The Protocol Wars: UCP vs. ACP vs. MCP in Agentic Commerce

Update, October 2026: OpenAI confirmed in March 2026 that it is ending Instant Checkout in favor of working with retailers on dedicated apps inside ChatGPT, so read the Instant Checkout parts of this post as history.

Everyone keeps asking which protocol will win. UCP? ACP? MCP? I spent three weeks mapping the competing standards, reading IETF drafts, and tracking what actually shipped. The answer I landed on surprised me: the protocol question is a sideshow. The real war is happening underneath it.

TL;DR: Four protocols — UCP (Google/Shopify), ACP (OpenAI/Stripe), MCP (Anthropic), and A2A (Google) — compete to become the HTTP of AI shopping. But OpenAI just killed its flagship checkout implementation after Walmart revealed 3x worse conversion rates. Meanwhile, Visa, Mastercard, and Cloudflare are quietly building the trust infrastructure that will matter more than any protocol: cryptographic identity, behavioral certification, and verifiable intent. We're in the pre-SSL era of agent commerce, and the padlock icon hasn't been invented yet.

Since I started investigating how AI agents actually shop, one thing has become clear: the technical plumbing between agents and stores is the bottleneck. AI traffic to US retail sites grew 805% on Black Friday 2025, per Adobe data reported by Retail Dive, but a study of 973 e-commerce sites found that affiliate traffic was 86% more likely to convert than ChatGPT referrals. The protocols are supposed to fix that. Here's what I found when I dug into whether they actually can.

What Are the Four Competing Protocols?

Four distinct protocols have emerged to connect AI agents with commerce, each backed by different tech giants and solving different parts of the problem. At NRF 2026, a live audience poll at Stripe's agentic commerce session found almost 75% of attendees were implementing or actively planning agentic commerce initiatives — but there's no agreement on which standard to use.

Let me try to untangle what each one is actually for.

UCP — Universal Commerce Protocol (Google + Shopify)

UCP handles the full shopping journey — browsing, cart, checkout, fulfillment tracking. Google and Shopify lead it. As of March 19, 2026, Google expanded UCP with three new capabilities: Cart (multi-item baskets per store), Catalog (real-time product data including variants and inventory), and Identity Linking (loyalty programs travel with the shopper across AI surfaces).

The coalition is big: co-developed with Shopify, Etsy, Wayfair, Target, and Walmart, and endorsed by 20-plus others including Stripe, Visa, Mastercard, Best Buy, and Macy's. When Walmart and Google announced a Walmart experience inside Gemini built on UCP in January, the plan was for customers to go from discovery to purchase without leaving the AI interface.

UCP's pitch is interoperability. One integration, every AI surface. Shopify President Harley Finkelstein calls agentic shopping "fundamentally merit-based": agents recommend what fits the shopper, not whichever big retailer a search engine surfaces first.

ACP — Agentic Commerce Protocol (OpenAI + Stripe)

ACP is narrower. It focuses on the transaction layer — the payment moment. OpenAI and Stripe built it as an open-source (Apache 2.0) standard with a key innovation: Shared Payment Tokens that let agents complete purchases without seeing full card numbers.

The flagship implementation was ChatGPT Instant Checkout — buy from Etsy and eventually 1 million-plus Shopify merchants inside the chat window. Merchants paid a fee on completed purchases; for Shopify merchants it was 4%.

I'll come back to what happened to that flagship in a moment.

MCP — Model Context Protocol (Anthropic)

MCP isn't a commerce protocol at all. It's Anthropic's general-purpose standard for connecting AI models to external data sources — databases, APIs, file systems. Shopify uses MCP to connect its product catalog to Claude and other AI assistants.

Think of MCP as the data pipe. It can feed product information to an agent, but it doesn't handle carts, payments, or fulfillment. It's the ingredient list, not the recipe.

A2A — Agent-to-Agent Protocol (Google)

A2A handles communication between agents. When your personal shopping agent needs to negotiate with a store's inventory agent, A2A manages that conversation. Nobody's really talking about A2A yet in the commerce context — there are no production deployments I could find. But if B2B agentic commerce takes off the way Gartner expects (90% of B2B buying intermediated by AI agents by 2028), agent-to-agent coordination becomes the whole game.

Which Protocol Is Winning? The One That Just Failed.

Here's where the narrative I expected fell apart. I assumed I'd be writing about protocol adoption curves and technical trade-offs. Instead, the biggest story of March 2026 dropped while I was mid-research.

On March 20, CNBC reported that Walmart found conversion on ChatGPT Instant Checkout — ACP's showcase implementation — was three times lower than when ChatGPT sent shoppers to walmart.com to check out. Three times worse. Walmart is moving off it and bringing its own Sparky AI assistant into ChatGPT instead, and OpenAI confirmed it is ending Instant Checkout.

