Agentic Commerce UK: What's Live, What's US-Only, What to Do Now

Read the US trade press and you would think AI agents are already doing the nation’s shopping. From a UK retailer’s desk in October 2026 the picture is more useful, and more sober: AI assistants already shape which products British shoppers consider, but the “buy it without leaving the chat” features are mostly switched on for US customers only — and the first big one, ChatGPT’s Instant Checkout, has already been withdrawn.

I write this as someone who has just made a website agent-ready. Aderlo City runs a 3D shopping mall in the browser with galleries in London and Warsaw, and this year we gave aderlo.city an MCP server, an OpenAPI-described API, Markdown versions of every article and an OAuth 2.1 server. Below is what is live in the UK, what is US-only, how ACP, UCP and MCP fit together, and a 90-day plan you can start on Monday.

What is agentic commerce?

Agentic commerce is shopping in which an AI agent acts on the customer’s behalf: it searches, compares and, increasingly, places the order. The customer states a need; the agent does the steps the customer used to do in a search engine and on your site. Our glossary has a short definition of agentic commerce.

It helps to separate three levels, which marketing tends to blur:

1. Shopping assistant — answers questions and recommends products with links. The human buys on your site.
2. Agent with a basket — gathers products, tracks prices and stock, prepares the order. The customer approves and pays.
3. Buying agent — within limits the customer has set (budget, brands, type of product), completes the purchase and payment.

Level one is mainstream in the UK. Levels two and three are live in the US and in pilots. For a UK retailer, the order of work follows from that: first be found and correctly described by agents; worry about in-chat payment later.

The stakes are real. The ONS puts internet sales at 27.3% of all retail sales in Great Britain in August 2026, and Google said in May 2026 that AI Mode in Search has passed one billion monthly users. AI Mode has been rolling out in the UK since 28 July 2025.

What agentic commerce is live in the UK today?

Discovery is live; in-chat checkout mostly is not. The table below is built from the platforms’ own documentation as of 17 October 2026.

Channel	Product discovery for UK shoppers	In-chat checkout — where it works	What it means for a UK retailer
ChatGPT (OpenAI)	yes, via ChatGPT search	Instant Checkout withdrawn in March 2026; purchase finishes on the retailer’s site or app	make your pages crawlable and your products well described
AI Mode and Gemini (Google)	yes, AI Mode in the UK since 28.07.2025	UCP checkout: US, Canada, Australia; UK “later”	Merchant Center and product data are the foundation
Microsoft Copilot	yes	direct checkout shown only to US customers (Shopify channel)	—
Perplexity	yes	Instant Buy with PayPal launched for US users	—

Which agentic checkouts are US-only, and why?

Almost all of them, for now. Agent checkout needs tax, delivery, payment tokens and consumer-law handling to be solved per market, and the platforms started with the US.

ChatGPT. OpenAI and Stripe launched Instant Checkout, built on the Agentic Commerce Protocol, in September 2025. In March 2026 OpenAI pulled it back. Forrester wrote on 7 March 2026 that native checkout in ChatGPT was gone and purchases were moving to merchants’ apps in ChatGPT or back to their own sites; citing The Information, it said only about a dozen Shopify merchants had been live with the feature (Shopify had told Forrester a month earlier it was closer to 30). Shopify now describes ChatGPT as a “discovery-focused referrer” whose customers pay in your own checkout.

OpenAI’s product feed is not open to everyone either. Its documentation says feed onboarding is for approved partners, and that a standard upload targets the US: the shared format currently represents the US, Canada and Mexico, and other markets apply only once OpenAI confirms them for your integration.

Google. UCP-powered checkout appears in AI Mode and the Gemini app for select merchants and products eligible in the United States, Canada and Australia, and merchants join through an interest form. At I/O on 19 May 2026 Google said UCP checkout would come to Canada and Australia “in the coming months” and then to the UK — no date. The same announcement introduced Universal Cart, which is rolling out in the US first.

Shopify’s agentic storefronts make the geography very concrete. Direct checkout in AI Mode and Gemini requires a store based in the US selling to US customers. Copilot requires selling to US customers. ChatGPT lets your store be based outside the US, but you must sell to customers in the US. Meta’s direct checkout covers the US, Canada and Mexico.

