The definition, and the split most definitions skip
Agentic commerce covers any purchase flow where an AI agent acts for the buyer. That flow has four stages, and they are not moving at the same speed. Discovery is the agent finding and comparing products. Selection is building a cart. Authorization is the buyer approving what the agent assembled. Settlement is money moving. Almost every explainer written in the last year treats these as one thing arriving at once. Discovery shipped, works, and is being measured. Authorization and settlement are still being argued over in standards bodies and, as of this month, in federal court. Sources: Yahoo Finance, reporting Adobe Analytics; Engadget.
An example makes the split visible. A shopper asks an assistant for running shoes under a certain price with a wide toe box. The assistant reads product data from several retailers, narrows to three, and explains the tradeoffs. That is agentic discovery, and it happens millions of times a day. Now the shopper says buy the second one. That instruction raises questions discovery never did: is this agent who it claims to be, did a human approve this exact cart at this exact price, who is liable in a dispute, and who is the merchant of record. Every 2026 protocol is an answer to those four questions. Sources: Google Cloud Blog; Visa Newsroom.
The evidence that discovery is working
Adobe Analytics, working from more than one trillion visits to United States retail sites, found AI-sourced traffic to those sites up 393 percent year over year in the first quarter of 2026, and up 693 percent during the November to December 2025 holiday period. The direction of travel matters less than the quality reversal underneath it. In March 2025 AI-referred visitors converted 38 percent worse than traffic from other sources. By March 2026 the same cohort converted 42 percent better, with revenue per visit 37 percent above non-AI traffic, 48 percent more time on page and 13 percent more pages browsed per visit. Sources: Yahoo Finance, reporting Adobe Analytics.
Visa cited growth of more than 4,700 percent in AI-driven traffic to United States retail sites over the prior year in an October 2025 release. Percentage growth off a small base is easy to overread, and neither number tells you what share of total retail traffic AI now represents. The conversion flip is the durable part: a visitor who arrives with a shortlist already built behaves differently from one who arrives mid-browse. Sources: Visa Newsroom; Yahoo Finance, reporting Adobe Analytics.
The checkout retreat nobody put on the explainer pages
On September 29, 2025, OpenAI and Stripe launched Instant Checkout in ChatGPT and open sourced the Agentic Commerce Protocol that powered it. The design was deliberately merchant-friendly. Merchants stayed merchant of record, kept their payment relationships, and could accept or decline each order, calculate and remit sales tax, and handle fulfilment and returns as normal. Stripe's Shared Payment Token let ChatGPT initiate a payment without ever seeing the buyer's credentials, scoped to one merchant and one cart total. It launched with United States Etsy sellers and a promise of over a million Shopify merchants. Sources: Stripe Newsroom.
It was withdrawn on March 4, 2026. By the time it ended, only about a dozen of Shopify's merchants had actually gone live, and the launch feature set had never grown past single-item purchases. OpenAI's replacement plan routes transactions through third-party apps inside ChatGPT rather than completing them in the chat itself. The protocol did not die with the product. The ACP specification kept shipping releases through the withdrawal, adding fulfilment in December 2025, capability negotiation in January 2026, and cart, feed, orders, authentication and MCP work in the April 17, 2026 release, all under Apache 2.0. Sources: Forbes, Jason Goldberg; OpenAI and Stripe, GitHub; Agentic Commerce Protocol, Stripe and OpenAI.
The lesson is not that agentic checkout failed. It is that the hard part was never the payment primitive. A five-month experiment with a working token standard and a billion-user surface still could not beat the retailer's own checkout page, because shoppers who had just used an agent to narrow the field were content to finish where their saved address, loyalty points and returns history already lived. Convenience at discovery is a real gain. Convenience at checkout was competing against something already convenient. Sources: Forbes, Jason Goldberg.
Three protocols, and what each one actually governs
The common mistake is to treat ACP, UCP and AP2 as competitors picking a winner. They govern different layers and are designed to compose. ACP, from OpenAI and Stripe, describes checkout and delegated payment between an agent and a merchant. UCP, announced by Google with Shopify on January 11, 2026, describes the whole commerce conversation: capability discovery through a JSON manifest at a well-known path, checkout, and post-purchase data, over REST, MCP or A2A transports. AP2 describes only payment authorization: how a merchant proves a human approved this cart at this price. Google's documentation states UCP is compatible with AP2 for payments and plugs into MCP and A2A. Sources: Google Developers Blog; OpenAI and Stripe, GitHub; Google Cloud Blog.
