Agentic commerce is shopping where an AI agent searches, compares and sometimes buys products for a person, within limits that person sets. "Agentic" means able to act toward a goal, not just answer a question.
McKinsey groups agentic commerce into three interaction models: agent to site, agent to agent and brokered agent to site. Agentic commerce is the shopping slice of a bigger shift, explained in our guide to what the agentic web is.
How Agentic Commerce Differs From Ecommerce Today
In traditional ecommerce, the shopper does every step. In agentic commerce, the agent does most of the steps and the shopper approves the result.
AI search assistance sits in between: a chatbot answers shopping questions and links to stores.
| Question | Traditional ecommerce | AI search assistance | Agentic commerce |
|---|---|---|---|
| Who searches? | The shopper, using a search box or store menus | The shopper asks, and the AI suggests products | The agent, based on the shopper's goal |
| Who compares? | The shopper, one tab at a time | The AI summarizes, and the shopper decides | The agent, using the shopper's rules |
| Who checks out? | The shopper, on the store's site | The shopper, after clicking a link | The agent or the shopper, after approval |
| What data does the system need? | Pages a person can read | Pages and reviews an AI can read and cite | Current, structured price, stock and shipping data |
| Who is the website built for? | Human visitors | Human visitors, with AI as a referrer | Human visitors and software agents |
McKinsey's three models describe how an agent reaches a store. In agent to site, the shopper's agent visits a store's website. In agent to agent, it talks to the merchant's own agent, and in brokered agent to site, a platform sits in the middle.
Why Agentic Commerce Matters Now
Shoppers already reach stores through AI tools. According to Adobe's 2025 holiday data, traffic to retail sites from generative AI tools (shoppers clicking on a link to a retail site) increased by 693.4% compared to the year prior.
The figure covers US retail sites in November and December 2025. It counts link clicks, not agent purchases, and Adobe notes the base of users remains modest.
By 2030, the US B2C retail market alone could see up to $1 trillion in orchestrated revenue from agentic commerce, according to McKinsey's agentic commerce research. Global projections reach as high as $3 trillion to $5 trillion.
"Orchestrated revenue" means spending that agents help steer. The estimate covers goods only, excludes services and B2B, and describes what could happen, not measured sales.
How Agentic Commerce Works, Step by Step
Every agent purchase follows the same data path, from the shopper's request to the delivered order. Each step needs different data, and each step can break.
Step 1: The Shopper Sets a Goal and Limits
The shopper tells the agent what they want in plain language. They also set limits: a budget, preferred brands, a delivery date and when the agent must ask first.
The agent turns that request into a structured task, which is a set of fields it can check. "A quiet blender that fits under my cabinets, arriving before the weekend" becomes a product type, a size limit and a delivery deadline.
Step 2: The Agent Discovers Products
The agent finds candidate products through web search, shopping results and merchant feeds. A merchant feed is a data file where a store lists its products, prices and stock for machines to read.
OpenAI says that through ACP, merchants share product feeds and promotions so their catalogs appear in ChatGPT. Feeds only cover merchants who opt in, so agents still search the open web for smaller stores and price comparison.
Feeds and web search work together. Builders can use Olostep's Google Shopping API for discovery and cross-merchant price comparison.
Step 3: The Agent Reads and Compares Product Pages
Next, the agent reads each product page for price, variants, stock and shipping. Variants are versions of one product, such as sizes or colors.
The agent fetches the page and renders JavaScript, the code that loads prices and buttons on many store pages. It also handles pagination, filters and "Load More" buttons.
The agent then extracts the facts into structured JSON, a format where each fact has a label like "price." Schema.org markup, often written as JSON-LD, helps by labeling name, price and availability inside the page.
Many product pages are still hard for AI to read. According to Adobe data reported by TechCrunch, around 34% of individual product pages can't be properly accessed by AI. The figure comes from Adobe's checker tool for US retail sites.
Olostep's real-time product page scraping fetches live pages, renders JavaScript and returns Markdown or structured JSON. With Olostep parsers, a team defines a schema once. The parser then returns consistent prices, variants and stock status across platforms like Shopify, WooCommerce and custom stores.
Step 4: The Agent Checks Out or Hands Off
Checkout happens in one of two ways today. The agent pays inside the chat through a protocol and a payment token, or it hands the shopper to the merchant's checkout.