Perplexity. PayPal and Perplexity launched Instant Buy on 25 November 2025 for US users.

So if you sell only in the UK, you are not missing a checkout integration you could switch on today. You may be missing the data work that decides whether agents recommend you at all.

ACP vs UCP vs MCP: what does a UK retailer need to know?

ACP and UCP are protocols for agent checkout (basket, payment, order). MCP is a general standard for connecting AI models to tools and data — including a shop’s catalogue. All three are open, and both commerce protocols support MCP as a transport.

	ACP — Agentic Commerce Protocol	UCP — Universal Commerce Protocol	MCP — Model Context Protocol
Maintained by	OpenAI and Stripe (founding maintainers)	Google, Shopify, Etsy, Wayfair, Target, Walmart and others; endorsed by Visa, Mastercard, PayPal, Adyen, Zalando and more	open standard started by Anthropic; current spec dated 2026-07-28
Covers	agent checkout and delegated payment; since the 2026-04-17 version also basket, feed, orders and authentication	catalogue, basket, checkout, account linking (OAuth 2.0), orders and returns	tools and data for AI models
Consumer deployment	ChatGPT (Instant Checkout, until March 2026)	AI Mode and Gemini in the US, Canada, Australia	AI assistants and developer tools
Merchant of record	the retailer	the retailer	not applicable
Licence	Apache 2.0, in beta	Apache 2.0	open

ACP defines a shared language between buyer, agent and business. The retailer exposes checkout endpoints (REST or MCP); card details reach the payment provider as a token, so the retailer’s existing payment rails stay in place. The specification repository shows how fast it is moving: from the first release on 29 September 2025 to the 17 April 2026 version that added basket, feed, orders, authentication and MCP.

UCP was announced by Google on 11 January 2026 as an open standard “across the shopping journey”, from discovery to post-purchase support. ucp.dev lists Google, Shopify, Etsy, Wayfair, Target, Walmart, Amazon, Microsoft, Meta, Salesforce and Stripe among the co-developers of its shopping part. The key principle for retailers: you remain the merchant of record and own the customer relationship. Google’s AP2 payments protocol adds a mandate in which the shopper sets the agent’s spending limit and purchase criteria.

MCP is not a commerce protocol at all; it is the plug between an AI model and any system. A retailer can use it to expose product search, stock checks or order status. That is why it turns up inside both commerce protocols.

Should UK retailers implement ACP or UCP now?

Only if you sell to customers in the US, Canada or Australia, or you are a platform or payment provider. For a UK-only shop there is no consumer-facing ACP or UCP checkout to plug into yet, and when UCP reaches the UK, most retailers will get it through their ecommerce platform and payment provider — the same way they got Merchant Center integrations.

The lesson of Instant Checkout is that channels change faster than integration projects. The investment that survives every change is data: an agent that does not understand your product will not recommend it, whichever protocol it pays with.

Worth doing now:

• if you ship to the US, check the agentic channels in your platform and the UCP interest form in Merchant Center,
• if you run your own platform or API, consider a small MCP server with product search and order status — it is days of work, not months,
• either way, fix your product data first (next section).

How do you make product data agent-ready?

A complete, fresh product feed, with exactly the same facts on the product page. Agents read fields, not photos. Where a field is missing, an agent guesses, or picks a competitor that filled it in.

Google Merchant Center supports the UK and GBP, and its feed is the foundation for AI Mode and Gemini too. In January 2026 Google announced dozens of new “conversational” attributes — answers to common product questions, compatible accessories, substitutes — starting with a small group of retailers.

