For twenty years, digital marketing assumed the customer eventually showed up. They might come from Google, from a friend, from an ad they half-remember, but somewhere along the way they landed on a page the merchant controlled. The merchant got to do something with that moment: improve the photography, cut a step out of checkout, test a new headline, rewrite the offer, build a better cart. The whole industry was built around making that moment work harder.
That assumption is starting to wobble.
Imagine the near version of this. A user asks an agent to find running shoes for wide feet, or to reorder the dog food that worked last time. The agent searches, compares, filters, and may eventually complete more of the purchase flow itself. The user may never visit the merchant's site, may not see the category, and may never know which options were considered and discarded along the way.
The rails are going in
This is still a small share of online purchases. But early usage is visible, and the rails for it are being built faster than most merchants are watching. Stripe is building agentic commerce infrastructure that lets merchants connect product catalogs, choose which agents they sell through, and use Stripe for discovery, checkout, payments, and fraud detection. Google has contributed its Agent Payments Protocol to the FIDO Alliance as part of a broader push toward standards for agent-led payments. OpenAI launched Instant Checkout in ChatGPT with Stripe in September 2025 and pulled it back in March 2026, shifting its focus toward product discovery while merchants handle their own checkout. None of this is settled, and the rails are still changing, but they are going in.
Most of the writing about what this means for merchants sits at one of two ends. On one end is readiness: can agents read your catalog, trust your inventory, call your checkout, and complete a transaction? Real work, but mostly plumbing. On the other end is conversion: does AI-referred traffic actually buy? Useful question, but asked too late: by the time you are measuring conversion, the agent has already decided whether you were worth considering.
The missing layer
That decision is the missing layer. It is not SEO. SEO ranked pages for people who would then choose for themselves. If agents absorb more of the choosing function, it will likely start in narrower categories: books, household repurchases, simple commodity buys. The deeper that pattern goes, the more the things agents weigh diverge from what merchants spent twenty years optimizing.
A human can be moved by a good photograph, by social proof, by the feeling that a brand understands them. An agent may pick up some of that through reputation, reviews, prior user behavior, and training data, but much of what it weighs is more boring: whether your attribute fields are filled in, whether your inventory data is fresh, whether your prices match what you actually charge at checkout, whether your returns policy is documented somewhere it can read, whether your checkout can be called without breaking the flow.
The old question was whether you could be found. The harder question is whether you can be selected.
There is no clean number of stages inside that, but it helps to pull the problem apart.
Memory is the underestimated layer. Once an agent knows what worked for this user last time, the next purchase does not begin as a search; it begins as a default: the size that fit, the vendor that delivered on time, the support team that resolved the issue without an argument, the brand the user never returned.
The merchant used to own that history through email lists, loyalty programs, and CRM. As more buying moves through agents, more of that history will move upward into the agent, the wallet, the browser, or a personal context layer that travels with the user. The merchant becomes one variable in a model the merchant may never get to see.
The new tax
This is where the new tax starts to take shape. The old internet taxed attention; Google found a more refined version of that and called it intent; Meta added identity to attention; Amazon taxed proximity to purchase. Each turn of the internet had its own tax, collected near the moment a user's attention narrowed into a decision.
As agents take on more of the narrowing, that decision shifts into a system the user does not watch. Whoever owns that system can collect rent on consideration, on the right to be inside the set before the user sees anything.
In its early forms, the tax probably will not look like an ad product. It may appear as preferred payment rails, verified merchant programs, certified checkout integrations, fulfillment-based ranking signals, and defaults shaped by past behavior. The public vocabulary will be safety, reliability, and user experience; the economic effect will be rent.
Who audits the agent?
Then comes the harder question, the one that has not fully landed yet: who audits the agent?
If the same system selects the options, explains the choice, and monetizes access to consideration, the conflict is built into the architecture. The user sees the recommendation, not the discarded alternatives. They see the answer, not the shape of the choice set.
This is where personal context becomes more than memory. A user-owned context layer could become the reference model against which an agent's choices are judged, not because it knows every product better than the agent, but because it knows the user outside the agent's commercial surface: prior fit, constraints, routines, preferences, friction, returns, timing, and what has actually worked before.
The audit question is simple: did the agent choose this because it fits the user, or because it fits the platform?
If the past two decades of consumer-internet platforms tell us anything, every agent platform will say it picks what is best for the user, and every agent platform will also need to monetize the architecture of that picking. Search had a version of this problem, but a search user could at least see the page and tell, more or less, which results were paid for. With agents, the decision is delegated, and paid influence inside delegated choice is harder to see and harder to opt out of.
If this becomes a major channel, paid influence will likely appear there too, only less visibly than it did on a results page.
What this means for merchants
For merchants, the practical change is not that marketing disappears. It moves into stranger territory. Structured data becomes distribution. APIs become demand. Fulfillment quality starts affecting whether you appear at all. Memory becomes the new loyalty layer, and much of it will live outside your CRM. The work shifts from persuading a person who has arrived to becoming the option an agent can justify picking.
The window to shape that architecture is open now, while the share is still small and the rails are still being negotiated. It will not stay open long.
That is the shape of the next advertising market. The unit being sold will not be attention; it will be consideration.