A payment token is a stand-in for card details. Stripe's version is the Shared Payment Token (SPT), which lets an app like ChatGPT start a payment without exposing the buyer's payment credentials.
The token also has limits. Stripe's Instant Checkout announcement states that SPTs are scoped to a specific merchant and cart total. In plain words, the token works for one store and one amount.
A careful flow adds a human approval step before money moves. The shopper confirms the item, price and store, and the agent pays only after that.
Step 5: The Agent Tracks the Order and Keeps Data Fresh
After the purchase, the agent can track order status and help with returns. The harder job is keeping product data current.
Prices and stock change often. An agent working from yesterday's data can pick an offer that no longer exists.
Freshness comes from re-checking sources on a schedule. Data goes stale when it is collected once and never re-checked, not because it came from the web. Olostep's price and page change monitoring sends alerts when prices or pages change.
The Protocols Behind Agentic Commerce
A protocol is a shared set of rules that lets agents, merchants and payment providers work together. Without one, each pair would need its own custom integration.
Each protocol handles one job in the data path.
| Protocol | Backed by | Job it does | Status |
|---|---|---|---|
| Agentic Commerce Protocol (ACP) | OpenAI and Stripe | Checkout first, then product discovery feeds | Launched September 29, 2025 (Stripe); OpenAI expanded it to product discovery in March 2026 |
| Universal Commerce Protocol (UCP) | Covers the shopping journey from discovery to post-purchase support | Launched January 11, 2026 (Google); checkout "soon" for eligible US retailers as of that announcement | |
| Agent Payments Protocol (AP2) | Agent-led payments | Introduced in 2025, per Google's UCP announcement | |
| Model Context Protocol (MCP) | Open standard | Gives agents access to tools and data | Listed by Google as UCP-compatible in January 2026 |
| Agent2Agent (A2A) | Open standard | Lets agents talk to other agents | Listed by Google as UCP-compatible in January 2026 |
| Visa Intelligent Commerce | Visa | Card network program for agent payments | Early, controlled transactions |
| Mastercard Agent Pay | Mastercard | Card network agentic payments program | New trust services in US testing |
These protocols are designed to fit together. Google's UCP announcement says "UCP is built to work across verticals and is compatible with existing industry protocols like Agent2Agent (A2A), Agent Payments Protocol (AP2) and Model Context Protocol (MCP)."
Protocols cover participating merchants. The open web covers the rest, so agents still need search and page reading for stores without a feed.
MCP is the protocol that connects an agent to outside tools, such as web search or a scraper, through one standard interface. The Olostep MCP server gives MCP clients live web search and scraping.
Where Agentic Commerce Stands in 2026
Today, agentic commerce mostly means discovering products with AI and buying on the merchant's site. Fully autonomous checkout inside a chat is still early.
As Stripe's announcement described, Instant Checkout launched in September 2025, starting with US ChatGPT users buying from US-based Etsy sellers. In OpenAI's March 2026 update, the company said, "we're allowing merchants to use their own checkout experiences while we focus our efforts on product discovery."
OpenAI also said it is expanding ACP to support product discovery. Google described UCP checkout as coming "soon" for eligible US retailers in January 2026.
The working pattern is "discover with AI, buy on the merchant's site." In our view, that makes discovery and data quality the main area of competition right now.
Risks and Open Problems
Agentic commerce still has open problems, and most of them trace back to data, identity or responsibility.
- Bad data: An agent that reads an old price or stock level recommends the wrong offer.
- Bot confusion: Stores struggle to tell helpful shopping agents from harmful bots.
- Unclear responsibility: When an agent buys the wrong item, it can be unclear who should fix it.
Stale or Wrong Product Data
Wrong prices and out-of-stock items are a practical risk for shopping agents. Each mistake costs the agent the shopper's trust.
One cause is data collected once and not refreshed. The fix is to re-check prices and stock on a schedule and again right before any purchase.
Telling Good Agents From Bad Bots
Stores can't easily tell a shopping agent from a harmful bot, so some block automated traffic. That can also block agents acting for real customers.
According to the Thales 2026 Bad Bot Report, bots made up 53% of internet traffic in 2025. That number counts all automated traffic, good and bad, so it is not a count of shopping agents.
