ACP vs UCP: which agentic commerce protocol should a retailer build first?

This article goes deep on one decision from my blueprint for agentic commerce architecture. If you are new to the topic, start there.

The Agentic Commerce Protocol (ACP) is an open standard released by OpenAI and Stripe in September 2025 under the Apache 2.0 licence. Stripe’s documentation now lists Meta alongside them as a creator. It defines how an AI agent reads a merchant’s product feed, manages a cart, creates and completes a checkout session, passes a payment token and, through OAuth 2.0, acts on a buyer’s behalf with the merchant. The latest published specification at the time of writing is dated 17 April 2026.

The Universal Commerce Protocol (UCP) is an open standard Google announced in January 2026 with Shopify, Etsy, Wayfair, Target and Walmart, also under Apache 2.0. A merchant publishes a manifest at /.well-known/ucp that declares its capabilities: checkout, identity linking, order and payment token exchange, with extensions such as discounts and fulfilment. Capabilities can be offered over REST, the Model Context Protocol (MCP) or agent-to-agent (A2A) transports. In April 2026 Amazon, Meta, Microsoft, Salesforce and Stripe joined its tech council.

Both keep the retailer as merchant of record. Both use tokenised payment rather than raw card details. Both can be reached over MCP. On paper they look like competitors; in practice they are front doors to two very different shopping surfaces.

Any honest comparison has to start with the most important event in this market so far. ChatGPT Instant Checkout launched on 29 September 2025, powered by ACP, with Etsy sellers first and over a million Shopify merchants announced as “coming soon”. Six months later OpenAI changed course.

OpenAI told TechCrunch in March 2026 that “the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide.” ChatGPT shifted towards product discovery (side-by-side comparisons, images, prices and reviews) and sends shoppers to the retailer’s own checkout or to apps inside ChatGPT. ACP continues to power how participating merchants’ product data reaches ChatGPT.

Trade press reported that only a handful of Shopify merchants had actually gone live before the change. I would treat those figures with caution, but the direction is not in doubt, and it holds a lesson for anyone planning a build: the protocol was not the bottleneck. Operations were. Live prices, accurate stock, fraud controls, returns and customer service all have to work when an order arrives from somewhere other than your own website.

Status reflects public announcements up to October 2026. Both specifications are versioned; check the current release before building.

The right sequence depends less on the protocols than on your platform, your channels and where your customers already shop. These are the four situations I see most often.

For a retailer on its own platform, the protocol adapter is the small part. This is how I would estimate the work for a mid-sized UK retailer, assuming the core commerce services do not yet exist as APIs.

One detail catches teams out. When checkout happens on the agent’s surface, your website’s analytics tags do not run. Shopify’s notes on Google’s native checkout, for instance, say client-side pixels do not fire and only server-to-server events are sent. Plan server-side order attribution before the first agent order arrives, or the channel will look smaller than it is.

Efforts are relative, not quotes. The point of the table is the shape: if the services exist, supporting a second protocol is a modest extension. If they do not, choosing between ACP and UCP is the wrong question.

The payments behind both protocols are covered in agentic payments explained, and verifying which agent is calling is covered in Web Bot Auth and the Trusted Agent Protocol. Being recommended in the first place is the subject of generative engine optimisation for ecommerce.