A woman tells her AI assistant to reorder the running shoes she bought last spring: same brand, under £90.
But the assistant doesn’t open a browser. Instead, it queries two merchants directly.
The first merchant runs a digital commerce standard the AI agent can read. Its systems tell the agent what it sells, confirm the size and price, and let the agent check out and pay, all in under four seconds.
The second merchant sells the same shoe for less, but the agent never finds it, because that store runs on dropdown menus, image carousels, and a checkout flow built for human fingers on a screen.
To the agent, that second website returns nothing useful. The cheaper shoe loses out to the readable one.
We’re now in the agentic commerce era, where autonomous agents act on a person’s behalf to discover products, compare options, and complete purchases without a human clicking through a website.
The buyer is now an AI agent reading structured data on commerce systems rather than a shopper with eyes on a screen.
In the first quarter of 2026, Adobe measured AI-driven traffic to US retail sites growing 393% year over year. That traffic converted 42% better than non-AI traffic in March.
Salesforce reported that AI and agents influenced 20% of global retail orders over the 2025 holiday season, around $262 billion in sales.
But many stores today aren’t built for AI shopping assistants.
Adobe found that around a quarter of retail homepage and category content can’t be read by large language models at all.
A new audience is arriving, and the question for any merchant is whether their storefront or commerce platform speaks the language agents read.
Two of those languages, Universal Commerce Protocol (UCP) and Agentic Commerce Protocol (ACP), let an agent discover your products and check out.
Two others, x402 and Machine Payments Protocol (MPP), handle a different job: letting an agent pay directly for a service, API, or digital resource.
The practical question for a merchant today is less “which files do I publish on my website?” and more “which AI channels do I want to sell through, and what machine-readable catalog, checkout, and payment data do those channels need?”
An agentic commerce suite with the right AI capabilities determines which agent transactions you can tap into.
How agentic commerce changes the buyer at your door
For two decades, ecommerce design has assumed a human visitor.
Retailers optimise for eyes and thumbs with hero images, persuasive copy, urgency banners, and a product catalog laid out for browsing.
But generative AI, conversational commerce, and agent interaction break that assumption.
Behind every agent stands a consumer, but the agent is the one reading the catalog, comparing prices, and forming purchasing decisions.
When a consumer hands a request to an AI shopping agent, the agent doesn’t see the page the way we do.
Instead, it reads whatever structured data your site exposes, weighs the options against the shopper’s intent, and acts.
This is already happening at scale. More than 900 million people use ChatGPT each week, and 39% of consumers told Adobe they’ve used AI for online shopping.
ChatGPT’s Instant Checkout lets a buyer complete a purchase inside the chat. OpenAI launched it in September 2025 with Etsy sellers and Shopify merchants like Glossier and SKIMS.
Walmart’s Sparky and Amazon’s Rufus push the same pattern inside their own storefronts. I’ve tested Rufus on a few orders and managed to facilitate a return and refund.
Per Adobe data, AI-sourced shoppers spend 48% more time on retail pages than other visitors. This suggests deeper engagement once an agent reaches a readable site.
But this digital commerce model is still settling. OpenAI has since moved toward dedicated retailer apps that route buyers back to the merchant’s own site rather than closing the sale in chat.
What digital commerce standards make a store agent-readable?
An alphabet soup of acronyms powers intelligent commerce for shoppers and agentic AI tools today.
Two standards, UCP and ACP, let a generative AI tool or agent find and buy your products. A merchant can support one or both.
Two more, x402 and MPP, let an AI agent pay for a service or resource over the open web. These overlap with each other, and they apply mostly when an agent buys digital access rather than physical goods.
Universal Commerce Protocol (UCP)
Imagine a shop where a board by the door tells you what’s sold, and behind that door the whole place is arranged so a visitor can pick items, prove who they are, pay, and track the order, without help from staff.
UCP is that shop, built for AI agents. Google developed it with Shopify, Etsy, Wayfair, Target, Walmart, and many others.
It covers the full shopping journey: product discovery, cart building, identity linking, checkout, and order management.
A merchant publishes a UCP profile at yourwebsite.com/.well-known/ucp that tells agents which of these capabilities it supports and where to reach them.