Card networks are building ways to identify trusted agents. Visa's December 2025 update reports hundreds of controlled, real-world agent-initiated transactions. Visa says its Trusted Agent Protocol is designed to help merchants tell malicious bots from legitimate AI agents. Mastercard's Agent Pay expansion adds a probability score, in US testing, that estimates whether an AI agent started a transaction.
Agents should identify themselves and respect robots directives, rate limits and site terms. Robots directives are the rules a site publishes, often in a robots.txt file, about what automated visitors may access.
Consent, Accountability and Trust
Responsibility can be unclear when an agent buys the wrong item. Several parties are involved: the shopper, the agent's maker, the merchant and the payment provider.
McKinsey warns that a single faulty prompt can lead to a purchase that occurs without consent. Spending limits, approval steps and scoped payment tokens are designed to limit that risk.
Privacy is the other open question. An agent needs sizes, addresses and payment details, so shoppers need to know who stores that data.
How to Prepare for Agentic Commerce
Preparation depends on your role. Merchants need agent-readable product data, and developers need a reliable way to find, read and refresh it.
For Merchants and Brands
Merchants should make products easy for agents to read and trust:
- Complete structured data: Add schema.org product markup with name, price, availability and shipping details.
- Accurate price and stock: Keep page data in sync with inventory so agents don't quote old offers.
- Product feeds: Share feeds through ACP, UCP or similar programs where available.
- Agent-view testing: Load your pages the way an agent would, and check which fields come through.
- Access decisions: Decide which agents to allow, and publish clear robots directives and site terms.
Merchkit works with companies like Walmart to help sellers optimize for agentic shopping. The Merchkit catalog enrichment case study reports 94% faster enrichment vs. manual workflows. It also reports 5x more SKUs optimized per month with the same team and a 10x cost reduction.
These are vendor-published results from one customer. A SKU is a single product listing, such as one size and color of a shirt.
For Developers Building Shopping Agents
A shopping agent needs a data layer that finds products, reads them into a fixed shape and keeps them current:
- Search for candidates through web search, shopping results and merchant feeds where they exist.
- Extract each product page to a fixed JSON schema, so every store returns the same fields.
- Process large catalogs in batches; Olostep's bulk catalog extraction handles thousands of URLs in one job.
- Re-check prices and stock on a schedule, and again right before checkout.
- Expose search and scraping to the agent as tools through MCP.
- Keep checkout on trusted rails, such as protocol checkouts or payment service providers (PSPs).
- Identify your agent, and respect robots directives, rate limits and site terms.
Olostep covers the data steps: search, scraping, batches, parsers and monitors. Checkout and payments stay with merchants, protocols and payment providers.
Frequently Asked Questions
What Does "Agentic" Mean?
"Agentic" describes AI that can act toward a goal, such as finding and comparing products, instead of only answering questions.
Is Agentic Commerce the Same as AI Shopping Assistants?
No. An AI shopping assistant suggests products, while an agent in agentic commerce compares options and sometimes buys within limits the shopper sets.
What Is the Agentic Commerce Protocol?
The Agentic Commerce Protocol (ACP) is an open standard OpenAI and Stripe launched in September 2025. OpenAI says merchants use it to share product feeds with ChatGPT.
What Is the Universal Commerce Protocol?
The Universal Commerce Protocol (UCP) is an open standard Google launched in January 2026 that covers shopping from discovery to post-purchase support. At launch, Google said UCP checkout would come "soon" for eligible US retailers.
Which Companies Are Using Agentic Commerce?
OpenAI, Stripe and Google have launched agentic commerce protocols, and Visa and Mastercard run agent payment programs. Etsy sellers were among the first merchants available through ChatGPT Instant Checkout in 2025.
Is Agentic Commerce Safe?
It can be safer when agents use scoped payment tokens, spending limits and human approval before buying. Fraud and bot detection are still developing, and card networks are piloting ways to identify trusted agents.
Why Might Agentic Commerce Fail?
Agentic commerce can fail when agents use stale product data, get blocked as bots or buy without clear consent. In-chat checkout is also still early, and OpenAI now lets merchants use their own checkout.
Where Do AI Shopping Agents Get Product Data?
AI shopping agents get product data from web search, shopping results, merchant feeds and the product pages they read directly. Feeds shared through protocols like ACP cover merchants who opt in, and the open web covers the rest.