An order lands in your dashboard. The line items are yours, the shipping address looks ordinary, the money is real. Next to it sits the name of an AI assistant you never signed a contract with. That is the shape of this week's news, and it arrived faster than most shops noticed.
- Meta's Muse assistant now completes purchases through Stripe's Link wallet. If you do not accept Link, Stripe issues a single use virtual card per purchase, so the charge arrives on rails you already take.
- The protocol underneath was written by Stripe, OpenAI and Meta together, so one seller setup points at several assistants rather than one.
- Selling this way is a feed job, not a design job. Product data once a day, inventory and prices every fifteen minutes, and a failure rate for unavailable items that Stripe asks you to hold under five percent.
- You must have a return policy on file before you can onboard as a seller. That is a document requirement, not a checkbox.
- The payment token you receive is scoped to one currency, one maximum amount and an expiry. You see the card brand and last four digits, not the buyer.
- Two optional hooks share a four second deadline and behave in opposite ways when they time out. One lets the sale continue, the other kills it.
What actually changed this week?
Meta wired a working checkout into its assistant. Small Business Trends reported the Stripe Link integration inside Muse on 25 September 2026, quoting Jay Shah, the business lead for Link at Stripe, saying the company is building the infrastructure that lets agents buy from businesses. Buyers attach a payment method once and the assistant can then pay on their behalf.
The reach matters more than the mechanism. TechCrunch reported that Muse launched on 8 September and had roughly 3.4 million downloads by late September on Sensor Tower's count, with 2.8 million in the first fortnight. Its download growth ran at 55 percent day over day early on, against 24 percent for ChatGPT across its first ten days. Meta began house advertising on 9 September across Facebook, Instagram and WhatsApp, and by 22 September sat inside the top ten brands by advertising spend.
So the surface is new and large, and a company is pushing it hard that owns the apps your customers already open. A shopping assistant with a working till is a different animal from one that hands out links, and the same pattern has now reached the browser itself, where a browser agent can finish a checkout on the shopper's behalf.
Can a purchase complete if you never signed up for any of this?
Partly, and this is the detail worth repeating to another shop owner. The Small Business Trends piece notes that for businesses not yet accepting Link, Stripe offers a single use virtual card for each purchase. A virtual card is an ordinary card number. It arrives at whatever checkout you already run, clears on the rails you already take, and settles into the account you already use.
That is not the same as being listed on the assistant's shelf, which does need the seller setup below. But the category of buyer changed without asking your permission, much as it did when a machine first started agreeing to your terms on somebody's behalf. Who consented to what is not settled by the payment clearing.
What the seller side actually asks of you
Stripe's documentation for selling through agents is unusually blunt about the work, and almost none of it is the part shops expect. There is no storefront to design. There is a catalogue to keep honest.
The Agentic Commerce Suite covers the United States, Canada and a list of European countries that includes the United Kingdom, France, Germany, Spain, Italy, the Netherlands, Ireland and the Nordics. Onboarding asks for your terms of service, your privacy policy and your return policy before you can trade. Tax codes are required for each taxable product, and Stripe says plainly that incomplete tax configuration can cause checkouts to fail.
Then the feeds. Four of them, on four different clocks:
| Feed | Cadence Stripe recommends | What it decides | What breaks when it is stale |
|---|---|---|---|
| Product data | Once per day | Titles, descriptions, images, categories | The assistant describes your item wrongly, or skips it |
| Inventory | Every 15 minutes | Whether the item can be bought right now | Agents keep offering something you sold yesterday |
| Pricing | Every 15 minutes | The number quoted to the shopper | Quoted price and charged price disagree, and checkout fails |
| Promotions | As needed | Codes, percentage deals, free shipping | Nothing breaks, but agents push discounted items harder |
One footgun deserves naming. Feed uploads run in two modes. In upsert mode, leaving a product out of the file changes nothing. In replace mode, anything absent from the uploaded file is permanently deleted from Stripe's catalogue. A shop that sends a partial file in the wrong mode deletes its own range. Stripe also warns that uploads process as independent asynchronous tasks with no guaranteed ordering, so firing three files in quick succession can land them out of order.