Data	Why an agent needs it	Where to put it
GTIN (barcode), SKU, brand, MPN	matches the same product across retailers to compare prices	feed; gtin, sku, brand in structured data
Price in GBP, sale price with dates	agents compare offers and stay within budget	feed; Offer.price, priceCurrency, priceValidUntil
Stock and condition	an agent will not recommend what cannot be bought	feed; availability, itemCondition
Delivery cost and time	“delivered by Friday for under £5” is a typical instruction	feed; shippingDetails or an Organization-level policy
Returns policy	shoppers ask the agent whether they can send it back	feed; MerchantReturnPolicy
Size, colour, material, dimensions, compatibility	answers the agent’s clarifying questions	separate attributes, not one long description
Factual description	models summarise; adjectives without facts add nothing	product description, on-page Q&A

Three rules that hold up in practice:

1. One source of truth. Price and stock must match across the feed, the markup and the page. A mismatch tells an agent your data cannot be trusted.
2. Facts in fields, not in images. A size chart saved as a picture is invisible to an agent.
3. Freshness. OpenAI’s feed documentation recommends full catalogue snapshots at least daily. The faster sold-out products leave your feed, the less often agents recommend something nobody can buy.

The Office of the Small Business Commissioner’s agentic commerce guidance (June 2026) makes the same point for smaller firms: audit how AI tools describe your business, organise product data with consistent identifiers, delivery and returns details, and keep business details identical across platforms.

What structured data should a product page have?

Product with an Offer in JSON-LD on every product page, plus your returns and delivery policy once, under Organization. That is Google’s guidance for merchant listings, and the same markup is read by other systems that parse pages.

Google requires a name, an image and an offer with a price above zero and a currency. Recommended: gtin, sku, brand, availability, itemCondition, reviews, shippingDetails and hasMerchantReturnPolicy. Google recommends declaring your returns and shipping policy at Organization level and using offer-level markup only for exceptions.

A minimal example:

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Brass LED table lamp",
  "image": "https://yourshop.co.uk/img/brass-lamp.jpg",
  "sku": "LMP-024",
  "gtin13": "5012345678900",
  "brand": { "@type": "Brand", "name": "Your Brand" },
  "offers": {
    "@type": "Offer",
    "price": "59.00",
    "priceCurrency": "GBP",
    "availability": "https://schema.org/InStock",
    "itemCondition": "https://schema.org/NewCondition",
    "url": "https://yourshop.co.uk/brass-led-table-lamp"
  }
}

Test with the Rich Results Test and watch the Products report in Search Console. Markup does not guarantee an agent picks you; missing markup nearly guarantees it has to guess.

Should you let AI crawlers into your shop?

AI search crawlers: yes, if you want to be recommended. Training crawlers: a separate decision. The most common mistake I see is a blanket block on every AI bot in robots.txt, followed by surprise that ChatGPT has never heard of the shop.

OpenAI separates its bots. OAI-SearchBot powers search in ChatGPT, and OpenAI says sites that opt out of it will not be shown in ChatGPT search answers (apart from navigational links). GPTBot collects content for training. The settings are independent: you can allow the first and block the second. Other AI companies split their crawlers in a similar way.

Check:

• robots.txt — can AI search crawlers reach product and category pages?
• firewall and CDN — is bot protection returning 403 to well-known AI crawlers?
• HTML — are the description and price in the page source, not injected by a script after a click?
• sitemap — does every product URL carry an up-to-date last-modified date?

On aderlo.city we allow search, answer and training crawlers and state it in robots.txt with Content-Signal: search=yes, ai-input=yes, ai-train=yes. That is our marketing choice: brand visibility matters more to us than protecting our text. A shop with unique product copy may decide differently on training, but I would keep search open.

What does UK law expect from retailers using AI agents?

Consumer law applies unchanged, and the CMA expects you to own whatever your own AI agent does. This is orientation, not legal advice; check the detail for your business with a lawyer.

• Your own agents. The CMA’s guidance of 9 March 2026 says you are responsible for what an AI agent does in the same way you are responsible for what an employee does — even if a third party built it. It asks firms to be open about AI use, not to present an agent as a person, to test before launch, to keep a human in the loop and to fix problems quickly. Breaches of consumer protection law can mean fines of up to 10% of worldwide turnover.
• Cancellation rights. A purchase made through an agent is still a distance contract. Under regulation 30 of the Consumer Contracts Regulations 2013, the cancellation period for goods generally ends 14 days after the day they come into the consumer’s physical possession. Your returns data in feeds and markup should say exactly that, or better.
• Merchant of record. In both UCP and ACP the retailer stays the seller, so pre-contract information, faulty-goods claims and refunds stay with you.
• Selling into the EU? Since 2 August 2026, Article 50(1) of the EU AI Act requires AI systems that talk to people to tell them they are interacting with AI, unless that is obvious from the context.

