In September 2025, OpenAI and Stripe published the Agentic Commerce Protocol and OpenAI launched Instant Checkout, letting people buy things without leaving a ChatGPT conversation. It was widely described as the beginning of AI-driven shopping.
On 5 March 2026, The Information reported that OpenAI was pulling back from Instant Checkout, and OpenAI confirmed the shift the same day. It published its own account on 24 March 2026. It was not converting.
The failure is more instructive than the launch was, and what replaced it is what retailers should actually be planning around.
What went wrong
The numbers are modest in a way the coverage was not. Shopify’s president Harley Finkelstein told a Morgan Stanley conference on 3 March 2026 that only about a dozen Shopify merchants had gone live with agentic checkout. Consumer usage stayed thin across the five-month trial. Etsy, one of the launch partners, told Modern Retail in March 2026 that it had not seen a large volume of sales from the feature.
Notice what did not fail. The technology worked. Payments cleared, orders were placed, the protocol did what it said. What failed was the assumption that people who were happily asking an assistant for recommendations also wanted to complete a purchase inside that conversation, rather than going to the shop they had just been told about.
That is a distribution and habit problem, not a capability problem. It is also the most common way that AI features fail, and it is almost never the thing anyone tests for.
What actually survived
The protocol outlived the product. The Agentic Commerce Protocol was released under an Apache 2.0 licence in September 2025 and is still very much alive. Its job changed. Rather than completing a purchase inside ChatGPT, it now handles product discovery, and the transaction completes on the merchant’s own store.
That is a smaller claim and a more durable one. Discovery is where AI assistants are strong, and checkout is where merchants already have working systems they have spent years optimising.
Meanwhile a second standard emerged. The Universal Commerce Protocol, from Google and Shopify, covers the full journey rather than just discovery, extending through cart, payment, order tracking, returns and customer service.
Two protocols, two surfaces
The practical situation for a retailer in 2026 is that these are not competing options you choose between. They serve different places your customers are.
ACP is the route to appearing in ChatGPT and Microsoft Copilot. UCP is the route to appearing in Google’s AI Mode. A merchant who wants to be visible in both needs to satisfy both, which is less alarming than it sounds, because the underlying work overlaps almost completely.
The work is boring, which is the good news
Nearly everything these protocols ask of you is something a well-run online shop should have anyway.
Complete, structured product data. Not marketing copy. Attributes an agent can compare against: size, material, compatibility, dimensions, whatever a customer would actually filter on. Vague descriptions written for humans who will look at a photo do not survive contact with a machine that cannot.
Accurate, current pricing. An agent that surfaces a stale price creates a customer service problem at best.
Real-time inventory. This is where most small retailers fall down. Being recommended and then being out of stock is worse than not being recommended, because you have spent the customer’s goodwill rather than saved it.
API coverage for cart, checkout and orders. Whether your platform supports UCP natively, or you need an integration layer on top, depends entirely on what you are running. This is the question to put to your platform vendor now.
None of that is exotic. It is the same product-data hygiene that has driven organic visibility for a decade, applied to a new consumer.
The lesson that generalises past retail
Instant Checkout is worth remembering even if you never sell a physical product, because the shape of the failure repeats everywhere.
A capable AI feature was built, it worked technically, and it was placed at a point in a journey where users did not want it. Five months and a dozen merchants later it was withdrawn. The organisation that built it has more capability and more usage data than almost anyone alive, and it still got the placement wrong.
If that can happen there, the odds that your first guess about where AI belongs in your business is correct are not good. Which is an argument for cheap, fast, measured experiments rather than long confident ones, and for treating the first version as a question rather than an answer.
What to do now
Audit your product data before you audit anything else. If your attributes are thin, no protocol work will help, because there is nothing for an agent to reason about.
Ask your ecommerce platform where they stand on UCP. Get a date, not a sentiment. Platform-level support is the difference between a configuration change and a project.
Check whether you are visible in AI answers at all. Ask ChatGPT and Google’s AI Mode what they recommend in your category and see whether you appear. Our guide to why a website can be invisible to ChatGPT covers the technical reasons this often goes wrong, and what GEO involves covers the broader discipline.
Do not rebuild your checkout for this. The direction of travel is discovery in the assistant and purchase on your own store. Your existing checkout is the asset here, not the obstacle.
Frequently asked questions
What happened to ChatGPT Instant Checkout?
The Information reported on 5 March 2026 that OpenAI was pulling back from Instant Checkout, and OpenAI set out its own account on 24 March 2026. Around a dozen Shopify merchants had shipped against it, and launch partners reported thin sales volumes.
What is the difference between ACP and UCP?
The Agentic Commerce Protocol, from OpenAI and Stripe, now covers product discovery with the purchase completing on the merchant’s own store. The Universal Commerce Protocol, from Google and Shopify, covers the whole journey including cart, payment, order tracking, returns and service. ACP is the route into ChatGPT and Copilot, UCP into Google’s AI Mode.
Do I need to support both protocols?
If you want visibility across both ChatGPT and Google’s AI surfaces, yes. The underlying preparation is largely the same for both, so the duplication is smaller than it appears.
What is the single most important thing to fix first?
Product data quality. Structured attributes, accurate pricing and real-time stock levels. Every protocol conversation depends on that foundation, and no amount of integration compensates for its absence.
Flux Dynamics builds and maintains ecommerce and web systems for UK businesses, including the product data and integration work that decides whether an AI assistant can see you. Tell us what you are selling and where.