Agentic commerce: the three camps behind UCP, ACP and MCP
Executive summary
Online retailers and marketplaces should connect to an agentic commerce protocol now. AI agents have started choosing and paying for products inside assistants, and a seller's visibility depends on which protocols it supports. ChatGPT Instant Checkout already completes purchases inside the assistant, and JD Sports became the first UK retailer to launch on Stripe's Agentic Commerce Suite. Three camps set the rules: Google's UCP and AP2 (UCP public since January 2026), OpenAI and Stripe's ACP, and the Visa and Mastercard card networks, with MCP and A2A shared across all three. Discovery, marketplace governance, agent-layer brand safety and attribution all change at once, so retailers should pick one camp, audit top product feeds, and measure agent traffic before volume scales.
The customer journey no longer starts at Google or your company’s landing page. The first online touchpoint has moved. Today a shopper looking for running shoes can ask ChatGPT and pay inside the chat. Agentic commerce is already live in the United States, with autonomous AI agents that research, compare and buy on a customer’s behalf. Three camps are now wiring up the protocol layer for the rest of the world.
That layer decides which retailers and marketplaces show up when an agent does the shopping. UCP, ACP and MCP are the standards in play. For UK and European retailers and marketplaces the work gets practical fast. There’s a camp to back, a funnel that warps when the buyer is software, and a brand to keep recognisable while an AI describes you instead.
McKinsey models USD 1 trillion in US B2C agentic revenue by the end of the decade. The plumbing being built today decides who captures it.
What changes when an agent does the shopping
When an AI agent runs the buying, most of the customer journey happens inside an AI assistant. The shopper sometimes never sees a product page.
Live examples already exist. ChatGPT Instant Checkout, also known as ChatGPT shopping, puts the whole purchase inside the conversation. A user picks something from Etsy or Shopify and pays without leaving the chat. Perplexity Pro runs its own verified-merchant shopping flow. Mastercard’s Agent Pay does the same on the card network using signed permissions. JD Sports went further. The UK retailer became the first to launch on Stripe’s Agentic Commerce Suite at NRF this year, with rollout starting in the US and the UK and Europe to follow.
The retail front door now lives inside ChatGPT, Gemini, Claude or Perplexity. Whether your product surfaces in the answer depends on the protocols you support.
The three camps wiring up agentic commerce
Google’s camp: UCP and AP2
Google’s approach is open standards. The Universal Commerce Protocol (UCP) acts as the shared language between agent, merchant and payment provider. The specification has been publicly available since January 2026. UCP is co-developed by Google with Shopify, Etsy, Wayfair, Target and Walmart, and endorsed by Mastercard, Stripe, Visa, Adyen, Carrefour, Zalando and Shopee, among others.
AP2, the Agent Payments Protocol, sits next to UCP. It handles the signed permission slip a network needs before an agent pays.
OpenAI’s camp: ACP
OpenAI and Stripe maintain the Agentic Commerce Protocol (ACP). It defines how an agent runs a full checkout for a buyer, from cart through payment and authentication. It ships inside ChatGPT, which is where most retailers will first see agentic commerce in action.
The card-network camp: Visa and Mastercard
Visa’s Trusted Agent Protocol identifies the paying agent on the card rails most retailers already use. Mastercard contributes two pieces of the trust stack: Agent Pay identifies the agent on Mastercard’s network, and Verifiable Intent, co-developed with Google, creates a tamper-proof log of what the user actually authorised. Both are AP2-compatible. Neither protocol replaces UCP or ACP. They make the payments side enforceable where money moves.
Underneath all three sits MCP, the Model Context Protocol agents use to read product feeds and call commerce APIs. Alongside it, the A2A protocol handles agent-to-agent messaging. Both are shared plumbing every camp uses, not a fourth competitor.
Practical tech mix
UCP and ACP both touch checkout, but they coexist. Stripe contributes to UCP and also ships ACP. For most UK and European online retailers, the practical mix is UCP plus AP2 for the Google side, ACP for the OpenAI side.
