What is agentic commerce?

An AI agent that compares prices or fills out a form is already touching your website. Agentic commerce goes a step further: the agent completes the entire transaction, from finding the product to paying for it, without the person having to open the merchant's site themselves. For the merchant, that means part of its sales can close without anyone ever visiting its product page the usual way, with their own clicks and their own cart.

There's no single standard for this yet. Two open protocols are competing for the same ground. The Agentic Commerce Protocol (ACP) was developed jointly by OpenAI and Stripe, released on September 29, 2025 under the Apache 2.0 license, and it's what ChatGPT already uses for its Instant Checkout feature. The Agent Payments Protocol (AP2) is led by Google, announced on September 16, 2025, backed by more than sixty payments and technology companies, including Mastercard, PayPal, and Adobe; its standardization keeps moving forward within the FIDO Alliance.

Both protocols solve the same underlying problem: how to let an agent buy something without handing it the person's real payment credentials, and how to keep a record that this person authorized that specific purchase.

Under OpenAI and Stripe's ACP, the agent reads the merchant's catalog, in a format built for machines, and builds a cart. Once the person confirms what they want to buy and with which payment method, Stripe issues a Shared Payment Token: a single-use, time-limited identifier tied to one specific merchant and one specific amount. The agent sends that token along with the cart to the merchant, who charges it like any other payment; the agent never sees the actual card number at any point. Before that step, the protocol includes a capability negotiation: the merchant declares which payment forms it accepts (Stripe's token, another provider's own token, or the seller's own handler), and the agent picks one of the ones the merchant allows. That leaves the merchant with real control over which methods reach its checkout, even though the whole prior conversation happened inside an AI assistant.

Google's AP2 solves the same problem with a different mechanism: mandates. An Intent Mandate is signed when the person makes the initial request, with whatever conditions they set (price cap, time window, allowed merchants). When the agent assembles the final cart, a Cart Mandate is generated that locks in the exact items and price the person saw and approved. A Payment Mandate, derived from the two prior ones, links the payment method to that chain of approvals, and flags whether a person was present at the moment of payment or whether the purchase was carried out under rules set in advance, without the person present at that instant.

The two mechanisms differ in form but share the same idea: the agent never decides to spend money on its own without leaving a verifiable trail that a person authorized it, whether through a limited token or a chain of signatures.

For a store that already sells well through Google and social media, this is the first concrete GEO question with a direct revenue effect: if a customer asks their assistant to buy something for them, can your store take part in that purchase, or does it get left out?

There's no public figure comparing how much volume each protocol moves today; any claim about which one is winning would be a guess, not a verified number. What does depend on a concrete decision is catalog readiness: the structured product data an agent needs to find and correctly describe an item is the same, whichever payment protocol gets chosen later. Putting off that preparation until it's clear which protocol wins in two years keeps a store out of agentic commerce during that whole stretch, whoever ends up winning.

There's a second, more operational decision: what happens with a chargeback when an agent triggered the purchase. ACP's token expires and is single-use, which limits the room for a claim if something goes wrong after the charge; AP2's mandate chain leaves a signed record of every step, built specifically to reconstruct what the person authorized if a charge needs investigating. A merchant that turns on either protocol without first deciding how to handle those disputes is putting off that decision until an angry customer is already on the line.

• Publish a machine-readable product feed before deciding which payment protocol to implement; that work serves both.
• Don't bet all your development on a single protocol while neither one clearly and verifiably dominates the market.
• Check whether your current payment processor already supports ACP or AP2 before building your own integration from scratch.
• Require the buyer to explicitly approve the final cart, even when the agent handled all the prior search and comparison.
• Keep the record of every agentic transaction, whether token or mandate, so you can respond to a chargeback.
• Keep the same fraud and stock rules you use at human checkout; no protocol replaces them on its own, no matter how reliable the token or mandate looks in a given case.

• Treating ACP and AP2 as if they were interchangeable or compatible with each other; they are two separate protocols with no official bridge.
• Assuming an agent can complete a purchase with no human approval at any point in the process.
• Exposing real payment credentials to the agent instead of the token or mandate the protocol requires.
• Ignoring agentic commerce on the assumption it doesn't generate volume yet, without checking that against your own store's data.
• Confusing agentic commerce with a customer-service chatbot that only recommends products without being able to charge for them.