A shopper asks ChatGPT for a waterproof jacket under $150. The assistant compares options, narrows them to three, hands back a short list. Somewhere in that exchange your store either appeared or it didn't, and there is no line in your analytics for the moment it happened.
That is agentic commerce as it exists right now. Not a robot with your credit card. A piece of software sitting between your shopper and your storefront, reading whatever it can find, and deciding who gets shown.
The term has collected a lot of noise, most of it from companies selling the future. Here is the plain version, plus the part vendor explainers skip: which half of this works, and which half quietly stalled.
Agentic commerce is any purchase journey where an AI agent acts on the shopper's behalf instead of the shopper clicking through your site themselves. The agent might only recommend. It might assemble a cart. In the fullest version it pays, too. None of that describes the support chatbot on your own site, which answers questions rather than acting for the shopper, and which is bought and billed on completely different terms: what an ecommerce chatbot actually costs to run is a separate question from anything on this page, however often a vendor pitch blurs the two.
OpenAI and Stripe put the shift plainly when they launched the first checkout protocol for it in September 2025: traditional e-commerce was built for humans, with the business controlling the interface and the payments, while in AI-led commerce the agent carries the buyer's identity, payment method and context into the transaction.
Google's agent payments spec is blunter about what that breaks. Today's payment systems assume a human is clicking buy on a trusted website. Take the human out and three questions open up, none of them settled. Who authorised this. Did the agent reflect what the buyer actually wanted. When it goes wrong, who pays.
Those three questions are the whole story of why progress here has been lopsided.
It helps to stop treating agentic commerce as one thing. There are three stages, at wildly different levels of maturity.
Discovery. The agent finds, compares and recommends products. The shopper clicks through and buys on your site, on your checkout, with your brand in front of them. This works, at scale, and it is generating real revenue right now.
Delegated checkout. The agent assembles a cart and the human approves that specific purchase without leaving the chat. Live, technically. Barely used.
Autonomous purchase. The agent buys against rules you set in advance, no human confirming the individual transaction. The specifications exist. Consumer appetite does not.
Gartner put a number on that last point. Among 322 US consumers surveyed in January 2026, willingness to let AI make purchase decisions topped out at 11%, and only for low-stakes things like household supplies. Openness to AI narrowing the choice ran roughly three times higher. Small sample, so read it as direction rather than precision. The direction is unambiguous: people want help choosing and they want to press the button themselves.
This is the fact that should reset how you plan.
Instant Checkout went live inside ChatGPT in September 2025, starting with Etsy sellers and a promised rollout to Shopify. Six months later, in March 2026, OpenAI wound it back in its own words: the initial version did not offer the flexibility they wanted to provide, so merchants would use their own checkout experiences while OpenAI focused on product discovery.
The reasons are ordinary commerce problems, not AI problems. Walmart, Etsy and Instacart each said a version of the same thing afterwards: volume through the chatbot was small, and goods sold inside the chat converted worse than the ones that sent a shopper to the retailer's own site. Some of it was as mundane as shoppers worrying their order would arrive as five separate boxes. What the in-chat flow could not carry was the ordinary machinery of retail: promotions, loyalty, real-time inventory, store pickup.
So the protocol survived and got repurposed. It now moves product data into ChatGPT rather than money out of it.
While the checkout story stalled, the boring half got good.
Adobe measures traffic arriving at US retail sites from generative AI tools across an analytics base it describes as covering over a trillion visits. In the first quarter of 2026, that traffic grew 393% year over year and converted 42% better than non-AI traffic in March. A year earlier the same measurement had AI traffic converting worse than average, so this is a reversal, not a continuation. Growth has cooled since as the base got bigger. The conversion advantage has not.
Read the two halves together and the strategy writes itself. The channel that pays is the one where the agent sends a well-qualified shopper to your storefront and your checkout. You keep the brand, the data, the promotions and the customer relationship. All you have to do is be findable, which is a narrower problem than it sounds and one we've taken apart in getting your products recommended by ChatGPT and Gemini.
Short version: more announcements than transactions.
The Agentic Commerce Protocol from OpenAI and Stripe is live for discovery, with Target, Sephora, Nordstrom and Best Buy among the retailers OpenAI names as integrated. The Universal Commerce Protocol, co-developed by Shopify and Google in January 2026, does support agent-completed checkout and carries what Instant Checkout couldn't: discount codes, loyalty credentials, subscription cadences. On the payment side, Google handed its agent protocol to the FIDO Alliance in April 2026 and the card networks have since backed an alliance of their own. No settled standard has come out of any of it, and none of it needs a decision from you this year. Which protocol wins is a question for your platform vendor, not your roadmap.
Here is the part most owners get backwards. They assume the work is payment integration. It isn't. Most of the agentic commerce work we take on turns out to be catalog work with a better name on the invoice: fixing feeds, making price and stock honest in real time, checking that the machines reading a store are allowed to. That sits inside our AI integration work and costs a fraction of what owners brace for.
An agent recommending your product needs four things to be true. A product feed with stable identifiers, complete attributes, and prices and stock correct at the moment they're read. Product pages that render their content to a machine, not only to a browser. Returns and shipping policies published as structured, machine-readable links, because that is a required protocol field and it is what an agent checks before recommending anything. And permission to read you at all, which lives in your robots.txt: OpenAI's crawler documentation states that sites opted out of its search crawler will not be shown in ChatGPT search answers.
That last one takes an afternoon. We have opened robots.txt on stores paying an agency for AI visibility work and found the relevant crawler blocked by a rule someone added in 2023.
The catalog side is where the real gap sits. In the same April 2026 analysis, Adobe scored how much of a page a large language model can read and put the average US retail product page at 66%, leaving roughly a third of product content invisible to the models doing the recommending. That is a vendor-defined score from a company selling the fix, so treat the exact figure with suspicion. The pattern matches our audits: price and availability are the fields that have quietly drifted, and a stale stock flag is worse than a missing one.
The checklist version lives in what agent-ready actually means for a catalog.
Liability is unresolved. If an agent buys the wrong size and the shopper disputes it, the rules for who eats that are still being written; the protocols offer an audit trail rather than an answer. One card network has committed to covering erroneous purchases by agents registered with it, and that is the only firm commitment of the kind we could verify.
The strategic risk is quieter and bigger. Every layer between you and your buyer takes something. If agents become the default front door, whoever owns that door decides how much of the customer relationship you keep. Which is the argument for making your own storefront the place the transaction completes — exactly where the market has landed anyway.
There is also the question of what to expose. Opening your catalog to machines is not the same as opening your systems to them, and that boundary deserves thought — we drew it in what to expose to an AI agent and what to lock down.
The work usually starts the same way: an audit of what an agent sees when it looks at a store. It is a short exercise, and the result is rarely what the owner expects. Most of the expensive-sounding problems turn out to be feed problems, and feed problems are cheap to fix.
Let’s build something powerful together - with AI and strategy.
