In the second half of 2025, four of the largest companies in technology and payments published, within a few weeks of each other, the plumbing for a future where an AI agent does your shopping. OpenAI and Stripe shipped one standard. Google shipped another. Visa and Mastercard each shipped a third and fourth. The announcements were dense, technical, and aimed at merchants and developers. Almost none of them used the word affiliate.
That silence is the reason this post exists. Affiliate marketing is a 14 billion dollar US channel, by eMarketer's 2026 estimate, and it runs on a single mechanical fact: a person clicks a tracked link, a cookie or click ID is set, and a sale within the attribution window pays a commission to whoever owned that link. Agentic commerce protocols were designed to move a purchase from a human clicking through a website to an agent calling an API. When that happens, the click does not happen. And almost everything affiliate depends on hangs off the click.
This is Part 1 of a two-part series. Here we explain what the protocols actually are, who owns them, and the specific mechanical reason they matter for affiliate. Part 2 takes the harder question head on: whether a tracked affiliate link can survive a purchase an agent completes without ever loading your page.
Origin: why payments needed a new protocol at all
For three decades, online checkout assumed a human at a browser. You see a page, you read a price, you type a card number, you click buy. Every fraud control, every tracking pixel, every affiliate cookie sits inside that loop and quietly assumes a person is in it.
An AI agent breaks the loop. If an assistant is going to compare three running shoes and buy one for you, it needs to do something a merchant has spent twenty years trying to stop: submit a payment without a human visibly present at checkout. That raises three problems at once. The merchant needs to know the agent is legitimate and not a bot draining inventory. The card network needs to know the charge was actually authorized by the cardholder. And the payment credential cannot simply be handed to a third-party AI. Solve those and you have agentic commerce. Fail and you have a fraud channel.
The protocols announced in late 2025 are four different answers to that problem. They are worth knowing by name, because the differences between them decide where, if anywhere, a publisher fits.
Present: the four protocols, by name and owner
The Agentic Commerce Protocol, or ACP. Co-developed by OpenAI and Stripe, announced on 29 September 2025 alongside Instant Checkout in ChatGPT. ACP is an open standard, published on GitHub under the Apache 2.0 license and maintained jointly by the two companies. Its job is to give an AI agent and a merchant a shared language: the agent reads a structured product feed, builds a cart, and submits an order, while the merchant keeps control of what is sold, how the brand appears, and fulfillment. The clever piece is the Shared Payment Token, a Stripe primitive that lets an app like ChatGPT trigger a charge without ever seeing the buyer's card. The token is scoped to one merchant and one cart total, so it cannot be reused or inflated. Crucially, ACP is not Stripe-only: a merchant on a different payment processor can still adopt it.
The Agent Payments Protocol, or AP2. Announced by Google on 16 September 2025, with more than 60 organizations signed on at launch, including Mastercard, American Express, PayPal, Adyen, Worldpay, Coinbase, and Salesforce. AP2 is not tied to one AI assistant. It extends Google's Agent2Agent protocol and works alongside the Model Context Protocol, and it is deliberately payment-agnostic: cards, real-time bank transfers, and stablecoins all fit. Its central idea is the Mandate, a cryptographically signed digital contract that proves what the user actually authorized. An Intent Mandate captures the request ("find me white running shoes under 120 dollars"). A Cart Mandate, signed once the agent presents a specific cart, locks the exact items and price. The chain from intent to cart to payment creates what Google calls a non-repudiable audit trail. In April 2026, Google donated AP2 to the FIDO Alliance, putting it under neutral standards governance rather than keeping it in-house.
Visa's Trusted Agent Protocol and Mastercard's Agent Pay. The card networks moved in October 2025, both within the same week. Visa's Trusted Agent Protocol is an open framework that lets a merchant verify an agent is trusted before checkout, signal the agent's intent, and recognize the real consumer behind it; its launch partners include Microsoft, Shopify, Stripe, and Worldpay. Mastercard's Agent Pay, first introduced in April 2025 and extended with a merchant acceptance framework in October, registers and authenticates agents before any transaction and leans on tokenized credentials and payment passkeys. Both build on Cloudflare's Web Bot Auth technology. Their focus is narrower than ACP or AP2: not the whole shopping flow, but the trust and identity layer underneath the payment.
Two patterns matter here. First, these are not really competitors in the way headlines imply. ACP governs the conversation between agent and merchant; AP2 governs the authorization and the payment instruction; the card-network protocols govern agent identity and trust. A real agentic purchase could touch all three. Second, look at who is absent from every list. There is no affiliate network among the launch partners. No impact.com, no Awin, no CJ. The protocols were specified by AI labs, payment processors, and card networks. The affiliate channel was not in the room.
The mechanical reason this matters for affiliate
Strip away the branding and an affiliate program is one rule: pay a commission to whoever the tracking says influenced the sale. The tracking, almost always, is a click. A reader clicks your link, the network sets a cookie or records a click ID, and any purchase inside the attribution window, typically 7 to 30 days, pays you.
An agent-mediated purchase, by design, never fires that mechanism. Walk it through. A shopper asks an assistant to find the best cordless vacuum under 300 dollars. The agent reads sources, picks one, builds a cart through ACP, and pays with a Shared Payment Token or an AP2 Cart Mandate. At no point does a browser load a merchant page carrying an affiliate parameter. There is no cookie drop, no click ID, no referral header from a publisher. The infrastructure affiliate has depended on for thirty years simply does not run. As one agentic commerce engineering write-up puts it bluntly, the entire attribution infrastructure affiliate marketing depends on does not fire.