OpenAI's response was essentially to retreat. The "buy it in ChatGPT" model is dead. The new approach: retailers build dedicated apps inside ChatGPT that route users back to the retailer's own checkout. Etsy is building one. Walmart is bringing Sparky.

This matters beyond OpenAI. It's a pattern that has now played out three times. Google Shopping Actions launched in 2018 with a universal cart across Google surfaces — Google shut it down in 2023 and moved to a redirect model. Facebook launched Facebook Shops in 2020 and pushed in-app checkout hard for years, then in 2025 moved Shops back to checkout on merchants' own websites. Now OpenAI.

Three platforms, three attempts to own checkout, three retreats. I'm starting to think this isn't a technical problem. It's a structural one.

I went back through my notes from Post 2 on the infrastructure gap. That conversion gap, with affiliate traffic 86% more likely to convert than ChatGPT referrals, suddenly makes more sense. It's not just bad infrastructure. Checkout inside someone else's interface keeps failing because the trust, the brand, the purchase context — it all belongs to the merchant. You can pipe product data through any protocol. But the moment of "yes, I'll pay" — that's earned, not transferred.

The War Nobody's Talking About: Identity and Trust

So if the protocol wars are partly settling themselves — UCP expanding, ACP's showcase dead, MCP staying in its lane — what's the actual fight? I kept pulling at this thread and ended up somewhere I didn't expect: a layer underneath the protocols that nobody in the commerce press seems to be covering.

The SSL Moment for AI Agents

In the 1990s, anyone could put up a website and claim to be a legitimate store. There was no way to verify identity. SSL certificates fixed that — a cryptographic proof verified by a trusted third party. The padlock icon became a trust signal billions of people understand.

AI agents are in that same pre-SSL moment right now. When Perplexity's Comet browser tried to shop on Amazon — the case I covered in Post 5 — Amazon had no standardized way to verify who or what was accessing its platform. The court had to step in because the technology hadn't.

What surprised me is how fast the fix is emerging.

Agent Identity: Who Is This Agent?

Visa launched its Trusted Agent Protocol (TAP) — open-source on GitHub, with agents signing each request using Ed25519 or RSA key pairs. Merchants can verify those signatures at the edge, before the request reaches their servers. It's digital identity for AI agents in commerce.

Cloudflare released Web Bot Auth, where agents sign their HTTP requests with keys published in a key directory, so Cloudflare can verify them at the edge before they reach the origin server. An IETF draft proposes Agent Name Service (ANS) — like DNS but for agents, where a name like shopping-agent.perplexity.ai could resolve to a verifiable identity with capability declarations.

And then there's World AgentKit. On March 17, Sam Altman's World project launched a tool that uses iris-scan biometrics to verify there's a real human behind an AI shopping agent's purchases. It integrates with Coinbase's x402 payment protocol for autonomous agent transactions with verified human backing.

According to HID Global's 2026 PKI study, 34% of organizations cite AI agent certificates as a top trend, and adoption stands at 18% in the United States and 13% in Europe. HID calls AI agents a new identity category.

Behavioral Trust — The Part That Keeps Bugging Me

Knowing who an agent is doesn't tell you whether it'll behave. And this is the gap I keep coming back to. Two competing standards are trying to fill it.

The Agent Certification Framework (ACF) — built by Josh Baker, who commented on my LinkedIn post about the Perplexity ruling — runs behavioral tests across four suites: commitment boundaries, consistency, hallucination detection, and adversarial resistance. Agents that pass get cryptographic certificates and appear in a public registry.

AIUC-1 is the heavier-weight alternative — think "SOC 2 for AI agents," developed with Stanford, MIT, MITRE, and others. On March 9, 2026 — the same day as the Perplexity injunction — UiPath became the first enterprise automation platform certified under AIUC-1, after more than 2,000 technical evaluations. That timing wasn't coincidental. The market is moving from "let courts decide" to "let certification decide."

The Cloud Security Alliance has its own Agentic Trust Framework. A2AS, a runtime-security framework for agents from researchers at Wallarm, AWS, Google, JPMorganChase, and others, is another. I count at least six separate initiatives that didn't exist 12 months ago — fragmented, yes, but that's what early infrastructure build-outs look like. HTTP had competing implementations too before the standards settled.

Authorization Records: What Is This Agent Allowed to Do?

The third gap — and the one the Perplexity case exposed most directly — is proving what a user actually authorized.

Mastercard launched Verifiable Intent with Google on March 5, 2026. It links a consumer's identity, their instructions, and the transaction outcome into a single tamper-resistant record, which serves as cryptographic proof of what the user authorized an AI agent to do. Baker's comment on my Post 5 nailed it: "If Comet had operated under a certification standard that required proving platform-level authorization before executing a transaction, this ruling might have looked very different."

