Agentic commerce for retailers: what's live, what's hype and what to fix first

Direct answer: As of October 2026, AI assistants research and compare products for shoppers worldwide, but in-chat checkout is live mainly in the US. Copilot Checkout, Shopify's direct checkout and Google's UCP checkout serve US buyers, Google's checkout also covers eligible products in Canada and Australia, and OpenAI has moved ChatGPT purchases back to merchants' own sites. In every market, the work that pays now is the same: let agents reach your pages, make your price, stock, delivery and returns machine-readable, and make sure an agent can get to your checkout.

This guide separates what is live from what is announced, market by market. It also shows what our own benchmark found across 225 UK and US retailers, and gives a prioritised plan for a mid-market ecommerce team. Every platform status below is dated, because it will change.

What agentic commerce means for a retailer

"Agentic commerce" covers three different things that often get blurred:

1. AI-assisted research. A shopper asks ChatGPT, Google AI Mode, Perplexity, Copilot or Claude for a recommendation, then clicks through and buys on your site. This is live in most markets. Google AI Mode, for example, reached the UK on 28 July 2025, and OpenAI's help centre describes how ChatGPT ranks merchants on availability, price, quality and whether they are the maker or primary seller.
2. Agents that act on the web. Browser agents open your site, use your filters, add to basket and fill in checkout on the shopper's behalf. They do not need your permission or a protocol. They just need your site to work for them.
3. Protocol-based checkout. The assistant talks to your backend through a standard (UCP, ACP) and the shopper pays inside the chat while you remain merchant of record. This is the part with the headlines, and it is live mainly for US buyers.

If you sell to US shoppers, all three matter today. Elsewhere, the first two matter today and the third is a roadmap item. We call the discipline of being chosen by, and transactable for, these agents agent conversion. It is different from AI visibility, which measures whether you are mentioned.

What is live, market by market (as of October 2026)

By channel

By market

A retailer based outside the US can still reach US in-chat checkout. Shopify's Copilot requirement is that "Your store must sell to customers in the United States": it is about where you sell, not where your business is based.

Three points stand out.

Checkout is moving back to merchants, not away from them. OpenAI launched in-chat Instant Checkout in September 2025. Its merchant page now says: "We're moving away from a standalone Instant Checkout experience in ChatGPT… purchases are completed on merchant-owned websites or apps." It also says there are "no fees on purchases that start in ChatGPT". Google states that with UCP checkout "you will remain the seller of record". Microsoft says Copilot Checkout "preserves the merchant as merchant of record". Even where checkout happens inside the assistant, your systems accept the order, take the payment and handle fulfilment and returns.

Rollout runs US first, then Canada and Australia, then the UK. In its May 2026 announcement, Google said UCP-powered checkout "will roll out across Canada and Australia in the coming months, and later to the U.K." Its Merchant Center help page now lists US, Canadian and Australian eligibility. Microsoft says it cannot give a timeline for other regions.

Feeds are the documented way in. OpenAI says most merchants "start with product feeds". You can appear without one if ChatGPT crawls your site, but "feeds give you greater control". Shopify and Etsy catalogues are already integrated. Google's discovery runs on Merchant Center data.

What is hype, or at least not yet evidence

Treat the following carefully in board papers and agency proposals.

• "Websites are dying." The best-documented in-chat checkout was scaled back within six months. OpenAI told CNBC that the first version "did not offer the level of flexibility that we aspire to provide". Every major platform keeps the retailer as merchant of record.
• Big forecasts. McKinsey's "up to $1 trillion" figure for US B2C retail by 2030 is revenue orchestrated by agents, which includes research that ends in a normal human purchase. It is not money agents spend on their own. Read the definition before quoting the number.
• "AI traffic converts 60% better." This is Adobe's July 2026 figure for US retail sites, from a vendor that sells the measurement. Shoppers who arrive from an AI answer are often further along in their decision, so higher conversion is partly a selection effect. Volumes are still small. Retailers interviewed by Digital Commerce 360 put AI at roughly 5% of referrals at leading retailers, and one reported AI conversion below its site average. Measure your own channel before moving budget.
• "Add llms.txt and you're AI-ready." Google's guidance on AI features says you don't need "AI text files" to appear in AI Overviews or AI Mode. An Ahrefs study of 137,210 domains found that 97% of the llms.txt files it observed received no requests in May 2026. Our own scoring spec includes it as a single light check (LEG-11) for that reason.
• "You need a protocol strategy now." The protocols are real and moving fast. UCP has shipped several dated versions in 2026, and ACP is still labelled beta. Outside the US, and for US retailers whose platform does not provide it, protocol work usually comes after the basics in this guide. For one emerging browser-side standard, see our guide to WebMCP for ecommerce.

