AI search, explainedPart 6 of 6
What is agentic commerce? How it works, and who is ready
Agentic commerce is buying and selling where an AI agent researches, compares and completes a purchase for a person, usually with the person approving the final step. It is live in ChatGPT, Google's AI Mode, Copilot, Perplexity and Amazon, but when we sent a polite, self-identified agent to 60 large US retailers in November 2026, 42% turned it away at the homepage.
What is agentic commerce?
Agentic commerce is shopping done by software acting for a person: an agent finds products, compares them, fills the cart and pays, within limits the person sets. Google's own AI Overview describes it as agents that "research, negotiate, compare, and complete transactions on behalf of consumers or businesses with little to no direct human intervention."
In practice most purchases still need a human yes at the end. Gemini's auto browse in Chrome is "designed to pause and explicitly ask for your confirmation" before buying, and Accenture's survey of 25,590 people found that 32% are ready to hand over purchase decisions but only 9% are open to fully autonomous shopping.
No standards body defines the term. The rules come from vendor-led specs, mainly the Agentic Commerce Protocol (ACP) from OpenAI and Stripe and the Universal Commerce Protocol (UCP) from Google and its retail partners. For the done-for-you side of this work, see our agentic commerce optimization service.
Does agentic commerce exist yet?
Yes, but the volume is small and the checkout layer keeps changing. In December 2025 Visa said "hundreds of secure, agent-initiated transactions have now been successfully completed", which tells you how early that was.
Traffic is growing faster than purchases. Adobe counted a 693% rise in visits to US retail sites from generative AI tools over the 2025 holiday season, across more than a trillion visits. Its Q3 2026 report found that AI's share of retail visits grew 138% year on year in May 2026, and that AI-referred visitors converted 54% better than other traffic.
This timeline uses the companies' own announcements.
| Date | What happened |
|---|---|
| Nov 2024 | Perplexity launches "Buy with Pro" checkout for US Pro users |
| Apr 2025 | Amazon tests "Buy for Me", its agent buying on other brands' sites; Visa announces Intelligent Commerce; Mastercard announces Agent Pay |
| Sep 2025 | Google publishes the Agent Payments Protocol (AP2); OpenAI and Stripe launch Instant Checkout in ChatGPT and ACP, with Etsy sellers first |
| Oct 2025 | Visa launches the Trusted Agent Protocol with Cloudflare; Walmart and PayPal sign up for ChatGPT checkout |
| Nov 2025 | Amazon sues Perplexity over its Comet browser shopping on Amazon; Google starts agentic checkout in Search and AI Mode |
| Dec 2025 | Instacart offers checkout inside ChatGPT; Shopify makes "every Shopify store agent-ready by default"; Stripe launches its Agentic Commerce Suite |
| Jan 2026 | Microsoft launches Copilot Checkout; Google announces UCP at NRF with Shopify, Etsy, Wayfair, Target and Walmart; Gemini in Chrome gets auto browse |
| Mar 2026 | UCP adds cart, catalog and identity linking; OpenAI moves away from standalone Instant Checkout; Shopify stores go live in ChatGPT, Copilot, AI Mode and Gemini; Amazon says tens of millions of products can be bought through Buy for Me |
| Apr 2026 | ACP publishes its 2026-04-17 version, still labelled beta |
| May 2026 | Amazon launches Alexa for Shopping, which includes Buy for Me |
| Sep 2026 | Meta launches Muse, a personal agent that checks out with Stripe Link |
| Oct 2026 | Google expands ads in AI Mode |
The March 2026 change is the one most guides miss. OpenAI's merchant page now says: "We're moving away from a standalone Instant Checkout experience in ChatGPT", and that "purchases are completed on merchant-owned websites or apps." CNBC reported that Walmart's in-chat checkout had converted three times worse than sending shoppers to its own site; Walmart has not published that figure itself.
How does agentic commerce work?
An agent needs four things from a merchant: product facts it can read, a way into the checkout, a payment credential it is allowed to use, and a site that lets it in. Each protocol covers a different part of that.
Product facts come from feeds and from the page. OpenAI's product feed requires fields such as item_id, title, price, availability and url, and asks merchants to update prices and stock "when they change." Google Merchant Center insists the price must "match with the price from your landing page, structured data, and at checkout."
Checkout runs through a protocol. ACP defines checkout session endpoints a merchant hosts (create, update, complete, cancel) and a delegated payment step that returns a single-use vault token with a maximum amount and an expiry. UCP has merchants publish a profile at /.well-known/ucp that tells platforms which capabilities and endpoints they support, over REST, MCP, A2A or an embedded flow.
