The 2026 evidence has settled a question businesses were anxious about all year: where does the AI actually sell for you, and where do you keep control? The answer is "discover in AI, buy on your site." AI assistants have become a powerful discovery engine that reaches buyers search does not. But when companies tried to close the sale inside the chat, conversion fell apart. The working model that emerged is a division of labour: the assistant handles discovery and research, and you keep checkout, order validation, tax, payment, and fulfilment on your own site. This is not a preference we are arguing for. It is what the field data shows and what the leading commerce standard already enforces.
The two Walmart numbers that tell the whole story
The clearest field evidence comes from Walmart, and it only makes sense when you read the two numbers together.
First, in-chat checkout underperformed. Walmart's EVP of product and design, Daniel Danker, disclosed that ChatGPT Instant Checkout converted at roughly one-third the rate of a normal Walmart.com transaction, about three times worse, and said Walmart was moving away from in-chat checkout (WIRED, 18 March 2026; Search Engine Land, 19 March 2026). Trying to complete the purchase inside the chatbot lost most of the sale.
Second, and easy to miss, the same channel was a discovery machine. Danker told WIRED the chatbot was "now bringing in about twice the rate of new customers as search engines." So the assistant was reaching people search could not, then failing to close them once you asked it to be the register.
Taken singly, these numbers mislead. Taken together, they are the whole argument. AI is excellent at getting you discovered by new buyers and poor at being your checkout. So do not chase in-chat checkout. Chase being the answer the assistant gives, then bring the buyer to a page you control to close the sale.
Discovery exploded, and it converts well on your own site
The Walmart story is one merchant. The market pattern is broader, and it points the same way.
Adobe Analytics, in an analysis of more than one trillion visits, found AI-driven traffic to US retail sites was up 393% year over year in Q1 2026 (Adobe Digital Insights, ~16 April 2026). That is the discovery surge in one figure. The more important finding sits underneath it: AI-referred shoppers, once they land on the retailer's own site, now convert comparably to or better than other traffic, with higher revenue per visit and stronger engagement.
Notice what this does not say. Adobe's better-conversion number is about AI-referral traffic arriving on the merchant's site, not about closing the sale inside a chat window. Walmart's three-times-worse number is about in-chat checkout. Hold them apart and they stop competing. Discovery happens in the AI. Conversion happens well on your own site. That is the settlement, stated as data.
The architecture already enforces it
If the field data were the only evidence, you could dismiss it as early and noisy. It is not the only evidence. The leading standard for AI commerce is built the same way.
OpenAI and Stripe's Agentic Commerce Protocol (ACP) keeps the merchant as the merchant of record (agenticcommerce.dev). Under ACP, ChatGPT calls your endpoints to create and update a checkout session, but your stack does the load-bearing work: you validate the order, determine fulfilment, calculate and charge tax, run payment and fraud risk, and charge the card. The assistant surfaces and negotiates. You remain the seller.
That is a design decision with a plain business meaning. The engineering task is not "integrate a chatbot checkout." It is "expose two clean surfaces." One is a discovery layer an agent can read. The other is a hardened confirm-and-authorize boundary where the transaction is committed on your systems, with a record of who agreed to what, at what price, at what moment. Even the standard that lets an assistant sell for you draws the line at your checkout.
The risk signal is worth stating plainly. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls (Gartner, 25 June 2025). The shakeout will be real. A model that keeps your durable assets on your own stack survives it.
Where AI helps and where the sale should close
The settlement resolves cleanly once you map each stage of the buying journey to who does it best and what you have to do about it.
| Stage | Who does it best | What the business must do |
|---|---|---|
| Discovery | AI answer engine | Publish clean structured data (JSON-LD, the shared vocabulary that tells an agent what your products, prices, and availability are) so the assistant can read and recommend you. |
| Research and comparison | AI answer engine | Serve fast, correct pages rendered on the server, so an agent reads your facts without having to run your site's scripts. Keep specs, prices, and stock current, because a cached wrong answer is worse than none. |
| Checkout | Your own site | Keep order validation, tax, payment, and fraud on your stack. Do not delegate the charge to a chat window. |
| Authorization | Your own site | Put a clear confirm step before anything is charged: a human or a bounded agent commits the sale, with an audit trail of what was approved. |
| The relationship after | Your own site | Own the destination the AI points to. First-party data and the direct customer relationship stay yours, not the assistant's. |
The non-obvious risk in this table is up-funnel. "Be the source AI recommends" quietly makes you a supplier to an intermediary you cannot fully see. If discovery consolidates inside a handful of assistants, they become the new gatekeepers, the way search engines once were, and they can reweight who gets recommended without telling you. The hedge is exactly the merchant-side architecture the settlement already forces. Because checkout, the customer relationship, and the transaction record stay on your stack, you keep the durable asset even as the discovery layer shifts under you. The maturing of cross-vendor interoperability, such as Google's A2A protocol (announced April 2025, now under Linux Foundation governance with more than 150 organizations, per Google Developers Blog), points to a future of many agents rather than one. That favours merchants who are readable by all of them over merchants locked into any single assistant's checkout.
Where Origin Pi stands
We read the 2026 evidence as a division of labour to design for, not resist. The AI handles discovery and research. The merchant keeps checkout, validation, tax, payment, and fulfilment. That framing is not something we imposed on the data. It is what the leading standard already enforces, and what the Walmart and Adobe numbers, held apart, jointly prove.
Our view of who wins: the merchants who are legible to agents and controlled at the boundary. Concretely, that is two layers we build for clients. First, the agent-ready layer, which is structured data and fast, correct, server-rendered pages, so you are the source the assistant recommends rather than the one it cannot parse. Second, the confirm step, a human-in-the-loop or bounded-agent authorization with bounded permissions and an audit trail at the checkout edge, which is exactly where the ACP standard already places validation and charge. Keeping the sale on your stack is also the defensible posture on liability and consent: you retain the audit trail card networks and regulators expect, and you are not handing authorization to a third-party assistant whose reasoning you cannot inspect.
We hold to one honest caveat. The Walmart figures are executive-stated and the Adobe figures are vendor-analysed. We treat the direction as strong and the magnitudes as estimates. The move is unglamorous and durable. Get discovered in AI. Get paid on your own site. Keep the record.
