ACP, UCP, and AP2: A Complete Guide to Agentic Commerce Protocols

AI agents are learning to shop on people's behalf. But when an agent walks into an online store, the problems it faces are entirely different from a human's: it can't find the buttons, can't fill out the forms, and can't hand over a bank card. Agentic commerce protocols were created precisely for this: a shared set of rules that lets AI agents discover products, complete payments, and prove to merchants that they're trustworthy.

According to a breakdown by the tech blog Genscore, today's mainstream protocols fall into four families: OpenAI and Stripe's ACP and Google's UCP handle "how to buy," Google's AP2 handles "who authorizes the payment," and the Visa and Mastercard protocols built on Web Bot Auth handle "who is allowed through the door." For small and mid-sized merchants, most of these protocols will be implemented automatically through e-commerce platforms and payment gateways. What merchants really need to prepare in advance is complete, structurally consistent product data.

Why protocols are needed: from HTTPS to agent checkout

A person browsing any online store knows the drill: find the button, fill out the form, enter the card number. An agent can't do that. If every store describes its products and checkout flow differently, the agent has to build a custom integration for each one, and the cost is too high to scale.

Protocols solve exactly this problem. They specify how merchants declare what they sell and how it can be purchased, how agents prove they're acting on behalf of a given customer, and how customers authorize a payment. The original article compares them to HTTPS in the e-commerce era: a set of conventions that lets two strangers complete a transaction. The analogy gets to the heart of the matter—the value of a protocol isn't the technology itself, but the reduction of trust costs between strangers.

Breaking down the mainstream protocols one by one

ACP: OpenAI and Stripe's checkout solution

ACP (Agentic Commerce Protocol) was released as open source by OpenAI and Stripe in September 2025, initially for direct payments within ChatGPT. In March 2026, OpenAI moved payment capabilities back into merchants' own apps, while stating that ACP would remain as infrastructure connecting users and merchants. This pivot shows that "paying directly in the chat box" has progressed more slowly than expected, but the underlying protocol has not been abandoned.

UCP: Google's product publishing standard

UCP (Universal Commerce Protocol) launched in January 2026, developed by Google together with Shopify, Etsy, Wayfair, Target, and Walmart, and backed by more than twenty companies including Visa, Mastercard, Stripe, Adyen, and American Express.

Its approach is straightforward: merchants publish a standard file on their own website declaring which services they offer, how payment works, and where the interfaces are. Any agent can read it without a custom integration. UCP is compatible with AP2 and MCP (the standard for connecting agents to tools), and currently runs in Search's AI Mode and the Gemini app, with payments first launching in the US and Canada, Australia, and the UK already announced to follow.

AP2: the authorization layer for payments

If UCP defines "how to buy," AP2 (Agent Payments Protocol) defines "who gives permission." It works through "mandates"—digitally signed contracts in which the customer specifies what the agent may buy, the spending limit, and a record for every transaction. The April 2026 version added unattended purchase scenarios, such as buying concert tickets the instant they go on sale. Google has handed AP2 over to the FIDO Alliance, the same organization behind passkeys for passwordless login.

Mandates are the core mechanism of AP2, literally meaning "delegation" or "power of attorney." In agentic commerce, they solve a fundamental trust problem: how does a merchant trust that a payment initiated by an AI was actually authorized by the cardholder? Each mandate is a digitally signed contract, explicitly issued by the customer before the transaction, specifying constraints such as the product categories the agent may purchase, the per-transaction amount limit, and the validity period. When a merchant receives a payment request from an agent, it can verify the mandate's signature and terms to confirm the transaction falls within the customer's authorization. After each transaction, a traceable record is generated for later auditing and dispute resolution. This design turns "customer consent" from a verbal assumption into verifiable cryptographic evidence, and ensures that even when an agent shops unattended, there is still a boundary set personally by the customer.

Visa, Mastercard, and Web Bot Auth: who is allowed through the door

Online stores receive thousands of bots every day, and most of them aren't there to buy anything. Visa's Trusted Agent Protocol and Mastercard's Agent Pay are both built on Web Bot Auth, proposed by Cloudflare: agents sign each request with a public key registered in a directory, indicating whether they're browsing or paying, and the merchant (or Cloudflare on its behalf) verifies the signature and lets them through. This lets authorized agents in without opening the door to crawlers.

Web Bot Auth is a bot identity verification specification proposed by Cloudflare in 2025, with the core idea of "cryptographically declaring who I am." Specifically, the agent's operator generates a public-private key pair, registers the public key in a public directory, and attaches a signature to each HTTP request sent to a website. When the merchant, or the CDN provider behind it (such as Cloudflare), receives the request, it can verify the signature against the directory information, confirming that the bot really comes from a known operator rather than being an anonymous crawler. The specification also lets agents declare their intent—for example, "browsing products only" versus "ready to check out"—so merchants can treat them differently: letting shopping assistants through while blocking price scrapers. It builds on existing standards such as HTTP Message Signatures, with the goal of establishing a verifiable, tiered minimum trust between bots and merchants, without resorting to blanket blocking or blanket allowing.