Payment uses scoped credentials, not raw card numbers. Stripe's Shared Payment Tokens "grant sellers scoped access to customers' payment methods for agent-initiated purchases" and can be resolved only once. AP2 records what the person approved as signed mandates; version 0.2 calls them a Checkout Mandate and a Payment Mandate, replacing the Intent and Cart Mandate names it launched with.
Browser agents skip the protocols and use the website like a person would. Gemini in Chrome, Perplexity's Comet and ChatGPT's agent all browse and fill forms. So the plain website still matters. If a product page needs JavaScript to show its price, or bot management blocks the agent, no protocol helps.
ACP vs UCP vs AP2 vs MCP: what is the difference?
ACP and UCP are checkout protocols backed by rival camps, AP2 is about proving what a person authorised, and MCP is the general way agents call tools. Most merchants will meet them through their platform rather than build them.
| Protocol | Led by | What it does | Status (Nov 2026) |
|---|---|---|---|
| ACP (Agentic Commerce Protocol) | OpenAI and Stripe, Apache 2.0 | Checkout sessions and delegated payment between an agent and a merchant; the merchant stays merchant of record | Beta; version 2026-04-17 |
| UCP (Universal Commerce Protocol) | Google, co-developed with Shopify and large retailers | Discovery through /.well-known/ucp, checkout, cart, catalog, orders and identity linking | Live for select US merchants through Merchant Center |
| AP2 (Agent Payments Protocol) | Google, with more than 60 launch organisations | Signed mandates recording what the person approved; used inside UCP | Version 0.2 |
| MCP (Model Context Protocol) | Agentic AI Foundation, Linux Foundation | Lets agents call tools; Shopify runs catalog, cart, checkout and order MCP servers | Spec version 2026-07-28 |
Which retailers are ready for AI agents?
Most of the largest US retailers are not ready for an unknown agent, and the main barrier is bot management, not robots.txt. On 14 November 2026 we sent a self-identified agent to 60 retailers: the first 50 qualifying names on the NRF Top 100 Retailers 2025 list and 10 well-known DTC brands.
- 25 of 60 homepages (42%) refused a plain request or came back unusable, and 23 of 60 (38%) still did when the agent ran a real headless browser. For the plain request, that was 46% of the NRF 50 and 20% of the DTC brands.
- 17 of 60 (28%) refused to serve robots.txt at all, so their rules never came into play.
- Of the 43 robots.txt files we could read, only Amazon's disallows all six AI agents we tested (OAI-SearchBot, ChatGPT-User, GPTBot, PerplexityBot, Google-Extended, ClaudeBot) across the site.
- Only 21 of 60 (35%) served a product page we could read. Of those, 16 had Product schema with a price in the raw HTML, so 16 of 60 (27%) gave an agent price data without JavaScript.
- On pages that did load, the price was in the raw HTML 90% of the time. Brooklinen was the only one adding its Product schema with JavaScript, and Target served our agent a product page with no price at all.
- 17 of 60 (28%) publish an llms.txt file: 11 of the NRF 50 and 6 of the 10 DTC brands. Our llms.txt guide explains what it does and doesn't do.
- 11 of 60 (18%) publish a valid UCP profile at
/.well-known/ucp. Among the NRF 50 that is 4: Ace Hardware, Chewy, Academy Sports and IKEA. Seven of the 10 DTC brands have one, all on Shopify.
Shopify's "agent-ready by default" shows in the numbers: the DTC brands were readable 80% of the time and most had llms.txt and a UCP profile without building anything. And several UCP co-developers we could reach, Walmart, Target, Best Buy, Wayfair and Home Depot, return a 404 at /.well-known/ucp. They may sell through Google's surfaces by other routes; the profile just isn't public at that address.
How big will agentic commerce get?
Forecasts for US agentic commerce in 2030 run from about $300 billion to $1 trillion, and the spread comes from how each firm defines the term. Treat every figure here as a forecast.
Bain forecasts $300 billion to $500 billion, about 15% to 25% of US e-commerce by 2030, counting purchases "initiated, influenced, or completed" by agents and excluding journeys that use AI only for search. McKinsey's widely reported estimate goes up to $900 billion to $1 trillion of US "agent-orchestrated" revenue by 2030; we could not open its report to check the wording, and orchestrated revenue is not the same as revenue merchants keep. Adobe forecasts AI traffic to US retail sites up 130% for the 2026 holiday season.
What are the risks of agentic commerce?
The open questions are fraud liability, who owns the customer, returns and whether you can say no to agents. None is fully settled.
