Two announcements this year, taken alone, look like routine product news. Google's AI Mode turns search from a list of blue links into a conversational decision surface that can reason, compare, and increasingly act. Rakuten Advertising's Mirai does the same to affiliate marketing, putting an always-on optimization agent between advertisers and offers, tuning creative and placement in real time.

It would be easy to file both under "more AI in marketing tools" and move on. That would be the wrong read. What is actually happening is narrower and more consequential: two of the most important performance marketing surfaces in the world are quietly becoming agent runtimes. The thing your customer interacts with is no longer a results page or an offer wall. It is a model deciding what to surface, what to recommend, and increasingly what to execute on their behalf.

That shift breaks the link click as the unit of performance marketing. And it forces a new question that most brands have not yet asked clearly: what are you publishing for the agent to select?

The artifact is changing

For twenty years, performance marketing has optimized one artifact: the link. A keyword, an ad, a landing page, a click, an attribution chain. Every metric, every team structure, every measurement vendor assumes that humans navigate a path of clicks and that you can measure intent by what they tap.

AI Mode does not work that way. When a shopper asks Google for the best running shoes for a flat-footed half-marathoner under two hundred dollars, the model is doing the comparison work that used to happen across a dozen tabs. The result is a recommendation, sometimes a shortlist, sometimes a direct action. The session may never produce the click that your dashboard expects to see.

Mirai pushes the same logic into affiliate. Instead of a publisher choosing between offers and a human deciding what to click, an agent evaluates offer mechanics, audience fit, and real-time performance, then routes accordingly. The advertiser is no longer optimizing for a static placement. They are optimizing to be selectable by a system that is itself optimizing.

The Yahoo Finance commentary on agentic commerce captured the consequence plainly: when a buyer delegates discovery to ChatGPT, Gemini, Claude, or Perplexity, the agent decides which products surface and which never get seen at all. There is no sponsored listing in the old sense. There is no rank. There is a decision, made inside the model, before any human attention is spent.

This is what becomes newly possible, and newly required: publishing for agents, not for clicks.

What an agent-consumable action actually is

If the agent is the new buyer's representative, brands need to produce something the agent can read, evaluate, and execute against. That artifact deserves a name. Call it an agent-consumable action.

An agent-consumable action has three parts.

First, a schema. A structured description of what the offer is, who it is for, what the price and terms are, what the eligibility constraints are, and what counts as a successful completion. This is not SEO metadata. It is the contract an agent reads to decide whether to recommend, route, or execute.

Second, an acceptance gate. A clear rule for which agents can act on the action, under what conditions, with what authentication, and with what reversibility. The gate is what separates an action you have published from an action you have actually authorized.

Third, instrumentation. Telemetry that captures the things the click model never captured: when an agent considered the action, when it shortlisted it, when it confirmed it with the user, when a state change occurred. The point is not to recreate last-click attribution inside the agent. It is to measure the new behavior the agent surface actually produces.

A brand that publishes well-formed agent-consumable actions can be discovered, evaluated, selected, and transacted with inside AI Mode, inside Mirai, and inside whatever surfaces follow. A brand that publishes only pages and ads will be summarized by someone else's model, on terms it does not control.

The deeper read

This is not the first time a distribution interface has restructured marketing. Mobile forced apps and responsive design. App stores forced ratings, ASO, and a new economics of distribution. Social feeds forced creative built for the scroll and measurement built for the platform. Each shift required brands to rewrite their assets and their measurement around a new selection layer.

The selection layer is now a model. The asset is now an action contract. The measurement is now agent behavior.

What makes this transition harder is that the previous shifts at least preserved the click as a primitive. You could still count taps and visits and conversions, even if the surface changed. Agent surfaces erode the primitive itself. By the time a recommendation reaches the user, the comparative work has already happened inside a model whose telemetry you do not own.

Adobe's expansion of AI agents inside Experience Cloud hints at the enterprise response: orchestration platforms that expect structured agent contracts on both ends, with Journey Optimizer identifying opportunities and agents executing against them. The CMS becomes a contract publisher. The journey tool becomes a selection optimizer. But that operating model only works if the brand has agent-consumable actions to publish in the first place.

The measurement objection

The predictable response from senior marketing leaders is that this is premature. Attribution is not solved. The standards are not settled. CFOs want clean numbers before reallocating budget. Wait, they say, until the measurement catches up.

The interface is not waiting. AI Mode is already routing queries. Mirai is already optimizing offers. Every month spent waiting for an industry-standard agent attribution model is a month spent optimizing for a surface that the platforms themselves are deprecating in favor of the new one.

The more honest read is that measurement lag here is integration debt, not a reason for delay. The same pattern showed up in the FTC's recent enforcement on AI claims, where the bottleneck was not whether to deploy but how to produce credible proof that a live claim matched current capability. I wrote about that as verification latency: the time required to make AI performance auditable in the same window in which it operates. Agent-consumable actions are the same problem at the front of the funnel. You do not solve it by waiting. You solve it by building the instrumentation alongside the artifact.

This is where the discipline of assent matters, and it is the only place this piece earns the word. Do not give assent to metrics you cannot defend, and do not refuse assent to a surface that is already routing your customers. Build for the first agent action you can measure credibly. Stage the rest. Treat each addition as a small commitment that has to earn its place in the proof loop.

A staged migration, not a moonshot

The practical move is narrower than most agentic AI roadmaps suggest. A brand does not need to rebuild its entire stack to participate. It needs a pipeline.

Start with an audit of existing offers and journeys. Identify three or four where the customer's intent is clear, the action is well-bounded, and a state change is observable. A subscription renewal. A trial activation. A configured product order. These are the candidates for the first agent-consumable actions.

Then draft the schema. Describe each action in structured form: what it is, who qualifies, what the agent must confirm, what the success condition is. Define the acceptance gate. Decide which surfaces you authorize, with what authentication, and what fallback if the agent's confidence is low.

Instrument the surface around it. Capture consideration, shortlist inclusion, user confirmation, and completion as distinct events. These become the early replacement metrics for the click funnel: time-to-shortlist, selection-fit, confirmation rate, completion-to-state-change.

Then pilot. Make a small number of agent-consumable actions live on the surfaces that already support agentic selection, watch the telemetry, and let the proof loop tighten before you expand the catalog. The portable, runtime-level shift I described in When Agent Deployment Becomes a ZIP File applies here. The scarce resource is not the agent. It is evidence of correct outcomes in your environment.

The brands that move first will not be the ones with the most sophisticated AI stack. They will be the ones that committed early to a small, measurable set of agent-consumable actions and learned how to instrument them while their competitors waited for an attribution standard that arrives after the surface has already reshaped the category.

The closing turn

The interface has moved. Search is becoming a decision runtime. Affiliate is becoming an optimization runtime. The performance marketing surface that produced the modern measurement stack is being quietly replaced by surfaces that select rather than display.

The right response is not panic and not a grand strategy memo. It is to publish the first agent-consumable action your team can defend, instrument it honestly, watch what the agent does with it, and expand from there. The artifact is the work. The proof is the discipline. The interface will keep moving. The brands that build for it, one measurable action at a time, will find their metrics arriving where their customers already are.

Sources