What our Agent Commerce Index found

The Agent Commerce Index Q4 2026 scores 225 UK and US online retailers across 8 sectors on whether an AI shopping agent can find them, read their offer, trust them and complete a purchase. The headline figures, as published:

The gap between 7% and 83% is the most useful finding for a retailer. The method notes explain how it was measured. The basket test added one in-stock product to an anonymous basket "as ChatGPT's agent" using plain page requests, then opened the checkout page, and nothing was ever submitted. Where the checkout turned the agent away, the same steps were repeated as an ordinary browser. In the published per-store data, the Index marked no store as "agents refused", meaning none let a plain browser request through while refusing the agent identity. What separated the two agent types was how the requests were made, not who claimed to be making them. Our interpretation is that bot protection which checks for a real browser on basket and checkout pages, and basket actions that only work after JavaScript runs, are the likely causes. Either way, an agent that fetches pages rather than driving a browser hits a wall, and the sale goes to a store it can read. The Index methodology guide explains how to compare your store.

The five things an agent needs from your store

Our open Agent Conversion Score specification (v0.6, CC BY 4.0) breaks readiness into 49 checks across five weighted dimensions. These map directly onto what to fix. (The Q4 2026 Index was scored against the previous version, v0.5.)

What to fix first: a prioritised plan

Ordered by what unblocks the most value for the least effort. Most mid-market teams can complete steps 1 to 4 within a quarter without new platforms. US sellers should run step 5 in parallel, because the channels are open.

1. Measure your baseline (week 1)

• Run a free Agent Conversion Score on your domain and two competitors. It takes about a minute and shows your blocker.
• In your analytics, create a channel for AI referrals (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and similar). Record sessions, conversion rate and revenue per session for the last 90 days. That is your channel size, not Adobe's.

2. Stop blocking the agents you want (weeks 1 to 2)

• Review robots.txt for the user agents that matter to shopping: OpenAI's OAI-SearchBot and ChatGPT-User, PerplexityBot and Perplexity-User, Claude-User, and Google-Extended. You can block training crawlers and still allow shopping and search agents. They are separate user agents.
• Check your CDN and firewall, not just robots.txt. Cloudflare now classifies AI traffic as Search, Agent or Training. Its documentation says that from 15 September 2026 new domains block Agent and Training traffic on pages that display ads by default.
• Prefer verifying signed agents to trusting user-agent strings. OpenAI documents that ChatGPT Work's Cloud browser signs requests with Web Bot Auth (RFC 9421 HTTP message signatures), and Cloudflare verifies these automatically.

3. Make every product page machine-readable (weeks 2 to 6)

• Product and Offer JSON-LD on every product page, with a numeric price, priceCurrency, availability, brand and an identifier (GTIN, MPN or SKU).
• Make sure the structured price matches the displayed price. UK and EU pages show VAT-inclusive prices while US pages show pre-tax prices. Stale cached JSON-LD, currency switchers and promotion banners are common sources of mismatch.
• Put name, price and stock in the HTML the server sends, not only after JavaScript runs.
• Add delivery cost and time, and returns terms, in structured data (OfferShippingDetails, MerchantReturnPolicy), and link policy pages from product pages. Delivery promises and returns windows, including statutory rights such as the 14-day cancellation period for distance sales in the UK and EU, are exactly the facts an agent compares.

4. Remove agent friction from basket and checkout (weeks 4 to 8)

• Offer guest checkout. Forced account creation stops an agent outright.
• Replace blanket captchas on basket and checkout with risk-based challenges that let verified agents through.
• Test it. Ask a browser agent to buy a specific variant and see where it stalls. Then fetch the product page without JavaScript and check whether price and the add-to-basket action are still there.