Connecting to each assistant is deliberate on both sides. You request it from the dashboard, the agent has to accept, and you can disable any single agent later from the same page. That is closer to a wholesale relationship than to open web traffic, and it is the mirror image of the position a shop is in when deciding whether to block a shopping agent altogether.
What do you learn about the buyer, and what stays hidden?
Less than from a normal order, and the gap is structural. Stripe's shared payment token documentation describes what the seller receives: a scoped grant against the customer's payment method, carrying a currency, a maximum amount set to match the transaction total, and an expiry timestamp. Retrieving the token gives you the card brand, the last four digits and those limits. The agent can revoke it, and a deactivation event tells you when the token is spent, expired or withdrawn.
Completed orders carry more. The checkout session exposes shipping address, billing address, SKU, quantities, tax rates and the last four digits, and the dashboard tags each transaction with the agent that sent it. An email address and a phone number appear as optional fields in the approval request, which means they can be absent. Plan for a customer you cannot email.
That reframes the economics. A sale with no contactable buyer is a transaction, not a relationship. It will not grow a list, it will not receive a second order prompt, and the product data you keep for assistants to read becomes the only marketing asset in the chain. For a shop whose repeat rate carries the business, price that in before chasing the channel.
The two four second deadlines that behave in opposite ways
This is the sharpest operational detail in the documentation, and it sits across two pages, so nobody states it in one place. Stripe's guide to managing an agentic integration describes two optional hooks, both with a four second timeout, whose failure behaviour is inverted.
The price and availability hook is called just before checkout completes. If it times out, Stripe falls back to your feed data and the sale continues. The order approval hook is your veto, where you approve or decline each purchase. If that one times out, Stripe declines the payment.
Read together: switching on the safety feature that lets you refuse an order also hands a slow server the power to refuse orders for you. Stripe adds that a non success response returns a 424 to the agent, that some agents retry, and that your endpoint therefore has to be idempotent. A shop without an engineer on call should probably leave the approval hook off and lean on frequent feeds instead. The same document asks sellers to keep checkout failures caused by unavailable items under five percent.
Refunds need no new plumbing. Stripe states that an existing refund flow through Checkout Sessions or Payment Intents works unchanged on agentic orders, and that a payment confirmed with a shared token behaves afterwards as though you had supplied the payment method yourself.
Why one setup reaches several assistants
Because the standard is shared. Stripe's page on the Agentic Commerce Protocol states that it is an open standard created by Stripe, OpenAI and Meta, covering agentic checkout, cart and feed, delegated payment, delegated authentication over OAuth 2.0, and order webhooks for confirmation, shipping, delivery and refunds.
Three companies that would normally build three incompatible systems wrote one. For a small seller that is the good news in the story: the catalogue discipline you build for one assistant survives the arrival of the next. It also explains why the Muse checkout appeared fully formed rather than as an experiment.
The shelf will recommend things that do not exist
Modern Retail's reporting on what it calls the AI shelf carries the most uncomfortable finding of the week. Tech writer Deana Burke created a completely fictional natural deodorant brand and got the assistants to recommend it. The piece, published 26 September 2026, frames the shift as brands working out how to be surfaced in ChatGPT and Gemini results, with machine readable product information as the lever.
Put that beside the checkout news and you get the thing neither article says on its own. The recommendation layer can be populated by a brand that does not exist, and the payment layer is now good enough to complete a purchase from it. Nothing in the token design checks that a product is real. It checks that a seller profile is scoped, that an amount is bounded and that a clock is running.
For an honest shop this cuts two ways. Your competition on the shelf includes things invented last Tuesday. And the trust you built elsewhere, the reviews and the corrections you make when an assistant states something wrong about your business, is what separates you from a plausible name with a tidy feed.