AP2 is the one that changed most. At launch on September 16, 2025 it described three mandates: an Intent Mandate capturing the request, a Cart Mandate creating what Google called a secure, unchangeable record of the exact items and price, and a Delegated Task Mandate setting price limits for autonomous purchases, all signed with verifiable credentials. By v0.2 the model had been reworked into Checkout Mandates and Payment Mandates, each with an open form setting constraints ahead of time and a closed form authorizing a specific transaction, with explicit human-present and human-not-present flows. Renaming the core object of a payments protocol seven months after a 60-organization launch tells you the working group is still finding the shape. Sources: Google Cloud Blog; AP2 Protocol.
| ACP | UCP | AP2 | |
|---|---|---|---|
| Origin | OpenAI and Stripe | Google and Shopify | |
| Announced | 2025-09-29 | 2026-01-11 | 2025-09-16 |
| Governs | Agent to merchant checkout and delegated payment | Discovery, capability negotiation, checkout, post-purchase | Payment authorization and proof of human intent |
| Core mechanism | Shared Payment Token scoped to one merchant and cart | JSON capability manifest at a well-known path | Cryptographically signed mandates using verifiable credentials |
| Current state | Beta, spec release 2026-04-17, Apache 2.0 | Spec at ucp.dev, wired into Google AI Mode and Gemini | v0.2, donated to the FIDO Alliance 2026-04-28 |
The real bottleneck is agent identity, not payment
Before a merchant can decide whether to sell to an agent, it has to know an agent is there. That turned out to be the load-bearing problem, and the answer converged faster than the payment protocols did. Cloudflare's Web Bot Auth uses RFC 9421 HTTP Message Signatures with Ed25519 keys, so an agent signs its requests and the site verifies against a public key in a well-known directory. It replaces user-agent strings and IP ranges, both of which are trivially spoofable. Visa's Trusted Agent Protocol, published October 14, 2025 with Cloudflare and a dozen early partners including Adyen, Checkout.com, Microsoft, Shopify and Stripe, builds on it and carries three things to the merchant: agent intent, consumer recognition, and payment information. Mastercard's Agent Pay uses the same authentication layer. Sources: The Cloudflare Blog; Visa Newsroom.
This is why the card networks, not the AI labs, are setting the pace. A merchant can now run a default-deny posture against automated traffic and still auto-allow agents registered with Visa or Mastercard, which is the only version a risk team will approve. Visa reported in December 2025 that hundreds of agent-initiated transactions had completed across pilots with over 100 partners globally, and that 47 percent of United States shoppers already used AI tools for shopping tasks. Hundreds of transactions is a pilot, not a market. Sources: The Cloudflare Blog; Visa Newsroom.
Governance moved too. On April 28, 2026 Google donated AP2 to the FIDO Alliance, saying the transition keeps the protocol platform-agnostic and community-led. FIDO opened an Agentic Authentication Technical Working Group chaired by CVS Health, Google and OpenAI, vice-chaired by Amazon, Google and Okta, with agent-initiated commerce in its Payments group chaired by Mastercard and Visa. When the people who run passkeys take over agent authorization, the industry has decided this is an authentication question wearing a commerce costume. Sources: Google, Google Pay blog; FIDO Alliance, via Business Wire.
The unsettled question: can a retailer say no?
Every protocol above assumes cooperation. Amazon and Perplexity spent 2026 testing what happens without it. Amazon sued over Comet, Perplexity's browser agent, which logs into a user's Amazon account and buys on their behalf. Amazon's complaint said Perplexity had deliberately disguised the agent as a regular Chrome session, evading detection rather than transparently identifying itself. On March 10, 2026, Senior United States District Judge Maxine Chesney granted a preliminary injunction, finding Comet accessed accounts with the Amazon user's permission but without authorization by Amazon. Sources: GeekWire.
On August 4, 2026 the Ninth Circuit vacated it. The panel held that when a user directs the assistant to complete a task on Amazon.com, it is the user and not Perplexity accessing Amazon's computers, so Amazon was unlikely to succeed on its Computer Fraud and Abuse Act claim, and reached the same conclusion on the parallel California statute. The opinion noted that an injunction against conduct that likely does not violate those statutes would not serve the public interest. The case was remanded and continues, and Amazon can seek rehearing or petition the Supreme Court, so nothing here is final. Sources: Engadget; eMarketer.
The practical read is narrow but real. Computer-fraud statutes look like a weaker tool for excluding user-directed agents than retailers assumed, which shifts leverage toward technical controls and away from terms of service. That is the gap cryptographic agent identity fills. A merchant that can verify which agent is knocking gets a business decision. One relying on a terms-of-service clause is relying on the part of the stack the courts just made shakier. Sources: eMarketer; The Cloudflare Blog.
