ACP, UCP, AP2 and friends: agentic checkout protocols explained

When a human buys online, decades of infrastructure quietly vouch for them: the browser session, the saved card, the device fingerprint, the address on file. When an AI agent buys on a human's behalf, every one of those assumptions breaks. The merchant sees a machine. The card network sees traffic that its fraud models were trained to block. The bank sees a transaction the account holder never directly initiated. Agentic commerce could not scale on handshakes and screen-scraping; it needed standards that let each party verify the others cryptographically.

Between September 2025 and early 2026, the industry produced five protocol families to do exactly that, and the pace tells you how seriously the platforms take the McKinsey-sized prize: agentic AI influencing an estimated 3 to 5 trillion dollars of global retail commerce by 2030. The names blur together, ACP, UCP, AP2, TAP, Agent Pay, but the underlying architecture is tidy once you sort them into layers: what the agent can buy, how the checkout executes, and how everyone proves the human said yes.

There is precedent for how fast this can normalize. Online payments themselves went through the same sequence in the late 1990s: bespoke merchant integrations, then a protocol layer, then invisibility, to the point that no store owner today thinks about how an authorization request travels. The agentic protocols are speedrunning that arc, with the notable difference that this time the retailers co-authored the standards instead of adapting to them after the fact. Walmart, Target and Etsy sitting inside UCP's design process is a structural admission that AI surfaces will be a first-class sales channel, and the retailers wanted the plumbing shaped before it shaped them.

This explainer takes each protocol in turn, then assembles them into a single purchase so you can see where each fires. If you want the wider commercial context first, the market shift is mapped in what is agentic commerce.

The Agentic Commerce Protocol is the standard behind the buy button in ChatGPT. Co-developed by OpenAI and Stripe and released as an open standard on September 29, 2025, it defines the full conversation between an AI agent and a merchant's commerce backend: presenting the item, confirming price and availability, collecting the buyer's approval, and executing payment. Its launch vehicle was Instant Checkout, live for US ChatGPT users with Etsy sellers on day one and over a million Shopify merchants announced as next.

The design choice that matters most to merchants is the shared payment token. The agent never sees raw card credentials; Stripe issues a scoped token authorizing that specific transaction, and the merchant remains the merchant of record. Your fulfillment, your customer relationship, your refund flow all stay intact. ACP also specifies the product feed side of the relationship, the catalog data that determines what ChatGPT can offer, which we cover operationally in how to get your products into ChatGPT Shopping.

Because ACP is open, its ambitions extend beyond ChatGPT: any AI surface can implement it, and Stripe positions it as generic agentic commerce infrastructure. In practice, through mid 2026, ChatGPT is where ACP meets real order volume, and the launch scope remains deliberately narrow: single-item purchases from US sellers, with multi-item carts and more regions on the public roadmap. Merchants should model it as a high-intent, low-friction channel for hero products rather than a wholesale replacement for their cart.

The Universal Commerce Protocol is Google's counterpart, and its authorship is the story. Rather than shipping a spec and inviting adoption, Google co-developed UCP with Shopify, Etsy, Wayfair, Target and Walmart, and gathered endorsements from more than 20 companies across the ecosystem, including Visa, Mastercard, Stripe and Adyen. The specification went public in January 2026. A protocol drafted with the retailers who will run it, and endorsed by the networks that will settle it, starts life with distribution ACP had to earn.

UCP is also broader in scope: it acts as a shared language between agent, merchant and payment provider, covering discovery, publishing what a business sells and how agents may interact with it, as well as the transaction itself. Its first consumer surface arrived in February 2026, when UCP-powered checkout reached Google's AI Mode with Etsy and Wayfair products, following the November 2025 agentic checkout rollout with Wayfair, Chewy and Quince. Shopify, Target and Walmart are named as the next wave. Given AI Mode's position inside the world's largest search engine, UCP is the protocol most likely to touch the broadest shopper base; the surface itself is profiled in our Google AI Mode guide.

Checkout execution answers how the money moves. The trust layer answers a harder question: how does anyone prove the human authorized this? The Agent Payments Protocol, announced September 16, 2025, solves it with cryptographically signed mandates. An intent mandate captures what the user asked the agent to do, with auditable context. A cart mandate signs the exact items and price the user approved. When a dispute arrives six weeks later, there is a verifiable chain from human instruction to executed transaction, which is what banks and merchants need before they will underwrite agent traffic at scale.

