What is agentic commerce? What changed in 2025-2026 and what a US store should do now

In short: agentic commerce is selling in which an AI agent does part or all of the buying work on behalf of a person: it searches, compares, builds the cart and, when the user authorizes it, pays. As of October 2026 there are open protocols for this (the Agentic Commerce Protocol from OpenAI and Stripe, and the Universal Commerce Protocol and Agent Payments Protocol from Google), and agentic checkout is already live in the United States through Google. If you sell online in the US, this is no longer a forecast: a program reads your catalog before a person does, and it chooses using the data it finds. The most useful thing you can do now is not to build a custom integration but to get your product data, Merchant Center account, payment stack and policies in order.

This article explains what agentic commerce is, what actually happened over the last year (including the steps backward), how the protocols work without the jargon, and what to do now if you run an online store in the US.

What agentic commerce is, in a definition that is useful

There are three ways AI can take part in a purchase, and it helps not to mix them up:

1. AI recommends. The shopper asks "what running shoes work for asphalt?" and the assistant answers with options and links. The purchase still happens on the store's site. This already happens every day in ChatGPT, Gemini, Perplexity and Copilot.
2. AI assembles the purchase. The assistant compares price, stock and shipping across several stores, builds the cart and leaves the shopper one click from paying, either inside the chat or in the merchant's checkout.
3. AI buys. The shopper leaves an instruction with limits ("when this water filter drops below 40 dollars, buy two") and the agent completes the payment without the person watching at that moment.

Agentic commerce in the strict sense is the second and third steps: the agent does not just inform, it acts. The first is usually called generative search or AI shopping, and it is the one driving the most traffic today.

The practical difference for a merchant is who makes the intermediate decision. In classic ecommerce, a person looks at your product page, your photos and your reviews. In agentic commerce, a program reads your structured data (price, stock, variants, shipping cost, return policy) and discards whatever is unclear before a person sees anything.

What actually happened between 2025 and 2026

The topic has been buried in predictions. These are the verifiable facts, in order.

Two readings of that table. The first: the infrastructure is moving fast, and it is pushed by the companies that own payments (Visa, Mastercard, Stripe, PayPal, Google). The second: paying inside the chat has not convinced everyone yet. OpenAI explained that the first version of Instant Checkout "did not offer the level of flexibility" it was aiming for, and let each merchant charge in its own checkout. What held firm is discovery: assistants read catalogs and send shoppers to stores. And Google, which owns the search results page and the payment wallet, is doing the part that OpenAI scaled back.

How the protocols work, without the jargon

A protocol is an agreement about how two systems talk to each other. Here, the AI agent and your store. There are three you should know by name.

Then there are the card networks, which are not commerce protocols but payment ones. Visa Intelligent Commerce and Mastercard Agent Pay issue special credentials for agents (tokens) that the bank can trace, limit and dispute like any other purchase.

One technical detail matters even if you never write code. In UCP, the store publishes a file at a fixed address on its site describing what an agent can do there (look up products, create a cart, apply a coupon, pay). In ACP, the merchant exposes a checkout through an API. In both cases the merchant is still the seller: you charge, invoice, ship and handle the complaint. The agent is a channel, not an intermediary that resells.

If your store runs on Shopify, a lot of this is meant to be handled by the platform. Shopify is one of the companies that developed UCP, and it offers a channel for publishing your catalog to AI assistants such as ChatGPT, Copilot and Perplexity. Availability varies by assistant, so check your admin panel rather than assuming. WooCommerce and BigCommerce merchants should check what their platform or payment provider offers, and custom stores will need to expose the data themselves.

What an agent reads from your store

Whatever the protocol, the agent needs the same data. OpenAI's product feed specification, for example, requires as a minimum an item identifier, title, description, URL, brand, seller name, image URL, availability and price, and recommends updating price and availability whenever they change. Google asks for an active Merchant Center account, approved products and declared return policies.

In practice, this means the agent compares stores using what you declare, not what your photos or copy say. If the price in your feed does not match the price on your site, if you do not declare shipping cost, or if the return policy is a vague paragraph, the agent has a reason to pick someone else.

