Agentic Commerce Shop

Finding and comparing products is the easy half of agentic commerce. The hard half is paying: how does a merchant know that an AI agent really has permission to spend someone's money, and how much? Several open efforts started in 2024 and 2025 to answer that. Here they are in brief, with what they mean for a store like this one.

The problem to solve

Online payments were designed around a person typing card details and clicking "pay". When software clicks instead, three questions need clear answers: who authorised the purchase, what exactly did they authorise, and who is responsible if something goes wrong.

The open efforts, in brief

• Model Context Protocol (MCP), published as open source by Anthropic in 2024, is a standard way for AI assistants to connect to tools and data. It is not a payment protocol, but it is one of the common ways agents reach the services they use.
• Agentic Commerce Protocol (ACP), released as open source by OpenAI and Stripe in 2025, lets an agent and a merchant complete a checkout together while the merchant keeps its own payment processing and its relationship with the customer.
• Agent Payments Protocol (AP2), announced by Google in 2025, uses signed "mandates", a record of what the person asked for and approved, so a purchase can be traced back to the person's intent.
• Visa Intelligent Commerce and Mastercard Agent Pay, announced by the card networks in 2025, are programmes for card credentials issued to AI agents, so their purchases can be recognised and limited.

They overlap and are still developing. What they share is the idea that an agent should carry proof of what it is allowed to do, and that the merchant should still own the sale.

What a store needs today

Whichever approach wins, a merchant needs the same foundation: an accurate catalog agents can read, and a real checkout connected to an established payment provider. This store supports Stripe, Square, Clover, Authorize.Net, Elavon, PayPal Payflow and Open Edge, each with a sandbox mode for testing.

aCOMMERCE Shop does not claim to implement any of the protocols above. It demonstrates the groundwork they all depend on. Its payment gateway is being connected; until then checkout stops before payment and no order is taken.

The human stays in charge

All of these efforts keep a person at the centre: you decide which agent to use, set its limits, and approve the purchases that matter. A good agent asks before it buys. A good store makes the facts clear enough that you can check what the agent chose.

More on the basics in Agentic commerce, in plain words.

When you open a product page you see a photo, a price and a button. When an AI agent opens the same page, it looks for something you never see: a block of structured data that states the facts precisely. Here is what that block contains on this store, field by field.

The hidden part of the page

Every product page here includes a JSON-LD script, a small piece of data in the schema.org vocabulary. Search engines have read this vocabulary for years; AI agents read it for the same reason: it removes guesswork. This is the offer of a sample product (since removed) as the store published it on 10 October 2026:

{
  "@type": "Product",
  "name": "65W Three-Port USB Charger",
  "sku": "65W-Three-Port-USB-Charger",
  "offers": {
    "@type": "Offer",
    "price": 39,
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock",
    "itemCondition": "https://schema.org/NewCondition",
    "priceValidUntil": "2026-11-09"
  }
}

What each field tells an agent

• name and description: what the product is, in the store's own words.
• sku: the store's item ID, a stable reference that does not change when the wording does.
• price and priceCurrency: an exact number and an ISO currency code, so there is no confusion between "$39", "39.00" or "USD 39".
• availability: whether the item is in stock, as a schema.org value rather than a phrase.
• itemCondition: new, used or refurbished.
• priceValidUntil: how long the price holds. This store sets it 30 days ahead, so an agent can tell a fresh price from a stale one.

Why the details matter

An agent comparing three chargers across five stores needs the same facts in the same shape from each of them. If one store hides its price in an image or its stock status in a sentence, the agent either skips it or has to guess, and guessing is how agents make mistakes for their users.

One detail to know if you build agents: on this platform the description field currently carries the description's HTML markup (paragraph and list tags), so strip tags before showing it to a person.

See it for yourself

Open any product in the store, view the page source and search for application/ld+json. The full list of what an agent can read here is on the For AI agents page.

Agentic commerce is shopping done by software that acts for you. You tell an AI agent what you need, and it searches, compares and, if you allow it, buys, within the limits you set. This guide explains what that means for shoppers and for stores, and what this demonstrator shows.

From search box to request

Today most online shopping starts with a search: you type, you open tabs, you compare, you decide. With an agent, you start with a request instead, for example "a USB-C charger that can power my laptop and my phone, under $50". The agent does the opening and comparing, and comes back with a short list and the facts behind it.

The decision stays with you. You choose which agent to trust, how much it may spend, and which purchases need your final yes.

What an agent needs from a store

A person can look at a photo and read between the lines. An agent cannot guess safely, so a store has to give it four things:

• A way to be found: stable product addresses, a store page that links to them, a sitemap, and a plain summary such as llms.txt.
• Facts it can read: structured product data in a shared vocabulary. schema.org Product and Offer is the one search engines and AI systems already understand.
• Facts it can trust: data that matches the page a person sees, prices with a validity date, and honest stock information.
• A way to pay: a checkout that software can complete safely, through a payment provider.

What this demonstrator shows

aCOMMERCE Shop is a small, working store built to show this on an ordinary website. Each product page publishes its name, description, picture, SKU, price in US dollars, stock status and condition as structured data. The site has an llms.txt summary, and an AI concierge answers questions on every page.

The cart works, and checkout runs up to the payment step. The payment gateway is being connected; until it is, no order is taken. The For AI agents page lists exactly what an agent can read here.

Why it matters for small stores

Agents favour stores whose facts are clear. That plays to the strengths of a focused shop with accurate product information, as long as that information is published in a form software can read. The good news is that most of it comes from doing the basics well: a clean catalog, honest prices and a working checkout.

Read more in Agentic commerce, in plain words, or see what an agent can read on the For AI agents page.

aCOMMERCE Shop sells a small, deliberate range of digital products: single-file software apps that open in a web browser, and a book in PDF. It is an online store only. The range is small on purpose, and every app is tested before it is listed. Here is what that testing covers, and how you can use the same checks on any store.

A small range, each item with a job

Each app here does one clear thing: plan a budget, track time and invoice it, log daily expenses, check a price before you pay, set a profitable price, organise tasks or keep a simple customer list. A small range means each listing can be checked properly, and each one can say plainly what it does and what it does not do.

Tested the way a buyer uses it

Before an app is listed, it is opened in a real browser window and used like a customer would use it: real typing and clicks, with fictional sample data. The test covers every feature the listing mentions, and the everyday mistakes people make, such as empty fields, unusual numbers and special characters.

• Offline: the app is started with the internet switched off, to find out exactly what still works.
• Your data: anything the app saves is reloaded, backed up, and restored in a fresh browser, to make sure nothing is lost or doubled.
• Screens: it is checked at phone width and on a computer screen.
• Privacy: every network request the app makes is listed, so the listing can say what, if anything, leaves your browser.

This testing finds real problems before a customer does, for example a printout that lost its main figure, or buttons that went blank without an internet connection. Those are fixed before the app goes on sale.

A listing that only says what was tested

Each app’s description is written from its test results. If a feature was not tested, it is not advertised. Limits are stated in the listing itself: what needs an internet connection, where your data is kept, and that the apps were tested in Google Chrome on a computer rather than on a real phone. Every product page also carries structured product data, the same facts in a form AI shopping agents can read, and it is checked as a visitor sees it.

Use the same checks anywhere

Whatever you buy and wherever you buy it, a few questions save trouble: What exactly will I receive, and in what format? What does it need to run? Does it work offline? Where is my data kept, and how do I back it up? If a listing does not answer these, ask before you pay. For this store, the Contact page is the place to ask.