30 Merchants, Millions Promised: The Early, Messy Reality of Agentic Checkout

The launch, in numbers

The headline was that anyone in the United States could now finish a purchase inside a chat. On February 16, 2026, OpenAI extended Instant Checkout, its "Buy it in ChatGPT" flow, to all US ChatGPT users. The rails were the Agentic Commerce Protocol (ACP), an open specification developed jointly by Stripe and OpenAI to let an agent initiate and authorize a transaction on a buyer's behalf.

The count of merchants who could actually receive those transactions told a quieter story. Reporting the same month, Forrester analyst Emily Pfeiffer put the number of Shopify merchants live on Instant Checkout at roughly thirty. Set against the stated ambition, which included PayPal positioning its own ACP server to onboard tens of millions of small businesses across 2026, the launch was less a mass rollout than a controlled pilot wearing the language of a platform shift.

These figures are press-sourced rather than audited, and they belong to a live 2026 story that is still moving. The right posture is to treat the numbers as directional signals of where the gap sits, not as settled totals.

ChatGPT Instant Checkout: what actually shipped

What shipped was a working transaction path, not a fully populated marketplace. Instant Checkout let a US ChatGPT user move from a product recommendation inside the conversation to a completed order without leaving the chat surface, with the merchant catalog and fulfillment sitting behind the ACP interface. The capability was real and generally available; the supply of merchants able to transact through it was the constraint.

That distinction matters because the merchant surface is where small businesses live. A protocol can be open, generally available, and technically sound, and still reach almost no one if the merchants who would sell through it are not connected. At launch, the roughly thirty live Shopify merchants meant the demand side had a front door and the supply side had barely walked through it.

The Agentic Commerce Protocol and its backers

The Agentic Commerce Protocol is the specification underneath Instant Checkout: a published way for an AI agent to discover a product, initiate a purchase, and authorize payment on a buyer's behalf. Stripe and OpenAI developed it jointly and released it as an open standard, which is what allows a payments network beyond Stripe to build against the same interface.

PayPal did exactly that. Its planned ACP server was positioned to bring tens of millions of additional small businesses into agent-initiated commerce over 2026. The presence of more than one serious backer building on a shared protocol is genuinely significant infrastructure. It is also the source of the gap being examined here: the protocol layer and its projected reach were declared well before the merchant layer had populated. Announced capacity and connected supply are not the same measurement, and conflating them is how a thirty-merchant pilot gets narrated as a tens-of-millions moment.

The March pivot to merchant-controlled checkout

The clearest evidence that the initial design under-served merchants is that its own author changed it. On March 4, 2026, roughly two weeks after the general-availability launch, OpenAI pivoted toward merchant-controlled checkout.

A move toward merchant control is an implicit concession that the first, more fully agent-led model did not give merchants enough authority over how a sale completed, what the buyer saw, and how the merchant appeared in the transaction. Read plainly, the sequence is a launch, a visible adoption gap, and a fast design correction toward the party the protocol most needs and had least served. That is not a failure so much as a normal early-cycle correction, but it is a useful corrective to any claim that agentic checkout arrived finished.

Why the gap is a pattern, not a surprise

This shape is familiar. In cross-border digital trade, the World Bank and WTO record a policy and technical layer that gets fixed faster than the operational-capability layer beneath it. The WTO frames its e-commerce moratorium and trade-facilitation provisions as small-business enablers, while its own MSME handbook names the binding constraints that persist: disproportionate customs burdens, unfamiliarity with e-commerce mechanics, and transport and logistics gaps. The door opens legally and technically before it opens practically.

Agentic commerce is the latest instance of the same pattern. The standard is published, the payment networks are committed, and the buyer-facing capability is live, all ahead of the merchant onboarding, catalog readiness, and control that would let ordinary small businesses actually transact through it. Each wave of "borderless" commerce, from globalization to e-commerce to agent-mediated checkout, tends to open the technical door first and the operational one later. Naming that pattern is what keeps a 2026 launch in proportion.

AI shopping agents and the visibility question

Even once merchant onboarding catches up, a second question sits underneath agentic checkout: being transactable is not the same as being chosen. If an agent completes the purchase, the agent first has to surface the merchant, and the surfaces that feed these agents already appear to be narrower than classic search.

The reading for a small merchant

The temptation is to swing between two errors: treating a thirty-merchant pilot as proof that agentic checkout is vaporware, or treating a tens-of-millions projection as proof that it is already the main event. Neither survives the evidence.

There is a genuine counter-thesis worth holding. Anderson's Long Tail argues that when distribution cost collapses, niche and small-catalog sellers can win share that a shelf-space economy denied them, and a mature agent layer that can reach any connected merchant is, in principle, a distribution-cost collapse. That is the optimistic case, and it is testable rather than automatic. It holds only if the small merchant is connected, controllable in the transaction, and visible to the agent in the first place, which is precisely what the 2026 launch did not yet deliver at scale.

The practical stance for a small business is therefore neither urgency nor dismissal. It is readiness held in proportion: understand what agent-buyable actually requires of a storefront, watch the merchant-adoption numbers rather than the announcement numbers, and decide what to prepare now against what to leave alone until the surface stabilizes. The measurable question is not "is agentic checkout the future," but "if a buyer's agent went shopping in my category today, could it find me, transact with me, and would it choose me," and that is answerable now.

How to weigh the evidence

Because this is a fast-moving 2026 story assembled from press and analyst reporting, the tiering below separates what is well established from what is still emerging or contested, so that nothing on this page is read as more settled than it is.