UCP vs ACP: The Two Protocols Fighting Over Agentic Checkout

Two open protocols. Both backed by companies you have heard of. Both promising that an AI agent will one day complete a purchase on a shopper's behalf without anyone touching your storefront.

One of them already ran the experiment and lost.

That is the useful frame for this, because the protocol comparison people keep publishing is mostly a features table, and features tables do not tell you what to do on a Tuesday. What tells you what to do is the adoption record.

ACP came first

On September 29, 2025, OpenAI and Stripe shipped the Agentic Commerce Protocol and, with it, Instant Checkout inside ChatGPT. US Etsy sellers went live that day. Over a million Shopify merchants, including names like Glossier, Vuori, Spanx and SKIMS, were listed as coming soon.

ACP was released as an open standard. Businesses not processing with Stripe could adopt it with their existing payment providers. On the fee, OpenAI said only that it would charge merchants a "small fee" on completed purchases, per TechCrunch's coverage that day, and stressed that results were organic and unsponsored.

The number reported everywhere since is 4%. That figure has never appeared in an OpenAI announcement we can find. Widely repeated, consistently repeated, not primary-sourced. Plan around it as an estimate, not a published rate card.

Then the retreat

On March 4, 2026, OpenAI pulled back Instant Checkout and moved purchasing into merchant-specific apps inside ChatGPT instead.

The number that came out afterwards is the one to remember. In Forbes coverage on March 10, Shopify president Harley Finkelstein said only about a dozen of Shopify's millions of merchants had actually gone live with it. A dozen. Finkelstein's position was that the holdup sat on the AI firms' side, not the merchants'.

Read that in context. The infrastructure shipped. The largest AI assistant on earth was promoting it. Payment rails were in place. Adoption rounded to zero, and shoppers who used ChatGPT to research kept finishing the purchase on the retailer's own site.

The protocol itself did not die. OpenAI still publishes ACP documentation covering feeds, products and promotions, with both file-upload and API integration paths. The consumer feature is what got pulled.

UCP came second and got bigger fast

On January 11, 2026, at NRF in New York, Google announced the Universal Commerce Protocol, an open standard covering the full shopping journey: discovery, cart, checkout, payment, and post-purchase things like tracking and returns.

Shopify published its side of the design the same day, describing UCP as co-developed with Google. Etsy, Target, Walmart and Wayfair were named as co-developers alongside Shopify. Twenty-plus endorsers came with it, including Adyen, American Express, Best Buy, Flipkart, Macy's, Mastercard, Stripe, The Home Depot, Visa and Zalando. Google also noted UCP is compatible with existing protocols it already had in flight: A2A, the Agent Payments Protocol, and MCP.

Then on April 24, 2026, Amazon, Meta, Microsoft, Salesforce and Stripe joined the UCP Tech Council, the body steering the standard. That put ten organizations on the council, including companies that had been pointing in different directions six months earlier.

Notice who is on both lists. Stripe co-authored ACP with OpenAI and sits on the UCP Tech Council. Microsoft runs Copilot, which reads Shopify Catalog, and joined UCP. The payments companies are not picking a winner. They are wiring themselves into whichever one gets volume.

Fee economics, and why they matter less than you think

Line the three models up.

ACP, as reported, takes roughly 4% of a completed purchase on top of your processing costs. That is cheaper than an Amazon referral fee and vastly more expensive than your own checkout.

Perplexity's merchant program charges merchants nothing. No commission, no listing fee. Perplexity funds shopping out of Pro subscription revenue rather than taking a cut of your sale, which is a structurally different bet about where the money comes from.

Shopify's own agentic distribution sits closer to the Perplexity end. Products syndicated through Shopify Catalog carry no transaction fees beyond standard processing rates, and the Agentic plan has no monthly fee at all, just card rates when something sells.

Here is the part that gets skipped. A 4% take rate on a channel doing a dozen merchants' worth of volume is not a business decision. It is a footnote. Fee structure will matter enormously in three years and matters almost not at all this quarter, because the denominator is tiny. Anyone telling you to pick a protocol based on take rate today is optimizing the wrong variable.

What fee structure does tell you is intent. OpenAI priced ACP like a marketplace. Google and Shopify priced UCP like infrastructure. Infrastructure tends to win standards fights, because nobody has to justify the line item.

What a merchant should actually do

Nothing rash. That is not a dodge, it is the position the evidence supports.

Do not build a custom protocol integration. If you are on Shopify, your storefront already publishes a UCP business profile at /.well-known/ucp and Shopify handles the protocol layer for you. Paying an agency to hand-roll agentic checkout in 2026 is paying to place a bet that is already placed.

What does pay out under every scenario is clean product data. Correct taxonomy, complete variants, real specs, honest availability. Both protocols read the same fields, and the same work makes your on-site search usable and your Shopping feed cleaner if agents go nowhere at all.

Three things worth checking this month:

• Whether /.well-known/ucp on your storefront returns valid JSON rather than an error.
• Whether any of your products are excluded from Shopify Catalog on a technicality, because if they are, the protocol debate is academic for you.
• Whether anyone on your team turned off an AI channel out of caution. Opting out of the clean feed does not stop the scraping. It just makes the scraped version the only version.

Then track adoption rather than announcements. Merchants live and orders placed, not council membership. Instant Checkout is what an announcement-rich, adoption-poor rollout looks like on the way down, and the announcements were excellent right up to the end.

What is genuinely unsettled

We do not know whether consumers will hand an agent payment authority at scale. The only real-world test so far said no, and said it fast.

We do not know whether UCP's council converges into a single standard or fragments once the members' interests diverge, which they will. Amazon joining a Google-initiated protocol council is not the same as Amazon shipping on it.

And we do not know what any of this looks like when the assistants start monetizing placement. Right now OpenAI describes its product results as organic and unsponsored. That is a statement about today, not a constitutional guarantee.

What is not speculative is the reading layer. Agents are already querying catalogs, building shortlists, and handing shoppers to your site to finish the job. We wrote about that split in more depth in agentic commerce and what happens when AI agents do the shopping. The buying may or may not arrive. The reading is here.

If you want to see what an agent currently gets when it reads your store, an AI-readiness check is a more useful place to start than a protocol decision: run one on your storefront.