5. Get your feeds channel-ready (weeks 4 to 10; now if you sell to the US)

• Keep Google Merchant Center in good standing with returns, delivery and customer support details complete. Discovery in AI Mode already draws on it. In the US, Canada and Australia, Google's UCP checkout requires products to carry the native_commerce(checkout_eligibility) attribute, ideally in a supplemental feed. Elsewhere, have it ready for launch.
• Apply for a direct ChatGPT feed. OpenAI's feed spec uses full daily snapshots matched on item_id. Shopify and Etsy catalogues are already integrated, and other applicants join a waitlist.
• US sellers not on Shopify: review the Copilot Checkout routes via PayPal or Stripe, and Microsoft Merchant Center, which needs a UCP-compliant feed.

6. Decide your protocol and dispute posture (quarter 2 onwards)

• Choose a protocol when a channel you sell into opens. If you are on Shopify, the platform handles much of this. On WooCommerce, Adobe Commerce or BigCommerce, support varies and changes monthly. Check vendor documentation before scoping.
• Plan your evidence for disputes. Liability for "my agent bought the wrong thing" is not settled. Worldpay's view is that no liability shift exists yet. It notes that the US rules for electronic payments (Regulation E) assume a transaction is either authorised or not, and that in the UK and EU, PSD2 strong customer authentication can stop an agent transaction completing at all. Keep order confirmations, the price and policy shown at checkout, and the agent identity for every agent-originated order. This is not legal advice.

Readiness checklist

Use this as a one-page audit. Each line maps to a check in the spec.

• robots.txt allows OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User and Claude-User on product pages
• Product pages and the homepage return 200 to agent identities, with no 403, 429 or JavaScript challenge
• No cookie banner, age gate or geo-redirect hides product content or price
• Product JSON-LD with numeric price, currency, availability, brand and identifier
• Structured price equals displayed price, including VAT and promotions
• Name and price in server-rendered HTML
• Variants expressed structurally, not only in a JavaScript picker
• Delivery cost and time, and returns terms, stated on the page and in structured data
• Shipping and returns policy pages discoverable and linked from product pages
• Guest checkout available; no captcha on basket or checkout
• A machine-readable product feed exists (Merchant Center at minimum)
• AI referrals tracked as their own channel

Common mistakes

• Blocking everything "AI" in one setting. This removes you from the research step where the shortlist is made, and that step is live in every market.
• Fixing robots.txt but not the firewall. A 403 or challenge from a bot-management rule stops an agent as surely as a robots.txt disallow, and it rarely shows up in your analytics because the agent never loads your tags.
• Treating schema as an SEO checkbox. Rich-result eligibility needs less than an agent comparing offers. Delivery, returns and variant availability are usually missing.
• Price mismatches. A structured price that disagrees with the page is worse than none, because it invites a misquote. In the UK, for example, giving consumers false or misleading information, including about price, can be a "misleading action" under section 226 of the Digital Markets, Competition and Consumers Act 2024, which replaced the CPRs from April 2025. US and EU consumer-protection rules also apply. Not legal advice.
• Waiting for in-chat checkout to reach your market. By the time it launches, the eligibility requirements (accurate feeds, complete policies, good Merchant Center standing) will be the same work described here.
• Buying a protocol integration before the basics. A UCP profile in front of a page that returns 403 to agents achieves nothing.
• Quoting vendor stats as your forecast. Use your own referral data and say whose numbers you are quoting.

How ConversionVector helps

• Agent Conversion Score: the free score shows your blocker in about a minute; the full report ranks every fix by how much it moves the score.
• Simulated Agent Buyer: real browser agents try to buy from you with set intents, so you see the exact step and reason they stall.
• Recommendation Monitoring: weekly share of recommendation on buy-intent prompts across ChatGPT, Perplexity, Gemini and Claude, with the reason you lost.
• Hosted Agent Layer: for stores whose platform lacks native support, a hosted structured catalogue, discovery endpoints and verified-agent access rules (waitlist).

Platform statuses in this guide were last verified against vendor documentation on 23 October 2026 and will change. Index figures are from the published Agent Commerce Index Q4 2026.