ChatGPT's in-chat checkout converted three times worse than Walmart's site. OpenAI moved on in March.

In November 2025, Walmart let ChatGPT shoppers buy from roughly 200,000 products without leaving the conversation. By March 2026, the retailer had largely walked away from it. Daniel Danker, Walmart's EVP of product and design, said purchases completed inside ChatGPT converted three times worse than purchases where a shopper clicked through to Walmart's own site. OpenAI had already confirmed it was scaling back Instant Checkout, the feature Walmart and Etsy helped launch five months earlier. The in-chat buy button is fading fast, and what replaces it, not the checkout mechanics themselves, is what will actually decide whether your products show up in ChatGPT.

What Instant Checkout actually was, and why it stalled

Instant Checkout ran on the Agentic Commerce Protocol, an open standard Stripe and OpenAI co-developed and launched on September 29, 2025. It let a ChatGPT user pay a merchant mid-conversation using a Shared Payment Token, so the chatbot never saw a card number and the order still landed in the retailer's own fulfillment system.

Etsy went live first, with millions of listings. Shopify followed, with named brands including Glossier, Vuori, Spanx, and SKIMS testing the flow. The protocol worked as designed. What it didn't produce was sales. Etsy told Modern Retail the feature never generated large sales volume and said it has been more useful as a discovery channel that drives referral traffic than as a point of purchase. Instacart reported that only a small fraction of its orders originated from AI agents at all. One unnamed executive at a large specialty retailer put it more bluntly in October 2025: the feature was far from ready for prime time.

The checkout model OpenAI is walking back toward apps

OpenAI confirmed the pullback in March 2026, after The Information reported internally that a scale-back was coming. In its own blog post, OpenAI said the original design lacked the flexibility it wanted to give merchants, so the company shifted toward letting developers build full ChatGPT apps instead.

Current developer documentation lists three ways a ChatGPT app can take payment. External checkout, sending the shopper to a merchant-hosted page on the brand's own domain, is the recommended path and the only one generally available; the merchant keeps control of pricing, tax, and refunds. A saved-payment-methods flow can show a returning customer the payment methods they already stored with that merchant, without collecting anything new. A ChatGPT-hosted payment sheet, which renders line items in the chat and calls a complete_checkout tool on the merchant's server, remains in private beta and is open only to select marketplaces. None of the three brings back the single universal buy button Instant Checkout promised in 2025. It fits a pattern GEO practitioners have seen before: ChatGPT's interface evolves faster than OpenAI explains how it affects who actually gets credited.

The product feed is still what decides whether you're found

Whatever happens to checkout, ChatGPT still has to find a product before it can show one, and that part runs on a structured feed OpenAI documents in detail. The feed uses JSONL records with nine required fields per item: item_id, title, description, url, brand, seller_name, image_url, availability, and price.

Availability only accepts five values, in_stock, out_of_stock, pre_order, backorder, or unknown, and anything else gets the row rejected outright. Two boolean flags, is_eligible_search and is_eligible_checkout, control whether a product can appear at all and whether it can be bought in-app, and turning off search eligibility disables checkout eligibility too. OpenAI's own documentation is blunt about what eligibility actually guarantees:

Eligibility does not guarantee display.

Review counts and star ratings go into the feed as data, but OpenAI doesn't say how, or whether, they affect ranking. That's consistent with how ChatGPT treats citations more broadly: the inputs are documented, the weighting isn't, the same gap that shows up in how ChatGPT chooses its sources for ordinary search answers. For brands, the practical takeaway is that feed hygiene, real brand names, current prices in the right format, HTTPS product and image URLs that actually resolve, is the one lever fully within your control, regardless of which checkout model wins.

Walmart chose its own app instead of OpenAI's checkout

Walmart's answer to a three-times-worse conversion rate wasn't to fix Instant Checkout. It was to leave the shared protocol altogether. On March 30, 2026, Walmart launched its own ChatGPT app, powered by its Sparky assistant, that takes shoppers into a dedicated Walmart environment instead of completing the purchase inside the conversation.

Inside that environment, a shopper can link their Walmart account and loyalty membership and check out through Walmart's own payment systems, things the original single-item Instant Checkout flow couldn't support. Danker said the move paid off in a different way than checkout ever did: ChatGPT brings Walmart new customers at twice the rate of search engines, even though those customers rarely bought anything without leaving the chat first. It's a similar calculation to the one Amazon made when it retired Rufus for Alexa for Shopping, treating the AI assistant as an acquisition channel rather than a transaction layer. Etsy is reportedly building a standalone app of its own for the same reason: control over the experience outlasted the novelty of a shared checkout button.

Google is betting on an open protocol, not its own checkout button

OpenAI isn't the only company redesigning AI checkout in real time. At NRF 2026 in January, Google announced the Universal Commerce Protocol, co-developed with Shopify, Etsy, Walmart, and Target and endorsed by more than twenty companies, to let shoppers buy directly inside AI Mode and Gemini using Google Pay or PayPal.

The contrast with OpenAI's approach is the real story. ChatGPT's commerce stack has been narrower and more OpenAI-controlled, and The Information reported that standardizing product data across millions of merchant catalogs was a real part of why Instant Checkout struggled to scale. UCP launched with the same major retailers already lined up as initial sellers. Neither company has solved agentic commerce yet; purchase volume through both protocols is still small. What it does mean is that a brand's visibility inside AI shopping results now depends on which protocol, or protocols, it feeds, and that answer is different in ChatGPT than it is in Gemini, which tracks the same inconsistency that shows up across AI engines generally, where the same brand can rank completely differently depending on which system is asked.

What this means for brand visibility beyond the checkout button

The lesson isn't that agentic checkout failed. It's that checkout was never the layer that mattered most for visibility. Three things are worth acting on now, regardless of which protocol or app model eventually wins the shopping war.

First, don't architect a commerce strategy around a single platform's checkout feature. Instant Checkout went from launch to wind-down in five months; whatever ChatGPT, Gemini, or Perplexity ship next will likely move just as fast. Second, keep investing in the parts of a product page that make it legible to an AI agent in the first place, since most product pages still aren't built for the AI agents evaluating them, feed or no feed. Third, measure the channel the way Walmart ended up measuring it: not by in-chat conversion, but by whether AI engines are sending qualified shoppers your way at all, since AI search traffic has converted meaningfully better than Google organic in several studies, even where in-chat purchases haven't. Tracking which products and brands actually surface in shopping-intent prompts, across ChatGPT, Gemini, and whichever protocol wins next, now matters more than tracking any single checkout flow.