Agentic commerce is shopping in which an AI agent does part or all of the buying work: it finds products, compares them against a shopper's limits and, in some cases, pays. It runs on shared protocols that let an agent read a merchant's catalog and start checkout, and on payment tokens that keep card numbers hidden. In 2026 the clearest pattern is discovery inside AI tools, with checkout often completed on the merchant's own site.
Last updated November 2, 2026.
Summary: Agentic commerce moves AI from recommending products to acting on a shopper's behalf. Several protocols compete to standardize how agents read catalogs, check out and pay, including OpenAI and Stripe's ACP, Google's UCP and AP2, and card network token systems from Visa and Mastercard. Bain and McKinsey forecasts for 2030 range from $300 billion in US sales to $5 trillion globally, depending on how each defines the term. Brands win by keeping product data complete, letting agents reach their sites, and earning the independent coverage that AI systems cite.
What does this guide to agentic commerce cover?
- What agentic commerce is and how it differs from AI shopping help
- How an AI agent completes a purchase
- Which protocols run agentic commerce
- What happened to in-chat checkout in 2026
- How big agentic commerce could get
- Whether retailers can block shopping agents
- A seven-step plan for brands
- The main risks
- Frequently asked questions
What is agentic commerce, and how does it differ from AI shopping?
Agentic commerce is a model in which AI agents act for a buyer across the shopping journey instead of only answering questions. Bain defines it as purchases that are initiated, influenced or completed by third-party AI agents or retailer-hosted agents, and it excludes journeys that only use AI-assisted search or discovery. That definition separates three stages that are often blurred together.
| Stage | What the AI does | Who makes the final decision |
|---|---|---|
| Research | Answers questions and compares products in a chat | The shopper, on the merchant's site |
| Recommendation | Shows products, side-by-side comparisons and prices | The shopper, who clicks through to buy |
| Action | Builds a cart and pays within limits the shopper set | The agent, inside those limits |
Most of today's activity sits in the first two rows. Bain's analysis found that 30% to 45% of US consumers already use generative AI to research and compare products, and it cited Salesforce's estimate that AI and agents influenced about $3 billion of US Black Friday sales. The same analysis noted that most shoppers are not yet comfortable letting AI handle a purchase from start to finish.
How does an AI agent complete a purchase?
An agent completes a purchase by moving through a fixed sequence: it takes a goal, searches catalogs, builds a cart, gets approval, pays with a limited credential and leaves fulfillment to the merchant. The steps below describe the common flow across the major systems.
- The shopper states a goal and limits, such as a product type, a budget and a delivery date.
- The agent searches merchant catalogs and product feeds and compares the options.
- The agent builds a cart with the chosen merchant.
- The shopper approves the cart, or the agent acts within limits set in advance.
- A payment credential scoped to that merchant and amount is issued, so the agent never handles the real card number.
- The merchant fulfills the order and handles returns and support as it does for any other order.
What is a shared payment token?
A shared payment token is a limited credential that lets an AI application start a payment without seeing the buyer's card details. In the Agentic Commerce Protocol, Stripe issues a token scoped to a specific merchant and cart total, and the agent passes it to the merchant through an API to complete the purchase. Visa and Mastercard use similar tokenized credentials on their own networks.
Which protocols run agentic commerce?
Six systems matter most today, and none has won outright. Some cover how agents read catalogs and check out, some cover how agents pay, and one covers how agents connect to tools at all.
| System | Backed by | What it covers | Announced |
|---|---|---|---|
| Agentic Commerce Protocol (ACP) | OpenAI and Stripe | Product feeds, checkout and payment handoff; extended in March 2026 to discovery | September 29, 2025 |
| Universal Commerce Protocol (UCP) | Google, with Shopify, Etsy, Wayfair, Target and Walmart | Discovery, cart, checkout and post-purchase steps | January 11, 2026 |
| Agent Payments Protocol (AP2) | Authorizing payments that an agent starts | September 2025 | |
| Visa Intelligent Commerce | Visa | Tokenized credentials, spending controls and agent verification | April 2025; Trusted Agent Protocol added October 2025 |
| Mastercard Agent Pay | Mastercard | Agentic Tokens issued for verified agents | April 29, 2025 |
| Model Context Protocol (MCP) | Anthropic, open source | How agents connect to tools and data, including merchant systems | November 2024 |
The pieces are designed to fit together. Google built UCP to work alongside AP2, the Agent2Agent protocol and MCP, and OpenAI and Stripe published ACP as an open standard so that merchants not using Stripe can adopt it with their own payment providers, according to Payments Dive. Large retailers do not have to pick one: Walmart announced a Gemini partnership built on UCP at NRF 2026 while also working with ChatGPT.
What happened to in-chat checkout in 2026?
OpenAI moved away from in-chat checkout in March 2026 and refocused ChatGPT shopping on product discovery. Instant Checkout launched on September 29, 2025, starting with Etsy sellers and with Shopify merchants promised next. On March 24, 2026, OpenAI said the first version did not offer the flexibility it wanted and that merchants could use their own checkout experiences instead, as TechCrunch reported.
OpenAI told TechCrunch that merchants could still offer checkout through apps inside ChatGPT, and that it would keep supporting several checkout paths, including merchants' own websites. What ChatGPT now emphasizes is richer product information: side-by-side pictures, prices, features and reviews, powered by merchant data shared through ACP.
Google took a different route. It announced UCP-powered checkout on eligible product listings in AI Mode in Search and in the Gemini app for US retailers. The lesson for brands is practical: the product data layer has outlasted any single checkout button, so the data is where the preparation belongs.
How big could agentic commerce get?