The card networks attacked the adjacent problem from their own rails. Visa's Trusted Agent Protocol, launched October 14, 2025, lets an agent present signed identity so a merchant can distinguish a legitimate shopping agent from hostile automation. Mastercard's Agent Pay applies the network's tokenization machinery to agent-initiated payments. Neither asks merchants to write new code; both flow through payment processors and issuers, the same way 3D Secure did. The architecture rhymes across all three: identity and consent become signed, portable artifacts instead of session-based inferences.

Protocol names make more sense in motion than in a glossary, so assemble the stack around a single transaction and the division of labor becomes clear. A buyer tells an assistant to order a specific coffee grinder when it drops below 120 dollars. The instruction becomes a signed intent mandate under AP2. The agent discovers the product and merchant through catalog and discovery machinery, UCP's territory on Google surfaces, the ACP feed on ChatGPT. When the price condition hits, the agent composes the order; the approved cart is signed as a cart mandate. Checkout executes through ACP or UCP depending on the surface. At the payment step, the agent presents verified identity via Visa's TAP or Mastercard's Agent Pay, and a scoped token moves the money without exposing credentials.

Each layer solves a genuine problem, and none of them touches the step that decided the revenue: why the agent surfaced that grinder instead of the four competitors at similar prices. The protocols are pipes. The selection happened upstream, in the AI answer, before any mandate was signed.

Ride your platforms for the pipes. Unless you run bespoke commerce infrastructure, you will meet these protocols through Shopify, your payment provider, or your marketplace. Supporting UCP and ACP through those channels covers the major consumer surfaces: ChatGPT on one side, Google's AI Mode and Shopping on the other. Confirm your processor's roadmap for AP2 and the card network protocols, then stop worrying about the plumbing. Protocol support is becoming table stakes precisely because it is being commoditized by the platforms.

Ask your providers three questions. When are we live on ACP surfaces, and what does our feed need first? When does UCP checkout reach our platform tier as Google expands beyond the launch retailers? And which of AP2, Visa TAP and Mastercard Agent Pay does our processor handle, with what dispute evidence available to us when an agent transaction is challenged? Ten minutes of vendor email replaces a quarter of speculative protocol reading, and the answers usually reveal that the plumbing is closer to done than the discourse suggests.

Compete where there is no protocol. No spec, working group or endorsement list governs which products an AI answer names. That layer is won with accurate product data, presence in the third-party sources engines cite, and content that answers buying questions the way engines quote. It is also measurable, which is where Reachroller earns its keep: it tracks your category's buying questions across engines on a schedule, scores visibility honestly with branded questions excluded, links every score to raw answer receipts, and writes the publish-ready fix for each question you lose, from 29 dollars per month. The pipes are becoming universal. The shelf is still up for grabs, and the brands measuring it now are taking it. Where to start is laid out in what is AI visibility.

Protocol roundups often pad the list with adjacent standards, so it is worth naming what does not demand merchant attention. MCP, the Model Context Protocol, is the general standard for connecting AI models to tools and data sources. Agents may use MCP internally to reach commerce systems, but it is infrastructure for developers building agents rather than something a brand implements to sell. x402 and its relatives target machine-to-machine micropayments, agents paying APIs and services in tiny increments, a genuinely interesting frontier that has nothing to do with whether a shopper's assistant can buy your product this quarter.

The filter to apply is simple: does a consumer surface with real traffic transact through it today? By that filter, ACP and UCP clear the bar, the payment trust protocols matter through your processor, and everything else is a watching brief. Revisit the list quarterly; do not staff against it.

Commerce standards wars have a consistent shape: incompatible early specs, a bridging period where intermediaries translate, then consolidation around whichever standard the distribution owners back. The endorsement lists already telegraph the bridge phase here. Visa, Mastercard and Stripe appear on UCP's endorsement roster while Stripe co-authors ACP and the networks ship their own identity protocols. The payments layer is deliberately hedging across every checkout layer, which tells you the players expect coexistence rather than a winner-take-all outcome.

For merchants, coexistence is fine news, because the translation burden lands on platforms. Shopify already speaks both dialects on behalf of its merchants, which is precisely why over a million of them could join ChatGPT checkout without engineering work. The historical analogy is accepting card networks: a store did not bet on Visa against Mastercard, it accepted both through one terminal, and the terminal vendor absorbed the complexity.

What does not consolidate is the answer layer. Each engine composes its own recommendations from its own retrieval, and no standards body will ever harmonize which brands get named. That asymmetry should drive budget allocation: minimum viable effort on the pipes, sustained effort on the shelf. If you want to see how uneven the shelf currently is, the tooling landscape for measuring it is reviewed in the best AI visibility tools.