For about six months, the story of AI commerce was that the transaction would move inside the assistant. Then it moved back.
OpenAI retired Instant Checkout in March 2026 and repositioned shopping in ChatGPT around discovery, routing shoppers to merchants to complete the purchase. Reporting on the retreat points at unglamorous causes: thin merchant participation, and product, inventory, and pricing data that was not reliable enough to transact on.
It is worth being precise about what that does and does not mean, because both overreactions are already circulating.
What it does not mean
It does not mean agentic commerce stalled. Shoppers still start in AI. The behavior that grew over the last two years, describing a need in plain language instead of typing keywords, did not reverse because one checkout flow was withdrawn.
It also does not mean in-assistant checkout is dead forever. Data quality problems get solved. The interesting question is what happens in the meantime, which for grocery is likely to be several years.
What it does mean
Discovery and transaction are splitting, and they are splitting in the grocer’s favor more than the current architecture of most grocery e-commerce can exploit.
If an assistant is going to hand a shopper off to your site with an assembled intent, the quality of the handoff destination becomes the constraint. A shopper who arrives having already described a week of dinners and lands on a search box has to start over. The intent, which is the valuable part, is discarded at the door.
That is the real cost of the split. Not lost discovery. Discarded context.
Grocery has a specific version of this problem
Most categories hand off a single product. Grocery hands off a list.
A shopper researching a jacket arrives with one item in mind, and a product page is a perfectly good destination. A shopper who has just described feeding four people for a week arrives with thirty implicit items, quantity math, dietary limits, and a budget. There is no product page for that. There is either a storefront that can assemble it, or a search box that makes them do it manually.
This is why the assistant retreat matters more to grocers than to most retailers, and why the response is not a better landing page.
Where Neomi sits, stated plainly
Neomi is an AI storefront that a grocer owns, running alongside the store they already operate. A shopper states a need and Neomi assembles a complete, editable cart against the live catalog, then hands it to the grocer’s own checkout. That is the on-site half of “discover in AI, buy on site,” and it is what exists today.
On the off-site half, we will be direct. Neomi can be configured to surface in third-party AI environments, and that is a capability we build per customer on request. It is not a live, generally available consumer feature today, and we are not going to describe it as one. When a grocer asks us for it, we scope it as work.
The practical takeaway
Assume discovery gets rented and transaction comes home. Then ask one question about your own site: when a shopper arrives already knowing what they want, how much of what they know survives the arrival?
For most grocery storefronts today, the answer is almost none of it. That is a fixable problem, and it is fixable without replatforming, because the fix is a second front door rather than a rebuild of the first.