What our own signal log has been showing
Two threads in the Daily AI Roundup's public signal log this month sit on this seam. On August 6 we tracked Google adding agentic food ordering and hotel booking to Maps, turning a navigation product into one that completes real-world transactions, filed under transaction agents rearranging who owns customer intent. On August 8 we tracked Cloudflare shipping Kitesurf, a cloud browser built for agents rather than people, filed as browser agents needing passports rather than better browsers.
Those are the same story at different layers. Transaction surfaces are migrating into assistants that already hold the user's intent, which is why Maps booking a table matters more than another shopping chatbot. The constraint on that migration is not payment plumbing, which three protocols now cover, but whether the site on the other end can tell a legitimate agent from a scraper. The passport comes before the wallet, and this year's protocol race has been the industry arriving at that ordering the expensive way. Sources: The Cloudflare Blog; Google Developers Blog.
What to do now, in the order that pays
First, stop blocking the agents you want. Many retailers still block major AI crawlers at the infrastructure layer while separately funding projects to capture AI-referred demand. Audit what your edge allows before buying anything. Second, make product data readable without JavaScript execution. An agent that cannot parse price, availability and variants from your product page will recommend a competitor whose page it could read. Sources: Yahoo Finance, reporting Adobe Analytics; The Cloudflare Blog.
Third, treat protocol adoption as sequenced. If you sell on Shopify you are already closer to UCP than you think, and if you use Stripe you are closer to ACP. Publish capabilities where the specification says they go rather than building a bespoke agent API nothing will speak next year. Fourth, put agent identity on the security roadmap now: default-deny plus a verified allowlist is where both card networks are heading, and it needs a risk owner before it needs a budget. Fifth, measure AI-referred traffic as its own channel, or you cannot tell whether any of the above worked. Sources: Google Developers Blog; OpenAI and Stripe, GitHub; Visa Newsroom.
Common questions
Can you give an example of agentic commerce? The clearest live examples are Google's AI Mode in Search and the Gemini apps, where United States shoppers can buy from participating retailers through UCP without leaving the conversation, and Google Maps completing food orders and hotel bookings. The best-known example, Instant Checkout in ChatGPT, is no longer live: it ran from September 2025 to March 2026. Sources: Google Developers Blog; Forbes, Jason Goldberg.
What does agentic mean in business, in simple words? Agentic means the software decides and acts on a goal rather than waiting for each instruction. A recommendation engine suggests; an agent goes and does. In commerce the practical test is whether the system can take an action with a financial consequence without a human clicking the button for it. Sources: Google Cloud Blog.
What is agentic commerce optimization? It is making a store legible and trustworthy to machines: structured product data an agent can read without rendering JavaScript, published capabilities where a protocol expects to find them, and an edge policy that verifies and admits legitimate agents instead of blocking all automation. It overlaps with classic technical SEO but the consumer is a program with a signing key, not a person with a browser. Sources: The Cloudflare Blog; Google Developers Blog.
Can retailers block AI shopping agents? Technically yes, and cryptographic agent identity makes selective blocking practical for the first time. Legally it is less settled than it looked in March 2026: the Ninth Circuit's August 4 decision suggests computer-fraud statutes are a weak tool against agents a user directed, though that case is still live on remand. Sources: Engadget; eMarketer.
Is agentic commerce the future? Agent-assisted discovery already is, on the numbers. Fully autonomous checkout inside an AI surface is an open question with one significant failure already on the record, and the honest position is that nobody has demonstrated durable consumer demand for it yet. The infrastructure is being built anyway, which is a reasonable bet but still a bet. Sources: Yahoo Finance, reporting Adobe Analytics; Forbes, Jason Goldberg.
Limits and uncertainty
This guide describes agentic commerce as of August 17, 2026, in a field where a core protocol renamed its central object inside seven months. Three limits are worth stating. The Adobe Analytics figures are cited from a syndicated report of Adobe's data, because Adobe's own pages did not respond on the day of publication; read them as Adobe's methodology over its own panel, not an independent measurement. Widely circulated 2030 forecasts putting a fixed percentage of commerce in agent hands are deliberately absent, because we could not verify any of them at the issuing analyst firm's own page today, and an unverified forecast is worse than none. Finally, the Amazon and Perplexity litigation is live and was remanded, so the holding described here concerns likelihood of success at a preliminary stage, not a final ruling on whether retailers may exclude AI agents.