Now add the second blow, the one that makes this acute rather than hypothetical. The agent still had to learn about that vacuum somewhere. It read reviews, comparisons, and roundups. And the open web's product-review content is overwhelmingly affiliate content. The UK's affiliate trade body, the APMA, cites CJ research finding that affiliate publishers account for up to 91 percent of the sources ChatGPT cites when answering questions about major brands' flagship products. The same body notes that when AI assistants ingest that content, they do not import the affiliate links embedded in it.
Put those two facts side by side and you have the defining problem of agentic affiliate. Affiliate content is doing more of the work than ever, because it is what the agents read. And affiliate links are capturing less of the credit than ever, because the agent strips the link and the click never happens. The publisher does the influencing and gets none of the attribution. That is not a tracking bug. It is the protocols working exactly as specified.
Be precise about what the protocols do and do not address. ACP, AP2, and the card-network frameworks all solve merchant trust, payment security, and authorization. None of them contains a field for an affiliate ID, a referrer, or a commission split. They are silent on attribution beyond the buyer and the merchant. That silence is not hostility; affiliate simply was not a design goal. But silence in a protocol is consequential: with no defined place to carry a publisher's identity through the transaction, the default outcome is that the identity is lost.
Future and impact: who captures the value the click used to
The protocols are young and the reality is messier than the launch posts. OpenAI's Instant Checkout is the clearest example. It launched in September 2025, and by March 2026 OpenAI pulled back to a discovery-first model: ChatGPT handles research and recommendation, the merchant keeps the checkout. Six months after launch only around 30 Shopify merchants were live, and OpenAI found users were happy to research in chat but preferred to buy in a familiar store. The agent-completes-the-purchase future is coming slower than the announcements suggested.
But the direction holds, and the money is already being priced. Where a purchase does complete through ChatGPT, OpenAI takes a fee from the merchant. Sam Altman initially described it as roughly 2 percent; reporting on the Shopify integration puts the figure at 4 percent of the sale, on top of Shopify's own fees. Read that number carefully, because it is the platform's answer to the same question affiliate is asking. The 4 percent is what the platform charges for delivering a buyer. It is, functionally, a commission. The agent platform has positioned itself in the slot the affiliate used to occupy: the party that gets paid for sending a purchase to a merchant. The difference is that the platform owns the surface, the conversation, and the checkout integration, and the affiliate owns a link the agent does not carry.
So the contested question for the next few years is not whether agentic commerce happens. It is who is recognized as the influencer when it does. Three rough scenarios are visible.
In the first, platforms keep the value. The agent reads affiliate content, recommends a product, completes the purchase, and the platform collects its fee. The publisher who wrote the review that swayed the agent gets nothing. This is the current default, and it is why affiliate trade press has spent 2026 sounding alarms.
In the second, a new attribution layer emerges that pays on influence rather than on the last click. If credit can shift from who set the cookie to who informed the decision, the most accurate and trusted product content becomes the most valuable input to an agent, and its authors could be paid for that. There are early gestures in this direction. Perplexity's publisher program shares subscription revenue with cited sources; some practitioners argue unique, structured voucher codes can survive an agent's summary where a tracked link cannot. None of this is a settled standard yet.
In the third, affiliate repositions around what agents cannot easily disintermediate. The advice from affiliate strategists in 2026 is consistent on this point: an mThink survival guide argues that community beats traffic, that owned channels like newsletters and members-only content buffer a publisher from zero-click volatility, and that generic comparison tables are dead weight when an agent generates one in milliseconds. First-hand testing, original data, and audiences who return for a specific voice are harder for an agent to flatten into a citation.
The most likely outcome is some blend, decided largely by whoever moves to define the missing piece. The protocols today have no affiliate field. Whether one gets added, whether affiliate networks integrate as recognized agents under Visa's or Mastercard's identity frameworks, or whether a separate attribution standard appears, is genuinely open. What is not open is the underlying shift: the click is no longer the reliable unit of credit, and any affiliate strategy still built entirely on it is building on ground that is moving.
For marketers running affiliate programs, the near-term work is unglamorous and worth doing now. Make product content genuinely useful and machine-readable, because that is what agents cite. Treat structured, unique discount codes as a tracking mechanism that can survive a summary. Watch the protocol specifications as they evolve, ACP and AP2 are public, and a future revision could add the field that changes everything. And read attribution reports knowing that agent-influenced sales are already arriving uncredited in the direct channel.
Part 2 follows a single tracked affiliate link into an agent-mediated purchase and asks, concretely, whether it survives the journey, where exactly it breaks, and what a publisher can realistically do about it.
Council summary
This post argues that the agentic commerce protocols, ACP, AP2, and the Visa and Mastercard trust frameworks, all solve merchant trust and payment security while leaving no field for an affiliate ID, so the click that affiliate attribution depends on stops firing. The council verified every protocol name, owner, and launch date against primary sources: ACP from OpenAI and Stripe on 29 September 2025, AP2 from Google on 16 September 2025, and the card-network protocols in October 2025. We corrected the AP2 donation to the FIDO Alliance, which the draft placed in November 2025 but which happened in April 2026, and tightened the CJ 91 percent citation to its specific claim about flagship products; the 4 percent ChatGPT fee and the discovery-first pivot were both confirmed in reporting. The takeaway is concrete: the click is no longer a reliable unit of credit, so make product content machine-readable, treat unique discount codes as tracking that survives a summary, and watch the public specs for the attribution field that does not yet exist.
Comments