What is agentic commerce, and what does your online store need to do now?
Agentic commerce is when an AI agent completes a purchase for a shopper, with their permission, instead of only recommending products. Agents buy in two ways today: through your normal checkout, like a person, or through checkout protocols such as Google's UCP and OpenAI and Stripe's ACP. Most stores don't need a protocol yet. They need product pages and a checkout an agent can use.

- Google's definition: agentic commerce is "where AI doesn't just help you find what you need, but actively helps you complete tasks on your behalf, with your permission".
- Most agent orders use the store's own checkout: in Forter's merchant network, agents that drive a store's own checkout pages went from 1% of agent orders in late July 2026 to 77% in the week of 19 October.
- Protocols are open, but access is limited: Google's Universal Commerce Protocol (UCP) and the Agentic Commerce Protocol (ACP) from OpenAI and Stripe are open standards. The checkouts built on them are pilots, mostly for US shoppers, and you usually have to apply or be eligible.
- Still small: DoorDash said it had seen "maybe several hundred orders in total" from third-party agents, against about 3 million deliveries a day.
- Agents make mistakes: in Visa's Agentic Commerce Benchmark, AI models failed 42.8% of shopping test runs, and 54% of failed runs still ended in a purchase.
What does agentic commerce mean?
Until recently, "AI shopping" mostly meant an assistant that answers questions: which running shoe suits flat feet, which laptop is lighter. The shopper still clicked through to a store and paid.
Agentic commerce is the next step. The AI acts as the shopper's agent: it finds the product, picks a size, fills in the delivery details and pays, after the person approves. Google put it simply when it launched agentic checkout in Canada on 4 November 2026: AI that "actively helps you complete tasks on your behalf, with your permission".
Three parties are involved:
| Who | Examples | What they want from your store |
|---|---|---|
| The agent working for the shopper | Meta's Muse, ChatGPT Agent, Google AI Mode, Microsoft Copilot | Clear product facts, prices, stock and a checkout it can complete |
| The store | You | Sales, the customer relationship, and protection from fraud |
| The payment layer | Stripe, Visa, Google Pay | A safe way to pass payment details without exposing card numbers |
In every protocol we looked at, the store stays the merchant of record. You still sell, ship, handle returns and own the customer.
The two ways an AI agent buys from a store
This is the part most guides skip, and it changes what you should do.
| Browsing agents | Protocol checkouts | |
|---|---|---|
| How they buy | Open your pages and fill in your checkout like a person | Call a checkout API the store has connected |
| Examples (per Forter) | Meta's Muse, Instinct, ChatGPT Agent | ChatGPT Instant Checkout (ACP), Google's agentic checkout (UCP) |
| What you need | Pages and a checkout that work for software | An integration, usually through your platform or payment provider, plus eligibility |
| Reach | Any store the agent can load | Only stores that have integrated |
| What you see | Looks like a normal visitor; hard to tell from a bad bot | Structured, identified orders |
| Share of agent orders (Forter, week of 19 Oct 2026) | 77% | The rest |
Forter, a fraud-prevention company, explains the gap plainly: browsing agents "can shop anywhere, while the protocols can only reach those stores that have integrated them." Through 7 October, Meta's Muse bought at 74% of the stores in Forter's network that received any agent order.
Read these figures as a signal, not a market count. They come from one company's network, mostly in North America. We covered the earlier release in our news post on agents using your normal checkout.
The main protocols, in one table
| Protocol or programme | Backed by | What it does | Who can use it now |
|---|---|---|---|
| Universal Commerce Protocol (UCP) | Google, co-developed with Shopify, Etsy, Wayfair, Target, Walmart and others | Catalog search, cart, checkout, identity linking and orders between agents and stores | Google's direct checkout in AI Mode and Gemini is live for select retailers in the US and, since 4 November 2026, Canada. Google says simplified UCP setup in Merchant Center starts "next year" |
| Agentic Commerce Protocol (ACP) | OpenAI and Stripe | Lets an agent start and complete a checkout with a store, and pass payment securely | Open source (Apache 2.0). ChatGPT was the first platform to use it, and you have to apply to take part |
| Copilot Checkout | Microsoft | Buying inside Copilot, using a UCP-ready Microsoft Merchant Center feed | English-language merchants selling to US buyers in USD. Needs a Merchant Center account |
Two things stand out. First, ACP's own FAQ says that implementing the protocol does not mean your products will automatically be listed by AI agents. Each AI platform runs its own process. Second, the protocols are converging on the same basics: accurate product data, clear policies, and the store keeping control of payment and fulfilment.
You don't need to read the specs. If you sell on a big e-commerce platform or use a major payment provider, any protocol support will most likely come through them. Microsoft, for example, says Shopify merchants were enrolled in Copilot Checkout automatically, with controls in the Shopify admin, and that Stripe offers a ready-made route.
Is anyone actually buying this way yet?