NIST launched its AI Agent Standards Initiative in February 2026 — early stage, but it signals the federal government sees this coming. Know Your Agent (KYA) frameworks, analogous to KYC in banking, are emerging from multiple groups independently.

Three Camps, Not Four Protocols

I kept trying to organize four protocols into a coherent picture and it never worked. The mental model that finally clicked has three camps, not four protocols. The protocols are just weapons each camp picked up.

Amazon and eBay are the walled garden. Block external agents, build internal ones, own every layer. Amazon's Rufus was used by more than 250 million customers in 2025, and Andy Jassy said it was expected to generate over $10 billion in annual incremental sales. Two days after the Perplexity injunction, Amazon opened Shop Direct to third-party feeds: over 100 million products from more than 400,000 merchants, with tens of millions of them purchasable through Buy for Me. They don't need open protocols. They already are the protocol.

Then there's the open coalition — Google, Shopify, Visa, Mastercard — building interoperable standards so every store is accessible to every agent. UCP is the backbone. Shopify just made its merchants' products discoverable in ChatGPT by default through Agentic Storefronts, with Copilot, Google's AI Mode, and the Gemini app managed from the same admin. Shopify merchants now show up in ChatGPT whether they know it or not. That's a fundamentally different philosophy from Amazon's walls.

And OpenAI? They're regrouping. The 4% fee model didn't survive contact with actual conversion data. The pivot — retailer apps inside ChatGPT — is an admission that AI platforms are discovery layers, not transaction layers. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027. I'd bet that number is low for commerce specifically.

What I Still Can't Figure Out

Here's where I'm stuck. The protocol you support probably matters less than your data readiness. Shopify merchants already show up in ChatGPT by default. WooCommerce and Magento merchants — the ones I flagged in Post 1 as invisible to AI agents — still need to act. But the act isn't "pick a protocol." It's "make your product data machine-readable."

I've been helping merchants assess their AI readiness through Ivinco. Last month I ran audits on 12 mid-market WooCommerce stores. Eight of them had product schema markup so incomplete that no AI agent could extract basic pricing or availability. The four that showed up in AI recommendations weren't the ones with the biggest ad budgets — they were the ones where a developer had properly implemented structured data. A $2M/year DTC brand was invisible to ChatGPT. A $400K Shopify store selling the same category showed up on the first query.

But here's the question I keep circling: the platforms are fighting, the protocols are splitting, the courts are ruling. What do actual consumers want? Do they even care about AI shopping, or is this an industry building infrastructure for a demand that doesn't exist yet?

That's what I'm digging into next — starting with a paradox. GenZ uses AI for shopping more than any other generation. They also trust it the least.

Frequently Asked Questions

Which agentic commerce protocol is winning in 2026?

Google's Universal Commerce Protocol (UCP) has the broadest coalition — co-developed with Shopify, Etsy, Wayfair, Target, and Walmart, and endorsed by 20-plus others including Stripe, Visa, Mastercard, and Best Buy. OpenAI's ACP lost its flagship implementation when Walmart revealed ChatGPT checkout converted three times worse than sending shoppers to its own website. Anthropic's MCP serves as a data layer, not a full commerce standard.

What happened to ChatGPT Instant Checkout?

OpenAI killed Instant Checkout in March 2026 after Walmart disclosed conversion three times lower than when ChatGPT sent shoppers to walmart.com to check out. The replacement model: retailers build dedicated apps inside ChatGPT that route users back to the retailer's own checkout. This effectively concedes that AI platforms are discovery layers, not transaction layers.

What is UCP (Universal Commerce Protocol)?

UCP is Google and Shopify's open standard for AI-powered shopping. It handles the full commerce journey — browsing, cart management, checkout, and fulfillment tracking. As of March 2026, UCP supports Cart (multi-item baskets), Catalog (real-time product data), and Identity Linking (loyalty programs across AI surfaces), with 20-plus retail and payment partners.

How do AI agents prove their identity to merchants?

Several competing systems are emerging: Visa's Trusted Agent Protocol has agents sign requests with Ed25519 or RSA key pairs. Cloudflare's Web Bot Auth verifies signed agent requests at the edge. An IETF Internet-Draft proposes Agent Name Service (ANS) — like DNS for agents. World AgentKit uses iris-scan biometrics to verify a real human backs each agent.

Should merchants pick a specific AI commerce protocol?

Protocol choice matters less than data readiness. Shopify merchants already show up in ChatGPT by default through Agentic Storefronts. For other platforms, the priority is making product data machine-readable — structured markup, clean product feeds, explicit pricing and availability, and return policies in formats agents can parse. The protocols will come to you; the data readiness won't happen automatically.