What changes for an online store

There are five concrete changes. The first two are already happening to every store.

1. The first reader of your catalog is a program

A share of your visits already comes from AI assistant answers, and those assistants chose which store to show by reading data. A listing with complete attributes (material, dimensions, compatibility, size, warranty) has a better chance of appearing in an answer than a listing with a nice photo and three lines of text. This happens with or without a payment protocol.

2. What used to take an afternoon is compared in seconds

An agent can check ten stores before the shopper finishes their coffee. Total price with shipping, delivery time, return policy and reviews end up in the same table. Differences that used to get lost between browser tabs are now visible.

3. Trust runs through data, not design

In a delegated purchase, the person trusts the agent and the agent trusts your data. A stock count that says available when it is not, a price that changes at checkout, or a shipping cost that shows up at the end are mistakes a human sometimes forgives. A system that buys at scale learns to avoid that store.

4. Checkout is no longer a single screen

With UCP and ACP, payment can start in the chat and finish in your store, or happen entirely in the assistant. Your payment stack and your fraud tools need to recognize agent-initiated purchases and not block them as if they were malicious bots.

5. After-sales becomes part of the product

If the agent made the purchase, the customer will ask the agent where the order is. The protocols already cover order status and returns. A store that does not report shipping in a structured way looks worse next to one that does.

What does not change (yet)

A few facts to lower the anxiety.

• Most purchases still end in the store. The retreat from Instant Checkout shows that both shoppers and merchants prefer, for now, the checkout they know.
• Not every merchant wants agents. The Amazon case against Perplexity is still open. The Ninth Circuit reversed the preliminary injunction in August 2026, but the case continues. It is a signal that the rules about who can buy on your site, and on whose behalf, are still being written.
• Outside the US the picture is thinner. Google announced the expansion of its checkout in Gemini to Canada, Australia and the UK, but it is not live there yet. In Latin America there are bank and card-network pilots (Visa and Santander ran agent-initiated payment pilots in March 2026, and Mastercard has reported real agent transactions in the region), not a consolidated sales channel. If you sell to customers in those markets, treat it as something to prepare for, not something that is already driving orders.

Agents that buy for shoppers and agents that sell for you

When people talk about agentic commerce, two different things get mixed together.

The good news is that both sides are prepared with the same work: organized product data, reliable stock and prices, clear policies and systems that can be queried through an API. What you do so that a shopping agent in ChatGPT understands your catalog is the same thing your own agent needs to answer "do you have this model in size 10?" without making something up.

For the seller side, AI agents for ecommerce covers the concrete uses: after-sales, repurchase, abandoned carts and customer service. For the buyer side, the full checklist is in how to prepare your store for AI shopping agents.

What to do now if you sell in the US

Because Google's agentic checkout is live in the US, the question for a US store is not whether the channel will exist but whether your store is eligible and ready. Here is a practical order.

1. Get Merchant Center in shape. You need an active account, approved products, and return and shipping policies declared. Check the diagnostics for disapproved items and price or availability mismatches between the feed and the site. This is the same data that feeds Google's shopping surfaces.
2. Check your checkout eligibility. Google's checkout in Search and Gemini has its own requirements and rolls out by merchant and platform. Look at what Merchant Center and your platform say about your store, and do not assume that being listed means being eligible to sell through the agent.
3. Review your platform's channel for AI assistants. On Shopify, look for the channel that publishes your catalog to assistants and see which assistants are available to you. On WooCommerce or BigCommerce, ask your platform or your payment provider what they support.
4. Talk to your payment provider about agent payments. Stripe, PayPal and Shop Pay are involved in this space, and the card networks issue agent credentials. Ask how agent-initiated purchases are identified in your account, what data you receive, and who is liable in a dispute.
5. Adjust your fraud rules. If your fraud tool blocks any order that does not come from a normal browser session, it may block legitimate agent purchases. Review rules for unfamiliar devices, new IP addresses and unusual checkout speed, and make sure agent-initiated orders are identifiable rather than treated as bot traffic.
6. Make your data trustworthy. Price, stock, shipping cost and delivery time in the feed have to match the site and the checkout, every time. Update them when they change, not once a day if you can avoid it.
7. Expose order status clearly. Tracking numbers, delivery estimates and return instructions should be available in a structured way, because the agent will be asked about them.