UCP works over standard APIs and connects to agent frameworks like Model Context Protocol (MCP) and Agent-2-Agent (A2A), and it leans on the new Agent Payments Protocol (AP2) for the payment step.
It already powers checkout inside AI Mode in Google Search and the Gemini app. For a merchant, supporting UCP puts your store on Google’s generative AI and agent surfaces.
Agentic Commerce Protocol (ACP)
ACP does the same job as UCP—finding products and checking out—but it comes from the other camp. Stripe and OpenAI designed it, and it powered Instant Checkout inside ChatGPT.
ACP works mostly through a product feed. A merchant sends OpenAI a structured feed of its catalog, with identifiers, pricing, inventory, media, and fulfilment details, refreshed daily.
When a buyer is ready, the agent runs a small set of checkout endpoints and settles the payment using a Shared Payment Token (SPT), so the merchant gets paid without the agent ever handling raw card details.
For eligible Shopify stores, merchants don’t need to publish a separate ACP discovery file for ChatGPT discovery.
Product data is already integrated into ChatGPT through Shopify Catalog, and Shopify says eligible products are discoverable by default through its ChatGPT agentic storefront.
UCP and ACP cover overlapping ground, so the practical move for many retailers is to support both and reach buyers on Google surfaces and ChatGPT at once.
x402
Now imagine a different job: you want a piece of premium content (say, a paywalled article) or a single API call.
You send your payment, the service checks it, and hands over the resource. No account or human cashier involved; your wallet and the machine talked directly.
x402 is that interaction, built for agents paying over the open web.
It revives HTTP status code 402, “Payment Required,” which had gone unused in the web’s plumbing since the 1990s.
If you’re not familiar with status codes, think of “404” (Not Found) and “403” (Forbidden). 402 works similarly: a status message sent to whoever, or whatever, is trying to reach a resource.
When an agent requests a paid resource, an x402-enabled server answers with a 402 carrying the price and payment terms.
Once the user has granted the agent authority within defined limits, the agent can fulfil the payment terms itself.
In x402 implementations, that often means an on-chain stablecoin payment, without card entry during the transaction.
x402 fits paid APIs, data, tools, and digital content better than it fits buying a pair of shoes.
In April 2026, the protocol moved to the Linux Foundation, which launched the x402 Foundation to govern it.
Coinbase contributed the protocol, and members include Stripe, Cloudflare, Google, Visa, and Mastercard.
Machine Payments Protocol (MPP)
MPP does the same kind of work as x402, agent-to-service payment, but wraps it in more structure. Stripe and Tempo launched it in March 2026.
Instead of requiring a separate payment for every request, MPP sessions let an agent open something closer to a spending tab.
The agent authorises a capped payment session upfront, then makes many small payments against it, which suits metered API usage, recurring charges, and streaming, pay-as-you-go services.
Stripe supports MPP payments in stablecoins and, through Shared Payment Tokens, through card and selected buy-now-pay-later methods such as Klarna and Affirm.
Where x402 is best understood as a per-request HTTP 402 payment challenge, MPP sessions are better suited to longer-running payment flows where an agent keeps consuming a service over time.
Like x402, this applies most naturally when an agent is paying for a digital service, API, tool, or resource rather than buying physical goods.
How a complete agent purchase works
Let’s trace the shoe purchase again from earlier using these standards.
Say the buyer uses ChatGPT. The agent receives the intent: “same brand, under £90.”
It finds the shoe in the merchant’s product feed, confirms the size and price, runs the ACP checkout endpoints, and settles the payment with a Shared Payment Token.
The merchant accepts the order and fulfils it. No human touches the keyboard after the first instruction.
Run the same purchase on Google’s AI Mode and it flows through UCP instead, from catalog lookup to checkout to the payment step handled by the Agent Payments Protocol.
Different standard, same outcome: the agent discovers, decides, and buys in seconds, because every step reads a structured file rather than parsing a messy page like a human would.
x402 and MPP enter a different scene. Picture an agent that needs to pay 50p for a single data lookup or a metered API while completing a task.
