OpenAI's Checkout Bet Failed: Walmart's Test Proves 3x Lower Sales

OpenAI's bet that it could own the retail transaction is over. The evidence isn't a rumor or a leaked memo — it's a conversion rate. When Walmart, the world's largest retailer, ran its products through ChatGPT's Instant Checkout feature, customers bought at rates three times lower than when the same purchase required clicking out to Walmart.com. Not slightly worse. Three times.

That number comes from Daniel Danker, Walmart's executive vice president of AI Acceleration, Product and Design, in an exclusive interview with WIRED's Paresh Dave published March 18. Danker called the results unsatisfying. Walmart is now killing Instant Checkout and replacing the entire architecture with something that tells you exactly what retailers have learned from the experiment: AI belongs in discovery, not checkout.

The failed experiment had real scale. Walmart put 200,000 products into ChatGPT's Instant Checkout starting in November 2025, following a high-profile partnership announced by Walmart CEO Doug McMillon and OpenAI CEO Sam Altman the previous October. The original press release promised "AI-first shopping experiences" and framed Instant Checkout as the future of how people buy things online. The categories were real — vitamins and supplements, automotive parts, beauty products, home goods. This wasn't a toy experiment. Walmart committed.

And it didn't work, for reasons that go deeper than interface design.

The Information broke the story around March 6 that OpenAI was scaling back its Instant Checkout plans. But Danker had already signaled it publicly at the Morgan Stanley Technology, Media and Telecom conference on March 4, two weeks before WIRED published the conversion data. He wasn't hiding the result. The question was what came next.

The answer is Sparky. Walmart's own AI shopping chatbot, built on a mix of open-source and proprietary retail-trained models, will be embedded directly inside ChatGPT beginning the week of March 25 — and inside Google's Gemini in April 2026. The critical design decision: Sparky routes you back to Walmart.com to complete the purchase. The AI surfaces intent and context. Walmart handles the transaction.

That's not a small tweak. That's a complete architectural reversal of the premise that made Instant Checkout interesting to OpenAI in the first place. OpenAI's commerce strategy had been built around a commission-per-transaction model — own the checkout moment, take a cut. Now, as CNBC's full landscape piece on March 20 makes clear, that model is effectively dead. OpenAI is pivoting to an Apps SDK approach where merchants build integrations that redirect users to their own checkout flows. The platform role shifts from destination back to discovery layer.

Danker's explanation for the failure is worth taking seriously. Conversion rates inside ChatGPT were lower not because customers couldn't figure out how to complete a purchase, but because they were more hesitant to do it in that context. Completing a transaction requires a level of trust in the environment — trust in what happens to your payment information, trust that the order is real, trust in returns and customer service. Those trust signals belong to Walmart.com, built over decades. ChatGPT, however capable as a conversational interface, hadn't earned them.

Analysts were watching this coming. Forrester, in a March 2026 blog post by Emily Pfeiffer, noted that its own March 2026 survey found in-app purchase is the least-adopted use case among answer engine users — a finding that contextualizes Walmart's 3x gap as a market-wide problem, not just a Walmart problem. Pfeiffer also highlighted a structural data issue: AI systems scraping product information struggle to keep up with real-time inventory and pricing changes, which creates accuracy problems that erode the experience further. Gartner's Bob Hetu has argued that OpenAI underestimated the full complexity of transaction infrastructure, from fraud detection to order management to customer service resolution.

The Sparky architecture is deliberately designed to avoid exactly the lock-in that made this failure costly. Sparky is intentionally model-agnostic, built to run across multiple AI platforms rather than depending on any single one. That explains why Walmart signed a deal with Google's Universal Cart Platform on January 11 — before the public announcement of Instant Checkout's death, while the experiment was still nominally live. Walmart had already hedged. It knew which way this was going.

Which makes the framing in some outlets — that Walmart "fired" OpenAI — straightforwardly wrong. The headline TheStreet ran isn't what happened. Walmart is still inside ChatGPT. ChatGPT is still driving twice the new customer acquisition rate for Walmart that search engines do, according to Danker. The value of the partnership didn't go away; the model for monetizing it changed. OpenAI lost the commission. Walmart kept the customer.

That distinction matters for how to read what comes next. The Instant Checkout failure wasn't unique to Walmart. The Information's reporting and CNBC's subsequent landscape piece make clear that only around 30 Shopify merchants were live in Instant Checkout when OpenAI pulled the plug. The footprint was tiny. The commission model had barely gotten started before it became clear it wasn't working.

OpenAI's commerce experiment has now produced a data point that every retailer and every AI lab working on commerce can use: customers will use AI to discover what to buy and even to identify where to buy it, but when they're ready to hand over payment information and confirm a shipping address, they want to be somewhere they already trust. AI as referral layer, retailer as destination. That's the equilibrium the data found.

Sparky inside ChatGPT and Gemini is Walmart's bet that this equilibrium holds. The open question is whether a Sparky-style integration — AI-surfaced context, retailer-owned checkout — can capture the discovery value ChatGPT is already demonstrating without losing it to the friction of the redirect. Danker's 2x new customer acquisition stat suggests it might. The conversion data from Instant Checkout suggests the friction of the redirect is actually an acceptable trade-off.

What to watch: whether OpenAI's Apps SDK pivot attracts merchants at scale, and whether any retailer attempts a native in-ChatGPT checkout that cracks the trust problem Walmart couldn't solve.