The AI Checkout Got Cancelled. The AI Customer Did Not.
OpenAI switched off in-chat checkout in March 2026 after five months. The protocol survived, the buying behaviour survived, and the work moved back to you.
What actually happened
For most of 2026 the advice to online sellers was that buying was about to move inside the chat window and you should prepare for it. That version of the future ran for about five months and then stopped.
| Date | What happened |
|---|---|
| 29 Sep 2025 | Instant Checkout launches with Etsy; the Agentic Commerce Protocol is published under Apache 2.0 |
| Late Jan 2026 | A 4% checkout fee takes effect, roughly 9.2% combined with card processing |
| 4 Mar 2026 | OpenAI pulls back from operating checkout inside ChatGPT |
| 24 Mar 2026 | The retreat is confirmed in the updated shopping announcement |
Only around a dozen Shopify merchants ever went live. The protocol stayed, the buying behaviour stayed, and the payment step went back to the merchant's own site.
If you spent this spring being told to get ready for in-chat checkout, that is why the advice went quiet.
Why it broke
The reasons are worth reading closely, because they are the same reasons most operational software is harder than its demo.
Sales tax. It is calculated per jurisdiction and per product category, and whoever takes the payment carries the obligation. Doing that correctly on behalf of thousands of merchants at once is a genuinely difficult problem, and getting it wrong is expensive in a way that surfaces months later.
Fraud. The signals that catch a bad order live with the merchant: order history, address patterns, what normal looks like for this catalogue. An intermediary sitting in front of many merchants has less of that context and still has to eat the chargebacks.
Inventory truth. A purchase needs stock to be correct at the moment of purchase, not at the last sync. Selling something you do not have is worse than not selling it, and syncing live availability across many catalogues at scale is exactly the kind of unglamorous plumbing that decides whether a system works.
None of these is a verdict on AI shopping. They are a reminder that the boring machinery around a process is usually the load-bearing part.
What survived
Three things, and they are the ones that matter to a shop owner.
The protocol is still there. ACP remains published under Apache 2.0 and maintained by OpenAI and Stripe, which means the integration you might build is against an open specification rather than one company's product decision.
The buyer is still there. People ask an assistant what to buy, compare options in the conversation, and arrive on a merchant site with the decision largely made. That did not depend on where the card details were entered.
And the work is still yours. Discovery happens in the assistant, checkout happens on your site. Which is to say the responsibility landed back exactly where it was before anyone promised to take it away.
What this actually asks of you
The assistant never sees your design. It reads your facts, and anything it cannot state confidently it leaves out of the comparison entirely, without telling you.
That makes the job concrete:
- Facts in machine-readable form. Price, availability, variants, dimensions, materials, delivery window, returns policy. On the page, in structured data, not only in the visual layout and definitely not only in a photographed spec sheet.
- Stock status that is true right now. A recommendation that lands on an out-of-stock page costs the sale and the next comparison too.
- Boring pre-purchase answers in plain text. Sizing, compatibility, what is in the box, how returns work. Assistants read pages, not chat widgets and not PDFs.
- One canonical page per product. If four URLs describe the same thing, an assistant will pick one, and it may not be the one you would have picked.
This is the same discipline that makes a site legible to any automated visitor rather than a human one, and none of it is wasted if the payment step moves again.
The honest read on the economics
The 4% fee is a useful number to keep, even though the flow it applied to is gone, because it is what an intermediary thought the introduction was worth.
Against a marketplace taking 25-30%, that is cheap. Against selling from your own site at roughly 4% all in, it is more than double. Neither comparison is the real question. The real question is whether the sale would have happened without the introduction, and whether the customer becomes yours afterwards or stays the intermediary's.
An introduction you convert into a repeat customer is worth paying a lot for. An introduction where somebody else keeps the relationship is a rented channel, and rent is not famous for going down. That is the same arithmetic you would apply to any other channel, which is to say the payback question does not change because the channel is new.
What to do this quarter
Nothing dramatic, which is usually the correct answer after a hype cycle deflates.
Fix the product data, because it pays off in ordinary search too. Make stock accuracy real rather than nominal. Keep watching where your traffic says it came from. And treat any vendor pitching an urgent agentic commerce integration with the scepticism the last twelve months earned: the standard is open, the buyers are real, and the checkout is on your own site, where it was to begin with.
Sources: OpenAI on Instant Checkout and ACP, Stripe on the open standard, the ACP specification, and Forbes on the March 2026 retreat.
01Should I still do anything about agentic commerce, given the checkout was switched off?+
Yes, but the work is different from what most people were told to do. The part that got cancelled was the payment step happening inside the chat window. The part that did not get cancelled is people asking an assistant what to buy and then arriving on your site with the decision mostly made. That behaviour is now the normal front end of a purchase for a growing share of shoppers, and it rewards different things than search did. An assistant comparing three products reads structured facts: price, availability, dimensions, materials, returns policy, delivery window. If your product page states those clearly in text and in structured data, you get compared accurately. If they live only in a photo of a spec sheet or in a PDF, you get skipped without ever knowing. That is the work, and it is the same work whether or not the payment ever moves back into the chat.
02Why did in-chat checkout fail if the demand was there?+
For three reasons that have nothing to do with AI and everything to do with running a shop. Sales tax is calculated per jurisdiction and per product category, and the party taking the payment carries the obligation, which is a genuinely hard problem to solve on behalf of thousands of merchants in one flow. Fraud prevention depends on signals the merchant has and the intermediary does not, and chargebacks land somewhere real. And inventory has to be correct at the second of purchase, not at the last sync, because selling something you do not have is worse than not selling it. Only around a dozen Shopify merchants ever went live, which tells you the integration cost was high relative to the volume it returned. None of that is a verdict on AI shopping. It is a reminder that the unglamorous parts of commerce are load-bearing.
03What does the 4% fee tell me about the economics?+
It tells you what an intermediary thinks the introduction is worth, and it is worth comparing against your existing channels before assuming it is expensive or cheap. OpenAI announced a 4% checkout fee for late January 2026, which stacked on top of card processing to reach roughly 9.2% all in. Against a marketplace taking 25-30% that is inexpensive. Against selling directly from your own site at around 4% all in, it is more than double. The right comparison is not the headline percentage but the incremental margin on a sale you would not otherwise have made, and whether the customer becomes yours afterwards or belongs to the intermediary. Introductions you can convert into a repeat customer are worth paying for. Introductions where the intermediary keeps the relationship are a rented channel, and rent goes up.
04How do I make my catalogue legible to an AI assistant?+
Start by assuming the assistant never sees your design, only your facts, and that anything it cannot state confidently it will leave out of a comparison. In practice that means four things. Put price, availability, variant options, dimensions, materials, delivery window and returns policy in machine-readable structured data on every product page, not only in the visual layout. Keep stock status truthful and current, because a recommendation that leads to an out-of-stock page costs you the visit and the credibility. Write the answers to the boring pre-purchase questions in plain text on the page rather than in a chat widget or a PDF, since that is where an assistant reads. And keep one canonical page per product so an assistant does not have to guess which of your four similar URLs is the real one.
Digital Transformation 2026: Why 70% Fail and What the Other 30% Do Differently
Responsible AI in 2026: The Compliance Reality Behind the Ethics Talk
