and avoid small-basket fees. Supplements performed particularly well because users new to GLP-1 weight-loss drugs were asking ChatGPT for nutritional advice and receiving recommendations to increase nutrient intake.

This is not evidence that conversational checkout works. It is evidence that consumers will tolerate a broken experience when the alternative is more friction, not less. The moment a shopper needed two items, the model collapsed.

For anyone who has spent time in payments infrastructure, the failure of Instant Checkout is not surprising. It is a confirmation of something the industry has understood for decades: checkout is not a commodity step at the end of a transaction. It is the point where loyalty programmes, bundle logic, upsell mechanics, returns policy, tax compliance, fraud detection, and customer identity all converge.

OpenAI tried to abstract that complexity away. It positioned itself as a frictionless intermediary, passing order details, payment information, and shipping preferences to merchants through the Agentic Commerce Protocol, built with Stripe. On paper, elegant. In practice, it left enormous gaps.

As Lengow reported, OpenAI had not built a system for collecting and remitting state sales taxes in the US as of February 2026. For any transaction flowing through ChatGPT’s checkout, that was not a minor oversight. It was a live regulatory exposure. Refunds, cancellations, fraud prevention, and customer service all remained the merchant’s responsibility, yet the checkout experience sat inside OpenAI’s interface, creating an accountability gap that neither party had fully resolved.

Forrester’s March 2026 ConsumerVoices survey confirmed the consumer side of this equation: among online adults in the US, UK, and Canada who regularly use answer engines, completing a purchase inside the engine is their least-adopted use case. Asking general questions and researching products ranked first and second. Forrester described the checkout moment as “emotional” and compliance-heavy, noting that it is easy to lose a customer’s patience, trust, and attention before an order is complete.

Only 23 percent of Gen X online adults in the US had even used ChatGPT to search for products in the past month, according to Forrester’s December 2025 Consumer Pulse Survey. Adoption climbed to 32 percent for Millennials and 35 percent for Gen Z, but these are discovery numbers, not purchase numbers. The gap between browsing and buying inside an AI interface remains vast.

Walmart’s response is not a retreat from AI commerce. It is a structural repositioning of who controls the transaction.

In the new model, Sparky operates as an embedded agent inside ChatGPT and, from next month, inside Google’s Gemini. Users log into their Walmart account the first time they encounter Sparky in either platform. From that point, their basket syncs across surfaces: items added on the Walmart app, the website, and inside ChatGPT all appear in the same cart. Checkout happens through Walmart, not through OpenAI.

The difference is substantial. A consumer who adds peanut butter on the Walmart app one day, aluminium foil the next, and a last-minute birthday gift through ChatGPT checks out once, through Walmart’s own infrastructure, with Walmart’s bundle logic, shipping optimisation, and loyalty integration intact.

Danker described the philosophy to WIRED directly: when Sparky travels to external platforms, it brings the Walmart store to the customer rather than forcing them into a disconnected experience.

The commercial case for Sparky is already building. Walmart US CEO David Guggina disclosed during the company’s Q4 earnings call that Sparky users have an average order value approximately 35 percent higher than non-Sparky customers. Roughly half of Walmart’s app users have engaged with the chatbot. And Danker told WIRED that ChatGPT is now bringing Walmart approximately twice the rate of new customers compared to search engines, likely because ChatGPT’s power users skew toward demographics that are not typical Walmart shoppers.

Sparky itself is built on open-source generative AI models combined with retail-specific models trained on decades of Walmart data. The system routes different queries to different models based on answer quality. It is designed to be visually flexible, adapting its look and feel to operate naturally inside external environments. This is not a white-label chatbot. It is a portable commerce agent designed from the ground up to travel.

The Walmart approach sits at one end of a widening strategic divide. At the other end stands Amazon.

Where Walmart is actively embedding Sparky into external AI platforms and welcoming third-party agents to shop on its website, Amazon has locked down. The company has blocked ChatGPT and dozens of other AI agents from accessing its marketplace. In November 2025, it sued Perplexity AI over the startup’s Comet browser agent, which was making purchases on Amazon on behalf of users.

A federal judge issued a preliminary injunction on March 9 blocking Comet from accessing Amazon’s password-protected systems. The ruling hinged on a distinction that will define the next decade of AI commerce law: Comet had the user’s permission, but not Amazon’s authorisation. Bloomberg reported on March 17 that the 9th Circuit Court of Appeals granted Perplexity a reprieve, pausing the injunction while the appeal proceeds.

The timing of broader moves adds another layer. Amazon announced a $50 billion investment in OpenAI on February 27. One week later, on March 4, OpenAI quietly removed Instant Checkout. Forrester analysts have speculated openly about whether the two events are connected, suggesting the OpenAI-Amazon partnership may have reshaped the entire competitive landscape for consumer agentic commerce.

Danker told WIRED that Walmart has no plans to block external AI agents from shopping on its website. The philosophy is straightforward: support whatever tools customers choose, as long as the experience is sound. No erroneous orders, no shocking bills, no excessive customer service burden.

The Instant Checkout pivot is not an isolated product decision. It is a market-level signal about where durable value sits in AI-mediated commerce.

An OpenAI spokesperson confirmed to Skift that the company now wants to focus on helping users research products while giving merchants more control over checkout. TD Cowen analysts, quoted by Futurism, called the retreat a “stunning admission” and suggested that the thesis of AI platforms replacing apps to become the new operating system is either not playing out or has been pushed back significantly.

We think the picture is more nuanced than that. Discovery is clearly the layer where AI adds immediate, defensible value. Consumers use chatbots to compare, research, and explore. They trust their existing checkout flows, saved payment methods, loyalty points, and delivery preferences to complete the purchase. The protocols, OpenAI’s Agentic Commerce Protocol and Google’s Universal Commerce Protocol, survive, but in a narrower form: supporting embedded apps from large, integrated retailers rather than enabling millions of merchants to sell natively inside a chatbot.

The real lesson is about infrastructure readiness. Walmart could execute this pivot because it had already built Sparky, already had basket sync across its digital surfaces, and already operated a fulfillment network capable of converting AI-driven intent into physical delivery. Most retailers do not have this. The gap between wanting to participate in agentic commerce and having the infrastructure to do so on your own terms is enormous.

For payments infrastructure providers, the implications are direct. Checkout orchestration, basket management, tax compliance, fraud detection, and identity verification are not commodities to be abstracted away by a chatbot layer. They are the moat. OpenAI’s five-month experiment proved that the most sophisticated language model in the world cannot substitute for the payments plumbing that makes a transaction safe, compliant, and complete.

If the most advanced AI in the world cannot close a sale, what does that tell us about who really owns the future of commerce: the platforms that surface products, or the infrastructure that moves money?