What this changes for e-commerce and marketplace teams
Four pieces of the buying journey change at the same time. Each one sits with a different team. Alongside SEO and website optimisation, AI assistants become the next channel to own.
Discovery and brand visibility
Without ACP or UCP integration, agents fall back to scraping your website. Agents still find your "pod coffee maker", but with thinner data and weaker placement than competitors who connected the protocols. Catalogue hygiene, structured data and API quality become brand-visibility levers.
Marketplace governance at platform level
A retailer protects one catalogue. A marketplace protects thousands at once. Every seller feed is a potential agent-visibility failure, and the platform operator carries the gap. Rules cannot be solved seller by seller. They have to live at the platform level. Feed standards, agent-identity defaults and brand-safety guardrails apply across the listings before sellers ever opt in. The work is closer to setting marketplace policy than fixing a product page.
Brand safety in the agent layer
The agent decides which substitutes to suggest, how to describe a product. Without rules at that layer, it can quietly recommend a competitor’s discount item and replace a hard-won category win. Codify these rules the way you would a paid-search ad-copy policy. Define acceptable substitutes. Set which review themes an agent can summarise and which must be quoted verbatim. Treat the agent’s tone as a brand asset with a named owner.
Measurement and attribution
The familiar funnel assumes the buyer visits the site. When the buyer is software, the referrer disappears. Session data thins out. The attribution model built around organic, paid and direct stops mapping to reality. A converted basket might leave no trace beyond a single API call. Some teams have already started treating agent-driven traffic as a fourth channel, instrumenting the API layer for it directly. Agent identity goes in the request, with a separate dashboard and revenue line. Build the dashboards before the volume scales. Without revenue tied to agent traffic, leadership will keep calling this a curiosity, not a channel.
Getting ready for agentic commerce?
Read our whitepaper on how AI is transforming commerce, and what an AI-first strategy looks like for retailers and marketplaces.
How retailers and marketplaces can prepare
Five moves while the protocols stabilise.
- Audit the product catalogue for agent readability. Test your top 50 product detail and category pages with an MCP-compatible probe. Marketplaces run the same audit on seller feeds and fix the worst offenders first.
- Decide which camp to pilot first. UCP plus AP2 connect you to Google’s agent side. ACP connects you to ChatGPT’s. Some platforms have already chosen for you. Shopify turned Agentic Storefronts on by default for millions of US merchants this spring. Most teams will end up supporting both, but pick one to start.
- Build an agent-identity layer. Whenever your own AI features speak for the brand or the platform, decide upfront how the agent introduces itself, what it can promise on your behalf and what it cannot.
- Codify brand-safety rules at the agent layer. Define which substitutes, descriptions and review summaries an agent may use. Marketplaces apply this twice. Once across the platform, and again as a setting sellers configure.
- Stand up the agent-traffic dashboard before the volume scales. Without baseline data you cannot tell if you are winning or losing the conversation.
The first-mover case for agentic commerce
The three camps will keep shifting until at least one goes live in Europe. That does not change the agenda for online retailers and marketplaces. Start now. Connecting to a protocol while it is still developing teaches you what transfers to the others. You learn how to design clean APIs, which product data agents actually need, and how a brand is experienced inside an AI assistant.
The teams that wait will spend 2027 catching up to brands that AI assistants already know. First movers set them.
FAQs about agentic commerce
Agentic commerce is online buying done by an AI agent on a customer’s behalf. The agent searches, compares and pays through a protocol such as ACP, UCP or AP2. The customer often never opens a product page.
Yes. Live US consumer flows already run, and the first UK retailer has launched on a public protocol. The protocols are open. What is uncertain is the European launch date, not whether agentic commerce works.
Traditional e-commerce assumes a person visits a storefront and clicks to buy. Agentic commerce hands that work to an AI agent. The agent compares across shops and pays through a protocol. The visible buying step shifts from your website to the inside of an AI assistant.