There’s no checkout session or shopping cart—just a direct machine payment for a resource.
| Standard | Who’s behind it | What it does |
| UCP | Google and Shopify | Full checkout on Google’s agent surfaces: discovery, cart, identity, payment, orders |
| ACP | OpenAI and Stripe | Product feed plus checkout inside ChatGPT, paid via Shared Payment Token |
| x402 | Coinbase, now the Linux Foundation | One-off agent payment for an API, tool, or digital resource over HTTP 402 |
| MPP | Stripe and Tempo | Ongoing agent payments for services, with sessions, recurring, and streaming |
The same checkout pattern covers more use cases than just footwear.
Grocery refills, printer ink, subscription renewals, and replacement parts all follow the discover-then-buy sequence.
In each of these use cases, the shopper already knows what they want, so delegating the purchase to an agent removes friction without adding too much risk.
An agentic transaction in those categories asks little judgment of the agent and saves the consumer a repetitive chore.
Consumer trust is the harder problem
The standards above handle discovery, checkout, and payment mechanics, but they don’t decide what an agent may buy without asking.
This comes down to the stakes involved. A wrong £5 grocery item is mildly annoying, but a wrong flight booking, annual software contract, or duplicated high-value order is costlier.
So agentic payments need authority boundaries: rules that tell a personal or business agent where it can act alone and where it must obtain confirmation.
Consumer trust in autonomous AI agents grows by category. Shoppers are most willing to delegate low-risk, repeatable purchases like groceries, which 41% would let an agent handle, and least willing on high-consideration purchases like financial services, at 15%.
Trust widens as consumers watch the agent get their buying decisions right. A shopper who lets an agent reorder coffee this month might let it handle a larger basket next month, and a different category the month after.
While that trust accrues to the agent, merchants inherit part of the problem. Clear return paths, honest product data, and predictable pricing all help an agent or AI assistant transact confidently on a shopper’s behalf.
This raises completion rates and lowers the disputes that follow bad transactions.
Retailers with trustworthy data give the agent fewer reasons to abandon the purchase, and give the consumer behind the agent fewer reasons to revoke the authority they delegated.
What this means for ecommerce merchants now
In the agentic commerce era, the merchants who’ll benefit first sell high-frequency, low-consideration items: consumables, refills, replacements, and repeat orders.
These are the purchasing decisions a consumer is happiest to delegate, and the agentic storefronts that serve them well will capture demand that browser-only sites currently can’t.
As we saw earlier, traffic data shows agents already shopping, while readability data shows most websites unready to serve them.
An agentic system can only act on what the merchant exposes, so your agentic capabilities rise or fall on the quality of your machine-readable data.
Start by choosing your channels. If you want to sell inside ChatGPT, you support ACP, usually through a product feed.
If you want to show up on Google’s AI surfaces, you support UCP. Many retailers will do both.
If you run on Shopify, you get a head start. Shopify rolled out Agentic Storefronts to millions of merchants in March 2026, giving them out-of-the-box access to ChatGPT and other AI channels from the admin.
The structured product catalog still depends on the quality of the data you expose.
The payment rails, x402 and MPP, only enter the picture if you sell digital access, APIs, or metered services that an agent pays for directly.
A clothing or grocery store can skip them for now.
| Move | What it covers | Priority |
| Support a checkout standard (UCP, ACP, or both) | Discovery and checkout on Google and ChatGPT | Foundational |
| Clean up your product feed and catalog data | Accurate pricing, inventory, media, fulfilment | Foundational |
| Set authority boundaries for delegated buying | Spending caps and confirmation rules | Standard |
| Add a payment rail (x402 or MPP) | Only if you sell APIs or digital access | Situational |
The bigger picture behind agentic AI commerce
Each wave of web infrastructure widened who could read a site.
Search engine optimisation (SEO) made pages readable by crawlers, while schema markup made content readable by knowledge graphs.
The standards I’ve described so far make commerce readable by AI agents. Early adopters will capture the traffic flowing through them, while laggards will get harder to find.
A retailer who treats agent readability as part of the customer experience, rather than a side project, gives the next buyer a reason to complete the transaction with them over a competitor.
The audit question for retailers is simple: point an agent at your own site and see what it can read.
If your catalog isn’t machine-readable and you haven’t connected to a single agent channel, the agents shopping today are passing you by.
Book an audit to learn how agent-ready your setup is today.