Forecasts for 2030 range from hundreds of billions to trillions of dollars, and the spread comes mostly from how each firm defines the term. Treat the figures as scenarios, not measurements.
| Source | Estimate for 2030 | What it counts |
|---|---|---|
| McKinsey's forecast | Up to $1 trillion in US B2C retail revenue; $3 trillion to $5 trillion globally | Revenue "orchestrated" by agents |
| Bain's forecast | $300 billion to $500 billion in the US, or 15% to 25% of e-commerce | Purchases initiated, influenced or completed by agents, excluding AI-assisted search alone |
Bain expects spec-driven purchases, such as household essentials where price, speed and availability decide the sale, to shift first. Apparel and travel are expected to move more gradually as trust grows. That ordering matters for brands: the more a purchase can be described by specifications, the sooner an agent can make it without a human weighing brand feel.
Can retailers block shopping agents?
The law is unsettled, and the first appeals ruling favored the agent. In March 2026 a district court granted Amazon a preliminary injunction against Perplexity's Comet shopping agent. On August 4, 2026, the Ninth Circuit vacated it, reasoning that because Comet needs user direction to use Amazon, the user is the one accessing Amazon's servers, according to Engadget's report.
The ruling did not end the case. Amazon said it disagreed and was evaluating next steps, and the lawsuit continues in the district court. For brands, the practical point is that agent traffic may arrive whether or not a retailer approves of it, so access policies, bot rules and terms of service need a deliberate decision rather than a default.
What should a brand do to be chosen by shopping agents?
A brand should make its products easy to read, easy to trust and easy to buy, in that order. The seven steps below apply whether the agent sits in ChatGPT, Gemini or a retailer's own app.
- Complete the product data. Fill every attribute an agent could compare: price, availability, size, materials, identifiers, shipping and return terms. Gaps are where agents skip a product.
- Keep feeds current. Update price and availability whenever they change, because an agent that quotes a stale price loses the shopper's trust in both the agent and the merchant.
- Write answer-first product and policy pages. Open each page with a direct answer in 40 to 80 words, and put returns, warranty and delivery rules in plain text. The method is explained in the guide to what AEO is.
- Let the right crawlers in. Check robots.txt and firewall rules against the agents you want. OpenAI documents that sites opted out of OAI-SearchBot will not appear in ChatGPT search answers, as covered in the guide to getting recommended by ChatGPT.
- Support more than one protocol through your commerce platform. With no clear winner, the lowest-risk path is a platform that supports ACP, UCP or both, and a payment provider that accepts tokenized agent credentials.
- Earn independent coverage and reviews. Agents weigh third-party evidence. A peer-reviewed study presented at KDD 2024, GEO: Generative Engine Optimization, found that adding citations, quotations and statistics raised visibility in generative responses by up to 40% in its tests, with results that varied by domain.
- Test with real prompts. Run a fixed list of shopping prompts monthly across the main assistants and record whether your products are named, how they are described and which sources are cited. The GEO audit checklist and an AI citation audit show how to set up that baseline.
| What an agent needs | Where it looks | Brand action |
|---|---|---|
| Structured product facts | Product feed, catalog API, page markup | Complete, accurate, frequently refreshed data |
| Proof of trust | Reviews, press coverage, reference sites | Earned coverage and consistent company facts |
| A clear way to buy | Checkout, merchant app or protocol endpoint | Platform support and tokenized payments |
| Permission to visit | robots.txt, bot rules, terms | Deliberate access policy for each agent type |
Programs that cover this work end to end are described on the generative engine optimization practice page.
What are the main risks of agentic commerce?
The main risks are trust, commoditization and control. Each one changes what a brand should protect.
- Trust. Bain found that most shoppers are not yet comfortable letting AI complete a purchase on its own, so adoption will depend on clear limits, receipts and easy reversal.
- Commoditization. When an agent compares on specifications, brand feel counts for less. Brands need distinctive, verifiable facts that an agent can weigh, not only creative claims.
- Control of data and checkout. Bain advises retailers to keep as much ownership of data, fulfillment and checkout as they can, because third-party AI platforms can otherwise sit between a brand and its customer.
- Accuracy. An agent can repeat an outdated price or policy. Monitoring answers and correcting the source page is part of the job.
Frequently asked questions
What is agentic commerce in simple terms?
Agentic commerce is online shopping in which an AI agent acts for the shopper. It can search, compare, fill a cart and, within limits the shopper sets, pay. The shopper stays in charge of the goal and the limits.
Is agentic commerce the same as AI shopping?
No. AI shopping often means research and recommendations, where the shopper still buys. Agentic commerce adds action, with the agent initiating or completing the purchase.
Which protocol should a merchant support first?
Most merchants should start with their commerce platform's support rather than a custom build. Shopify co-developed UCP with Google, and ACP works with Stripe and other payment providers, so the right first step is to check what the existing platform and payment provider already support.
Can an AI agent buy without the shopper's approval?
Only within limits the shopper has set. Systems such as ACP, Visa Intelligent Commerce and Mastercard Agent Pay rely on scoped, tokenized credentials with spending controls, and most current experiences ask the shopper to confirm before an order is placed.
Will agentic commerce replace websites?
Nothing in the forecasts says so. Bain expects AI to play a role in most online shopping journeys by 2030 but to complete up to a quarter of transactions, and OpenAI's 2026 shift sent checkout back to merchant sites. Websites become one of several endpoints an agent can use.
How do you measure agentic commerce results?
Judge it by visibility, not clicks alone. Track how often products are named and cited across a fixed prompt list, how accurately they are described, referral sessions from AI assistants, and impressions and position in search. Clicks understate the effect when answers are given inside the assistant.