Fraud is rising before the rules are written. Visa reported a 25% rise in malicious bot-initiated transactions over six months, and 40% in the US, and says cardholders are never held responsible for unauthorised charges. No card network has published how chargebacks for agent purchases will be split.
The merchant keeps the customer on paper. ACP says businesses "maintain their customer relationships as the merchant of record", and Google tells UCP merchants "You remain the Merchant of Record". But the agent owns the conversation and the comparison, which is where brand preference forms.
Returns stay the merchant's job. Google requires return cost, return window and a policy link before a product can use UCP checkout, and OpenAI's feed has return fields. Adobe's survey found that 69% of AI shoppers say they are less likely to return an item bought with AI help, which is self-reported.
Blocking agents is legal territory now. Amazon blocks the AI agents in robots.txt and sued Perplexity over Comet. A law firm's summary of the August 2026 Ninth Circuit ruling says the court vacated the injunction on the computer-fraud claims, finding it was "the user who 'accessed' Amazon's computers", while the contract claims continue. User-triggered fetchers such as ChatGPT-User and Perplexity-User say robots.txt rules may not apply to them.
How should retailers prepare for agentic commerce?
Start with the website, because every protocol and every browser agent depends on it. In our view, these come first:
- Make product facts readable without JavaScript: name, price, availability and return policy in the HTML and in Product schema, matching your feed.
- Check your bot management as well as robots.txt. Decide which agents you want, and make sure your CDN or bot vendor lets them through. That was the biggest gap in our test.
- Claim your channels: Google Merchant Center, and the ChatGPT merchant listing (Shopify and Etsy catalogs need no extra setup; other merchants join a waitlist).
- If you are on Shopify, check that Agentic Storefronts are on and read the llms.txt and UCP profile Shopify generates for you.
- Measure: run the questions your buyers ask in ChatGPT, Gemini and Perplexity, and note whether your products appear and which sources are cited instead.
A custom ACP or UCP build can wait if your platform doesn't support one yet, and so can agent-specific loyalty, account linking and negotiation features, which are still on roadmaps. OpenAI's six-month turn on Instant Checkout is a good reason not to bet heavily on one checkout flow.
To see how an agent reads one of your product pages today, run Agent Ready, our free extension. Our page on agent-ready websites covers the fixes, and if you'd rather we did the work, that's our agentic commerce optimization service. For how this fits with getting cited in AI answers, read AEO vs GEO vs SEO, or go back to the AI search series.
How we measured
We checked 60 retailers on 14 November 2026 between 06:36 and 07:31 UTC from one US data-centre server. The sample is not random: the first 50 qualifying names on the NRF Top 100 Retailers 2025 list (we skipped grocery, convenience, fuel, telecom and chains without US e-commerce) and 10 DTC brands we chose: Allbirds, Warby Parker, Glossier, Bombas, Brooklinen, Away, Everlane, Casper, Gymshark and SKIMS.
Each request named itself as WorkPodsResearch/1.0 (+https://workpods.ai), one at a time with a pause between them. For each site we read robots.txt, /llms.txt, the homepage, one product page and /.well-known/ucp, fetching pages both as raw HTML and in headless Chromium. We reviewed the edge cases by hand, for example counting Nordstrom as blocked after its pages redirected to a closed-site notice.
Sites may let verified crawlers from OpenAI, Google and others through by IP address or signature while blocking an unknown agent like ours. So "blocked" here means blocked for a self-identified agent that is not on a verified list, which describes most new agents and tools. It was one product page per site, at one moment, from one IP address.
Questions
Questions buyers ask
What is agentic commerce in simple terms?
It is when an AI agent shops for you: it finds products, compares them and buys, usually after you approve. The agent might be in ChatGPT, Google's AI Mode, Copilot, Perplexity or Amazon's Alexa.
Is agentic commerce the same as ACP?
No. ACP, the Agentic Commerce Protocol from OpenAI and Stripe, is one of the protocols agentic commerce runs on. Google's UCP is the other main one, and many agents simply use the website like a person would.
Do I need to build ACP or UCP myself?
Usually not. Shopify, Stripe, PayPal and Google Merchant Center wrap the protocols for merchants, so most stores should fix their product data and bot access first and build only if their platform lacks support.
Can I stop AI agents from buying on my site?
You can try with robots.txt and bot management, but user-triggered agents such as ChatGPT-User say robots.txt may not apply to them, and blocking also turns away shoppers who send an agent. Amazon is blocking and litigating; most retailers in our test blocked unknown agents by accident through bot management.