When does the money actually land?
Immediately, unless you change it. Stripe captures agentic payments as soon as the purchase completes, and the manage page describes a setting that switches capture to manual, after which you call capture on the payment intent yourself.
For a shop selling finished stock off a shelf, leave it alone. For anyone making to order, manual capture is the single most useful switch in the whole setup. It lets an assistant place an order, holds the authorisation, and leaves you to take the money when the work is actually starting. A maker with a three week queue who captures on day one has taken payment for something they have not begun, and the refund conversation that follows is worse for having no human on the other end to explain it to.
Promotions are the other lever the documentation flags, and it is unusually direct about the effect: sellers with active promotions consistently see higher conversion rates than equivalent products without them. The promotion feed takes discount codes, percentage deals, fixed amounts and free shipping, targeted at single products or the whole catalogue, applied automatically or behind a code. Stripe ignores promotions that do not match an active product id, which is the quiet failure mode. A code aimed at a SKU you renamed does nothing and reports nothing.
What you can see, stage by stage
Because the buyer detail is scattered across three documents, it helps to lay out what exists at each point in the sale. This is assembled from the seller guide, the token page and the hook payload, and it is the table worth keeping:
| Stage | What you receive | What is missing |
|---|---|---|
| Approval request, before the charge | Line items, amounts, tax rates, billing and shipping details, card brand and funding type, optional email, optional phone | Any guarantee that email or phone is there at all |
| Payment token | Currency, maximum amount matching the total, expiry timestamp, card brand, last four digits | The underlying card, the wallet, the buyer identity |
| Completed checkout session | Order date, SKU, quantities, unit prices, tax, shipping and billing address, receipt number, the agent that sent it | A relationship you can market to later |
| After the sale | Refunds and disputes through your existing flow, deactivation events on the token | A way to reach the buyer that does not go through the agent |
One practical consequence: the approval request is the earliest and richest view you get. If you do run an approval endpoint, log its payload even when you approve everything. That log is the only place some of those fields will ever appear together, and it costs nothing to keep.
Where does this not work yet?
Outside the listed countries, and that list is worth reading rather than assuming. The suite covers the United States and Canada plus a European set that runs from Austria and Belgium through to Sweden, Slovenia and Slovakia, taking in Switzerland, Norway, Gibraltar and Liechtenstein along the way. Australia, New Zealand, Japan, Brazil, India and the whole of Africa are absent as of today.
That matters for two kinds of shop. If you sell from an unlisted country, this channel is not available to you regardless of how good your feed is, and the effort belongs elsewhere for now. If you sell from a listed country to buyers in unlisted ones, nothing here changes your shipping problem, and the tax and duty questions that already come with a cross border parcel arrive unchanged. The assistant makes the sale easier to start. It does nothing about the part that goes wrong at a border.
Is this worth your attention this month?
For most one person shops, not as a project. As a decision it is worth an afternoon.
The cheap half is housekeeping you probably owe anyway. A return policy written and published. Tax codes that are genuinely set. Titles and descriptions that read as product facts rather than atmosphere. Stock numbers true within the quarter hour. Every one of those pays off on your own site too, which is the test of whether a channel is worth preparing for: if the work is wasted when the channel fails, skip the channel.
The expensive half is the automation around the feeds, and that is where a small shop should wait. A fifteen minute inventory push means a system that knows your stock in real time. If your stock lives in a spreadsheet you update on Sunday, the feed will lie to agents and your failure rate will sit above the five percent Stripe asks for. Fix the stock record first. A shop that owns the code behind its own storefront and database can write that push once and point it at whatever surface comes next, which beats renting the connector from whoever is fashionable this quarter.
The honest summary: an assistant you did not choose can already pay you with a card you already accept, and it can already recommend a competitor that does not exist. The first fact is a small opportunity. The second is why your product data and your reputation are now the same asset.