Three open standards are competing to define how AI agents discover, evaluate, and transact with businesses in 2026 — and most brands have not yet decided which ones to implement, or why. ACP (Agentic Commerce Protocol, from Stripe and OpenAI), UCP (Universal Commerce Protocol, from Google and Shopify), and MCP (Model Context Protocol, from Anthropic and the Linux Foundation) each solve a different slice of the same problem: making your business queryable by AI agents rather than only by humans. Picking the right one — or the right combination — depends on which AI systems send you traffic and what kind of transaction you are trying to enable.
This is not a theoretical question for 2027. ACP went live in production in late 2025 with PayPal and Worldpay as payment partners. UCP launched publicly at NRF 2026 with Google, Shopify, and Walmart as founding partners. MCP crossed 97 million monthly SDK downloads in mid-2026. The protocol layer is live, and businesses without any implementation are already absent from the agentic commerce layer that runs on top of it.
What the Three Protocols Actually Do
The easiest way to understand the distinction is by what each protocol connects to.
ACP is a centralized discovery layer for ChatGPT commerce. Merchants submit product catalogs to OpenAI’s platform, and those catalogs become queryable inside ChatGPT when users shop or ask product-related questions. ACP handles the discovery query (“which of these products meets my requirements?”) and hands off to the merchant’s own checkout for the transaction. Because it routes through OpenAI’s platform, ACP is effectively ChatGPT’s answer to “how does a brand get into our agent’s product knowledge.” Launch partners include Wayfair, Revolut, and Salesforce.
UCP is a decentralized protocol where merchants host product data on their own servers and any compatible AI agent can discover and query it. Google and Shopify co-developed UCP with Walmart at NRF 2026 to cover the full commerce lifecycle: product discovery, catalog browsing, cart management, checkout, and post-purchase support including returns and order tracking. Unlike ACP, UCP does not route through a single platform — it is designed so that any AI agent (Google, ChatGPT, Perplexity, a third-party shopping assistant) can query a merchant’s UCP endpoint directly, provided the endpoint is discoverable.
MCP is the general-purpose layer. Created by Anthropic and now governed by the Linux Foundation’s Agentic AI Foundation, MCP provides a standard for connecting AI models to any external tool or data source — not just commerce. A business with an MCP server can expose product information, a FAQ, a booking endpoint, or any structured capability to any AI agent that supports MCP. It is the broadest protocol and the least opinionated about what you expose. As of mid-2026, 41% of software organizations have MCP servers in production, and the public registry at mcp.so lists over 10,000 servers.
Who Each Protocol Serves Today
The practical audience for each protocol is different enough that “which one should I implement” is really a question about which AI systems matter most to your business right now.
ACP is the right starting point if ChatGPT is your primary AI traffic source and your business model involves product transactions. ChatGPT’s 1 billion active users (as of May 2026) represent the largest single pool of AI-assisted shopping intent. Brands with ACP catalog submissions appear in ChatGPT product responses and instant checkout flows. Brands without submissions do not. If your analytics show ChatGPT referral traffic, ACP converts that traffic into structured agent visibility instead of relying on ad hoc citations.
UCP is the right priority if you sell through Shopify, if Google Shopping is a meaningful traffic source, or if you want to be discoverable by any AI agent rather than just ChatGPT. UCP’s decentralized architecture means a single endpoint implementation covers multiple AI systems simultaneously. The trade-off is higher implementation complexity — you are hosting and maintaining your own UCP endpoint rather than submitting a catalog to a platform. For Shopify merchants, the UCP integration is significantly simplified through Shopify’s native support.
MCP is the right first step if you sell services rather than physical products, or if you want agent visibility across all platforms before choosing between ACP and UCP. An MCP server with product info, FAQ, and contact endpoints covers Claude, any MCP-compatible AI assistant, and the growing set of agentic workflows built on the MCP ecosystem — without requiring the commerce-specific infrastructure that ACP and UCP assume.
The Walmart Approach: Why the Answer Is Usually “Both”
Walmart’s implementation is instructive for brands large enough to run multiple protocols. In October 2025, Walmart launched ChatGPT Instant Checkout through its ACP partnership with OpenAI, making its catalog queryable inside ChatGPT. At NRF 2026, it announced Gemini integration through UCP, covering Google’s AI surfaces. This is what protocol analysts call the “yes, and” approach — implementing both centralized and decentralized layers because AI agent traffic is fragmented across platforms.