Competition or complement? Both

ACP and UCP overlap heavily—both address "how agents buy in a store"—and behind them is competition between the ChatGPT ecosystem and the Google ecosystem. AP2 and the Visa and Mastercard protocols address a different layer: customer authorization and agent identity, and by design they can be paired with either of the former.

One telling signal: Visa, Mastercard, and Stripe all appear in multiple protocols. According to the original article, this precisely shows that no giant is willing to bet on a single standard.

Some players aren't waiting for protocols to mature at all. A group of personal agents that emerged in fall 2026—Meta's Muse, Grok Bot, and Instinct—shop through a browser like a human, paying with single-use cards from Stripe's Link or with Shop Pay. Shopify accommodates them by opening up its product catalog and checkout through UCP, while Amazon chose to block them.

MCP (Model Context Protocol) is a protocol released as open source by Anthropic in late 2024 for connecting AI models to external tools and data sources in a standardized way. You can think of it as "the USB port for agents": whether the backend is a database, an e-commerce catalog, or a payment API, as long as capabilities are exposed according to the MCP specification, any agent that supports the protocol can call them without writing a separate integration for each service. In e-commerce contexts, MCP often appears alongside UCP—UCP defines how merchants declare their products and checkout flow, while MCP provides the universal channel through which agents read that information and call the relevant interfaces. The division of labor differs: the former is oriented toward business semantics, the latter toward tool invocation. That is precisely the significance of UCP declaring compatibility with MCP: a merchant publishes its catalog once, and it can be read by agents in the Google ecosystem as well as by other agents that support MCP, reducing the cost of redundant work in an era of multiple coexisting protocols.

Practical advice for merchants

The original article provides an action checklist aimed at the Spanish market, most of which applies equally to merchants elsewhere:

• Merchants using major platforms like Shopify only need to watch for platform announcements; protocols will arrive as dashboard settings rather than development tasks.
• Self-hosted stores don't need to implement protocols on their own yet—as of the time of the original article, agentic payments had no generally available solution in Spain.
• What you can do right now is something every protocol will read: complete and consistent product data, including price, inventory, shipping, and return policies.
• Confirm with your anti-bot provider whether it will recognize agents signed with Web Bot Auth.
• The top priority: check whether AI mentions you when asked about the products you sell. No protocol will bring an agent to a store it doesn't even know exists.

Structured data refers to describing product information in machine-readable markup (usually embedded in web pages as JSON-LD) according to public vocabularies such as Schema.org—for example, name, price, stock status, shipping method, return policy, and reviews. Search engines and AI agents rely on these tags to quickly understand what a page is selling, without having to parse messy HTML or take screenshots. For agentic commerce, the significance of structured data is amplified further: protocols solve "how to buy," but the agent first has to know "what is available to buy." Missing product fields, prices and inventory out of sync, and inconsistent descriptions across channels will all cause an agent to skip this store during comparison and recommendation. Therefore, organizing product data into a structurally consistent, fully populated format is foundational work that merchants can complete before protocols are implemented and that all protocols benefit from.

Protocols belong to platforms; recommendations belong to you

The original article's core judgment is worth remembering: in a few years, paying with an agent will be as effortless as mobile payments, and almost no merchant will need to know which acronyms underpin it. Platforms can solve protocol integration for you, but they can't solve "whether AI chooses you." Tools like Genscore measure exactly the step before the protocol—whether ChatGPT, Gemini, and Claude mention your brand when asked your core search questions, and whether your website allows AI crawlers and carries structured data.

This also highlights the real threshold of agentic commerce: protocols are the ticket to entry, but visibility is the gateway to traffic. For merchants, getting product data in order and getting the brand into AI answers is more worthwhile than rushing to implement a particular protocol.

One-line Summary

Agentic commerce protocols fall into four families, but the real threshold for merchants is product data and AI visibility.

Paragraph Summary

The article maps out the protocol landscape for AI agent shopping: OpenAI and Stripe's ACP and Google's UCP handle "how to buy," Google's AP2 handles "who authorizes the payment," and the Visa and Mastercard protocols built on Web Bot Auth handle "who is allowed through the door." ACP and UCP overlap functionally, reflecting the rivalry between the ChatGPT ecosystem and the Google ecosystem, while AP2 and the card networks' protocols position themselves as a neutral layer that can be paired across ecosystems. Visa, Mastercard, and Stripe all betting on multiple protocols at once shows the industry has yet to converge on a single standard. For small and mid-sized merchants, most protocols will be implemented automatically through e-commerce platforms and payment gateways, with no need for self-implementation; what truly needs advance preparation is complete and consistent product data, plus confirming whether anti-bot providers can recognize signed agents. The article emphasizes that protocols are only the ticket to entry—brand visibility in AI answers is the more fundamental gateway to traffic.