A rule worth repeating: if your stock on the site does not match your warehouse, no protocol will help you. That is the first fix.

Where to focus, by store type

You do not need the same level of preparation everywhere.

Risks to keep in mind

Fraud and chargebacks. A purchase initiated by an agent can be legitimate or not. The Visa and Mastercard agent credentials exist precisely so the bank knows what the cardholder authorized. Before accepting agentic purchases, ask your payment processor how it identifies them and who is responsible in a dispute. Also ask how your chargeback evidence works when the shopper never saw your checkout page.

Agent mistakes. An agent can buy the wrong size or an incompatible product. Your return policy has to account for that, and your product pages have to make attributes clear so it happens less.

Platform dependence. If a good share of your sales comes from one assistant, whoever controls that assistant controls your visibility. It is the same lesson Google and the marketplaces taught: keep your own channel (your site, email, SMS, your customer list) as well as being where the agents are.

Personal data. If an agent shares shopper data with your store, you are responsible for that data under the laws that apply to you and to your customer. Review what you receive and what you use it for, and check with your lawyer if you are unsure.

What it costs to get ready

Market reference ranges as of October 2026, for a small or mid-size business. These are not quotes: they depend on your catalog and the systems you already have.

Other processes where an agent of your own pays off, beyond the store, are in AI agents for business: what they actually automate.

Frequently asked questions

What is agentic commerce?

It is commerce in which an AI agent searches, compares and, with the user's authorization, buys on their behalf. It differs from AI recommendation in that the agent carries out actions: it builds carts, applies coupons and pays within limits the person sets.

What is the Agentic Commerce Protocol?

It is an open standard created by OpenAI and Stripe in September 2025 so that an AI agent can start a purchase in a store and pass a payment method without exposing the card data. The merchant remains the seller. Since March 2026, ChatGPT uses it mainly to discover products and sends the payment step to each store's own checkout.

What is agentic checkout?

It is a checkout that an AI agent can start or complete on a shopper's behalf, using a protocol such as ACP or UCP and a tokenized payment credential rather than raw card details. In the US, Google offers it through Universal Cart in Search and the Gemini app, with payment through Google Pay.

Can you buy things inside ChatGPT?

As of October 2026, ChatGPT shows products from stores and sends the shopper to the merchant's checkout. The full in-chat purchase that existed with Instant Checkout stopped being OpenAI's priority in March 2026. Conditions change quickly, so check OpenAI's documentation for merchants.

What is the difference between UCP, ACP and AP2?

ACP (OpenAI and Stripe) focuses on the checkout between agent and merchant. UCP (Google and large retailers) covers the whole journey, from the catalog to order status. AP2 (Google and payments partners) deals with proving that the user authorized the payment. UCP and AP2 are designed to be used together.

Is Google's agentic checkout available in the US?

Yes. Universal Cart has been live in the US in Search and the Gemini app since October 15, 2026, with payment through Google Pay, and Google has announced expansion to Canada, Australia and the UK. Eligibility depends on your Merchant Center setup and your platform, so check both.

Do I need to integrate these protocols right now?

In most cases, no. What is urgent is having complete product data, reliable stock and prices, and clear policies. That improves your visibility in assistants today and leaves you ready to integrate a protocol when your platform offers it or when volume justifies it.

Will AI agents replace online stores?

No. They replace part of the journey: searching and comparing. The store still sells, charges, ships and handles support. What changes is that some of your customers arrive with the decision already made by an agent that chose you by reading your data.

If you want to see what your own after-sales and repurchase agent would look like connected to your store, we build a free working demo in 24 to 48 hours. You can book a call or see how we build AI agents for business.