How did we make aderlo.city agent-ready?

We gave agents everything aderlo.city shows people, in machine-friendly form, without changing a single byte for browsers. aderlo.city is the information site of a 3D shopping city, not a shop, so there is no checkout. But the discovery layers are exactly the ones a retailer needs.

What	Address on aderlo.city	Why an agent cares
API catalogue (RFC 9727)	/.well-known/api-catalog	one place to discover the API, the MCP server and the docs
Read-only JSON API + OpenAPI 3.1	/api/v1/, /openapi.json	facts about the city, MVP status, article list
MCP server	/mcp	tools get_aderlo_city_info, get_mvp_status, list_articles, get_article
Markdown of the home pages, articles and glossary	Accept: text/markdown or /en/blog/<slug>/index.md	clean text without menus and footers, fewer tokens
Link header, llms.txt, sitemap	every HTML page	the agent finds the API description and sitemap without guessing
OAuth 2.1 (PKCE, client registration, consent, revocation)	/.well-known/oauth-authorization-server	two actions on a person’s data: checking and joining the new-districts update list

What carries over to a shop:

1. Agents cannot see 3D. The Aderlo Galleries in London and Warsaw are real-time 3D in the browser, which crawlers do not run. So everything that matters — what the galleries are, MVP status, articles — also exists as HTML, Markdown and JSON. In a shop, the equivalents are configurators, try-on tools and image carousels: whatever drives the decision must also exist as text.
2. Public stays public; personal needs consent. The API and most of the MCP server need no login and are rate-limited per IP address. A token is required only where an agent acts on a specific person’s data — for us, joining the update list, which the person confirms in their browser and can revoke on a “connected apps” page. Linking a customer account to an agent will work the same way in a shop; UCP specifies OAuth 2.0 for it.
3. Browsers get the same HTML. We serve Markdown only for Accept: text/markdown, with Vary: Accept, so caches never hand it to people.
4. The shop directory is a separate service. The list of shops and units in the galleries comes from a separate MCP server run by our platform, GAPLI, listed in our API catalogue as an external resource. A retailer’s catalogue can live on its own server too, with the brand site simply pointing to it.

You can try it yourself — list the tools of our MCP server:

curl -s https://aderlo.city/mcp -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

What should UK retailers do in the next 90 days?

Audit visibility, fix data, open the doors to search crawlers, and let your platform carry the protocols. This is the plan I would run in a small or mid-sized UK shop:

1. Week 1 — audit. Ask ChatGPT, AI Mode and Copilot about your brand, your categories and your three best sellers. Note what they say, where they get it from and what they get wrong.
2. Weeks 1–2 — access. Check robots.txt, firewall and CDN: AI search crawlers should get product pages with a 200.
3. Weeks 2–4 — feed. In Merchant Center, fill in GTIN, brand, stock, delivery cost and time, and returns policy; remove products you cannot sell.
4. Weeks 3–6 — structured data. Product + Offer on every product page, Organization with returns and shipping policy; test it.
5. Weeks 4–8 — content. Rewrite the descriptions of your 50 best sellers: facts, dimensions, materials, use cases, plus on-page questions and answers.
6. Weeks 6–10 — measurement. Separate traffic from AI assistants (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com) in your analytics. In our GA4 it is its own channel, placed before Referral.
7. Weeks 8–12 — protocols and your own agents. Ask your platform and payment provider about UCP, ACP and MCP for UK customers. If you deploy your own AI assistant, run it against the CMA’s checklist: disclosure, testing, human oversight.

Where do 3D galleries fit next to AI agents?

They solve different moments of shopping. An agent is best when the customer knows what they want — “reorder the same batteries as last time”. Walking through a gallery is better for inspiration, gifts and shopping with someone else, which I cover in What Is a 3D Virtual Mall?.

Aderlo City does both: the Aderlo Galleries in London and Warsaw are for people, and data about the city, shops and articles is for agents, through our API and MCP server. We plan to make brands with units in the galleries as visible to AI assistants as they are on a 3D walk. If you want your brand in the gallery, you can reserve a unit — the same immersive commerce story, seen from the retailer’s side.