What is agentic commerce?

Agentic commerce is online retail in which an AI agent searches for products, compares them, and completes the purchase on behalf of a person. The person only states a goal, such as “waterproof hiking boots under 150 dollars”; selection, price comparison, and payment are handled by software. The buying decision therefore happens inside a system that reads product data by machine before any human opens the product page.

Term profile at a glance

How does an AI agent buy something?

An AI agent buys in 4 steps: it understands the request, finds matching offers, builds a cart, and triggers the payment. The person stays the instance that issues the request and confirms the purchase; the research in between runs without them.

1. Understand the request: The user states a goal along with conditions, meaning product type, price range, delivery time, or brand. The language model breaks that request down into criteria it can check.
2. Find offers: The agent pulls product data from feeds, interfaces, and web pages. What it can read there about price, availability, and attributes decides whether a store enters the shortlist at all.
3. Build the cart: Through a standardized interface, the agent creates the cart at the merchant and retrieves the binding values for shipping costs, taxes, and final price.
4. Trigger the payment: The user confirms, and the agent hands a single-use payment token to the merchant. The agent never sees card details.

Step 2 is where most stores drop out today. An agent evaluates in seconds whatever is available as machine-readable data; information that only a human eye picks up on the rendered page counts as missing.

Which standards govern purchases made by AI agents?

3 open protocols govern purchases made by AI agents. The Agentic Commerce Protocol handles checkout, the Agent Payments Protocol handles payment, and the Model Context Protocol handles access to tools and data.

• Agentic Commerce Protocol (ACP): OpenAI and Stripe published it on September 29, 2025 under the Apache 2.0 license. It describes how a merchant exposes a product catalog, how an agent creates a cart, and how payment data is passed along as a token. ACP powers instant checkout in ChatGPT.
• Agent Payments Protocol (AP2): Google introduced it on September 16, 2025 with more than 60 partners, among them Mastercard, American Express, PayPal, and Adyen. AP2 works with 3 signed mandates for intent, cart, and payment, issued as W3C Verifiable Credentials. That makes it possible to prove afterwards what a user actually authorized their agent to do.
• Model Context Protocol (MCP): Anthropic developed the standard for connecting agents to tools and data sources. In retail, MCP provides the channel an agent uses to query inventory, price lists, or order systems.

The 3 protocols govern how a transaction runs. The question before that stays open: Whether an agent considers a store at all is decided by the data the store itself publishes.

What does a store have to publish as machine-readable data?

A store needs 3 data layers for AI agents. The first one is a complete product feed, the second one is structured data in the source code of every product page, the third one is access for the crawlers that collect both. If one of the 3 layers is missing, the chain breaks at exactly that point.

Beyond the required fields, the optional ones decide whether an offer wins the comparison. Anyone who supplies gtin, sale_price, shipping_price, accepts_returns, return_deadline_in_days, star_rating, and review_count hands the agent the criteria people filter by anyway. Without a return window in the feed, an offer quietly drops out of the question “Which seller takes it back?”

A second point regularly costs stores their visibility: Agents do not execute JavaScript. Price, availability, and variant have to be present in the delivered HTML, otherwise they stay invisible. The same rule has applied for years in JavaScript SEO and now reaches product data.

For the 3 layers to line up, every value has to read the same everywhere. If the feed says 149 dollars, the Schema.org markup says 159, and the cart says 154, an agent discards the offer as unreliable. This consistency is the first thing we check when we work on visibility in AI answers, because it takes little effort and decides every recommendation that follows. The groundwork is explained in the entry on structured data.

Who is collecting the data shows up in the access log. Which AI crawlers are out there and which of them you allow on purpose belongs to the preparation; an llms.txt adds a map through the catalog.

What is the difference between agentic commerce, agentic SEO, and ChatGPT shopping?

The 3 terms differ in whose agent is acting: in agentic commerce the agent works for the buyer, in agentic SEO it works for the website owner, and ChatGPT shopping is a single interface in which agentic commerce takes place.

ChatGPT shopping is the best-known case of agentic commerce and still covers only part of it. Perplexity, Google, and Microsoft are building their own shopping features, and every interface brings its own requirements for feed and checkout. How content shows up in ChatGPT answers in the first place is covered in the guide to ChatGPT SEO.

For merchants the terms belong together: An agent recommends what it understands. The brand has to be recognizable as an entity of its own, and the products have to exist as clean data. Both together decide whether a store shows up as a seller in AI search.

Whether the effort pays off shows in the conversions from agent channels, once a store tracks them separately. That separation is worth setting up before the first agent orders, because it cannot be reconstructed after the fact.

Frequently asked questions about agentic commerce

Are AI agents already buying in Europe?
Purchases completed by AI agents are still at an early stage in Europe. Instant checkout in ChatGPT launched in September 2025, initially for buyers and merchants in the United States, while European stores mostly appear through product search and recommendations in AI answers. According to a Bitkom survey from April 2026, 61 percent of retailers see a competitive advantage in using AI.

Is my Google Shopping feed enough for agentic commerce?
A well-maintained Merchant Center feed is a usable foundation, because the ACP spec supports a Google-compatible mode. Field names and conventions differ there, however, and details about the seller and about checkout eligibility are added on top. Plan the feed as a stage of its own.

Will my website lose visitors to agentic commerce?
Some purchases will be completed without a page view, because the agent processes the product data directly. Sessions and page views then reflect success less accurately. Anyone who tracks orders from agent channels as a separate metric early on spots the shift before it surfaces in overall traffic.