For brands without Walmart’s resources, the equivalent is a staged approach: MCP first (broadest coverage, lowest implementation cost), then ACP or UCP depending on where your agent traffic is coming from, then the other when resources allow. Adyen’s June 2026 announcement of Adyen Agentic — a modular API suite that supports UCP, ACP, AP2, and Meta’s AI checkout simultaneously — signals that multi-protocol support is becoming infrastructure rather than a competitive differentiator.
The Minimum Viable Agent Endpoint for Non-Enterprise Brands
For businesses that cannot immediately implement ACP or UCP, the minimum viable path to agent discoverability runs through MCP. The three capabilities that produce the most agent interactions for a typical SaaS or service business:
get_product_info — Structured data about your offering: name, description, use case, pricing tiers, key features, target customer. This is what an AI agent queries when it is evaluating whether to recommend you. If this endpoint doesn’t exist, the agent has to infer the answer from your website text, which is less reliable and slower.
get_faq — Answers to the questions that appear most often in AI queries about your category: “Does it integrate with X?”, “Is there a free trial?”, “What’s the minimum contract?” Agents use these answers to respond to user questions before the user has expressed intent to purchase — early-stage visibility that can persist through the buying process under high-reasoning AI modes.
initiate_contact — A booking link, demo scheduling URL, or contact endpoint. This is what turns agent visibility into a lead. For agent workflows that are executing a task (not just answering a question), this endpoint is the conversion point.
The Anthropic MCP SDK (available via npm or PyPI), FastMCP, or the growing set of no-code MCP builders can expose these three capabilities in a day without custom infrastructure. The server should be registered on mcp.so and accessible via a /.well-known/mcp.json path on your domain for discoverability.
Discovery Is the Prerequisite
Having an endpoint does not automatically produce agent traffic. Both ACP and UCP require active submission or registration steps. MCP requires listing in registries and having your server referenced in sources AI agents learn from — documentation, registry entries, and mentions in content that AI training pipelines index.
The MCP roadmap for late 2026 includes work on MCP Server Cards, a proposed standard for exposing server metadata via .well-known URLs so AI crawlers and registries can discover capabilities without requiring a developer connection. This would make MCP discovery work more like how search engines discover sitemaps — passively, through crawl rather than active submission. Until Server Cards are live, active registration on mcp.so and documentation that AI systems can find remain the primary discovery paths.
For ACP, discoverability is handled through OpenAI’s merchant platform submission process. For UCP, discoverability relies on the same entity signals that affect chatbot-mode AI visibility — Wikidata entries, schema markup, and cross-platform brand mentions that help AI systems identify your domain as a known merchant entity before attempting a UCP query.
What Changes When AI Agents Can Transact
The shift from AI answering questions to AI executing transactions changes the commercial stakes of being present in the agent layer. Chatbot-mode AI citations produce brand awareness and referral traffic — valuable, but the user still has to navigate to your site and convert. Agent-mode AI with ACP or UCP integration can complete the discovery-to-purchase cycle inside the agent’s interface, with the merchant’s system receiving a structured order rather than a human visitor.
The Agentic Commerce for WooCommerce plugin appearing in the WordPress repository (new entry in the llms.txt search category as of September 22, 2026) signals that the tooling layer is catching up to the protocol layer. Commerce plugins that expose ACP and UCP endpoints without custom development work are beginning to appear, which will significantly lower the implementation barrier for small and mid-market merchants over the next 12 months.
The Implementation Sequence for 2026
For most businesses starting from zero on agent visibility, the sequence that produces the fastest coverage at the lowest cost:
1. Deploy an MCP server with product info, FAQ, and contact endpoints. Register on mcp.so. Add /.well-known/mcp.json to your domain. This produces broad agent coverage across Claude, MCP-compatible assistants, and agentic workflows, and takes less than a week with existing frameworks.
2. Submit your product catalog to ACP through OpenAI’s merchant platform if you sell physical or digital products and ChatGPT is a traffic source. This is a centralized submission rather than a self-hosted endpoint.
3. Implement UCP if you are on Shopify or if you want decentralized discoverability across Google AI Mode and any other UCP-compatible agent. This requires more implementation work but covers a broader set of AI surfaces with a single endpoint.
4. Connect all endpoints to live data. An MCP server or UCP endpoint that returns stale pricing or discontinued products creates trust failures that affect future agent interactions. Real-time data connections are not optional once agents start transacting.
The window for early-mover advantage is real but closing. The brands that establish agent presence in 2026 — through MCP, ACP, or UCP — will be the ones whose names appear in AI agent responses as trusted, queryable entities when their categories become routine targets for agentic shopping in 2027.
To see where your brand currently stands in AI agent discoverability — and identify which protocol gaps are costing you the most — run a free visibility audit at ai-visibility.llmagnet.com.