Yes, but from a low base, and the numbers point in different directions.
- Growing fast: Forter says agent orders are up 7.2 times since the first week of August 2026, and the average agent order was $227.
- Still rare for some retailers: Amazon's CEO Andy Jassy said few customers buy through outside shopping agents because the experience is "still not great, nor more convenient than coming to a retailer directly". DoorDash reported only several hundred agent orders in total.
- Not reliable yet: Visa tested 22 AI models from 12 labs on 15 shopping tasks. The best models passed 82.2% of attempts, the worst 28.9%. All three attempts at the same task passed just 34.2% of the time.
- Trust is the bottleneck: Visa's September 2026 Trust Index found 72% of US consumers had used an AI assistant, but only 23% trusted AI to handle payments for them.
So agentic commerce is real, early and uneven. For more on each data point, see our posts on Amazon and DoorDash's agent order numbers and Visa's AI shopping benchmark.
What your online store should do now: a 7-point checklist
Everything here helps human shoppers too, so none of it is wasted if agents grow slowly.
- Put the deciding facts in text. Price, stock, sizes, variants, delivery cost and times, and the returns policy should be readable text on the product page, not only in images, tabs that load later or pop-ups. An agent can't buy what it can't read.
- Mark up products and policies. Product structured data with price and availability, plus shipping and return details, gives agents a second, machine-readable copy. Our guide to showing shipping and return policies to Google and AI shows how.
- Test your checkout as a stranger. Allow guest checkout, label every field clearly, and avoid steps that only work on hover. A forced account sign-up is one more step where an agent can get stuck.
- Keep your data consistent. The price and stock on your page, in any Google or Microsoft Merchant Center feed and at checkout should match, so an agent never quotes a price that your checkout then changes.
- Look before you block. Check your firewall and bot settings. Cloudflare now shows personal agents as their own traffic type and lets you allow or block each one. In its words, blocking agents altogether "can feel like the safest choice, but it can mean losing customers." See our note on Cloudflare's per-agent controls.
- Ask your payment and fraud providers. Find out how they treat agent orders, whether they support agent payment tokens, and whether a real customer's agent is blocked by default.
- Ask your platform about protocols. If you're on a hosted platform, ask whether it supports UCP, ACP or Copilot Checkout and what you need to switch on. If you run a custom stack and sell to the US, watch Google's Merchant Center UCP rollout next year.
What you can skip for now
- Building your own protocol integration from scratch, unless you're a large retailer selling to the US. The checkouts are pilots and the platforms do most of the work.
- Panic. The share of sales coming through agents is still small for most stores. The work above is ordinary good e-commerce, done a little more carefully.
Check your site
The free Web2Agents AI check shows whether AI tools can reach and read your homepage, with results on screen. The €19 AI-Ready Kit runs 36 checks on up to 25 pages, including product pages and their schema, and ranks the fixes. For the wider list, use our free GEO audit checklist.
FAQ
What is agentic commerce in simple terms?
It's shopping where an AI agent does the buying for a person. The agent finds the product, fills in the order and pays, after the person approves. Google describes it as AI that helps you complete tasks on your behalf, with your permission.
Do I need UCP or ACP to sell through AI agents?
Not to start. In Forter's data, most agent orders came from agents that use a store's normal checkout. The protocol checkouts are mostly US pilots that you apply for or get through your platform or payment provider.
Is agentic commerce available in Europe?
The protocol checkouts are mostly not available there yet. Copilot Checkout is limited to US buyers, and Google's direct checkout covers the US and, since November 2026, Canada. Browsing agents are used more widely: Forter says ChatGPT Agent placed nearly a quarter of its orders outside the US.
Should I block AI agents from my online store?
Not by default. An agent buying for a real customer is still a customer. Check which agents visit, whether they reach your product pages and checkout, and decide per agent. Cloudflare, for example, now lets you allow or block each personal agent.
Will AI agents buy the wrong product and cause returns?
Sometimes. In Visa's benchmark, AI models failed 42.8% of shopping test runs, and 54% of failed runs still ended in a purchase. Clear specs, sizes and compatibility details on your pages lower the risk.
Sources
- A new way to shop: agentic shopping for consumers and retailers in Canada, Google Canada blog (4 Nov 2026)
- AI Agent Order Data: Agentic Commerce on Forter's Network, Forter (updated 2 Nov 2026)
- Universal Commerce Protocol, ucp.dev
- Agentic Commerce Protocol, agenticcommerce.dev
- Agentic Commerce, Microsoft Advertising
- Visa Releases First-of-Its-Kind Industry Benchmark for AI Shopping, Visa (5 Nov 2026)
- Amazon CEO says shopping agents like Meta's Muse 'not great', Fortune
- DoorDash said it's only received 'several hundred' orders from AI agents, TechCrunch
- See the agents, set the terms: new visibility and control for personal agents, Cloudflare blog


















