Your next customer is an agent

A few months from now, a sale will close on your site and no person will have seen it. Someone told their assistant what they needed. The assistant read your page, checked your terms, weighed you against two competitors, and paid. Your hero image, your testimonial carousel, the pop-up you timed so carefully: the buyer loaded none of it.

We have spent twenty years writing for a person reading a screen. Search rewarded the pages people clicked. Then answer engines arrived and the goal shifted to being the source a model quotes back. That work has a name now, answer engine optimization, and it still matters. But it assumes the agent is a reader. The agent is turning into a buyer.

What actually changed

A customer needs three things, and agents now carry all three.

An identity, so there is a “who” behind the request and a record of whom they act for. A wallet: the new agent-payment rails let an assistant spend inside a bounded mandate rather than holding a person’s card. OpenAI and Stripe shipped agentic checkout; Google published an Agent Payments Protocol; Visa and Mastercard each named an agent-pay program. And an inbox: agents increasingly have their own address, send and receive mail, and act on what lands there.

Put those together and you get a customer who can find you, judge you, and pay you, with no person in the loop for the routine parts. That is not a future memo. The protocols are shipping this year.

Marketing to an agent is a different craft

An agent does not feel your brand. It does not linger on a good photo. It reads structure. It wants to know, in terms a machine can check: what you do, what it costs, whether it can act through you without getting its person burned, and whether your claims survive a look. You do not win it with adjectives. You win it by being verifiable.

The old funnel was attention. The new one is trust you can check. There is a name forming for the work, agent experience optimization, the step after SEO and AEO. It is less about copy and more about plumbing. The honest version: the best marketing you can do for an agent is a clean interface and a claim that holds.

The four moves

Discoverable. The agent has to find the machine-readable you. A llms.txt at your root that states your products, your pricing, and how to act. Structured data so an agent can match a real query to your real offer. And, more and more, a surface it can call rather than only a page it can read: an MCP server is becoming the front door agents knock on.

Usable. Once found, can it act? Documentation an agent can parse. Pricing it can read without scraping a PDF. Errors that explain themselves. An interface that behaves the same way twice.

Buyable. Can it pay without risk? Short-lived, scoped authority, so an agent spends inside a mandate and not a cent past it. Terms a machine can read. A clear record of who acted, and on whose behalf.

Trustworthy. Agents are unsentimental. They notice latency, downtime, and a promise that does not survive a check. The cheapest marketing you will ever do is being honestly fast and honestly available.

The line most advice skips

Agent-ready is not scrape-ready.

An agent is not a crawler hoovering your data. It is a customer acting for one particular person, inside that person’s consent. Blur those two and you have built an exposure, not a storefront. The work is not to open everything to every bot. It is to know which person an agent stands for, what they are cleared to do, and where the line sits.

This is the same problem we usually write about, seen from the other side of the table. Most of our writing is about securing the agents you run. The agents coming to buy from you are the mirror image, and they need the same primitives: identity, scope, consent. Those are not the tax on agent-commerce. They are what make it possible. A buyer you cannot identify is a fraud you have not met yet.

So the business that wins the agent customer is not the one with the loudest page. It is the one whose identity and consent layer is real, so an agent can act with confidence and the person it serves stays in control. That is the whole bet behind building a private operating system for AI instead of one more place to pour data.

Where to start

You do not need an agent-commerce department. You need to walk your own talk, in this order: publish a llms.txt and structured data so an agent can read you; expose one thing it can actually call; make your pricing and terms machine-legible; and put scoped, revocable authority around anything it can spend or change. Then watch what reaches you, because the traffic will tell you what the agents want before any forecast does.

None of this is finished. The protocols are weeks and months old, and plenty will change. But the direction is not in question, and the move is the one we ask of everyone we work with: get ready before you have to, and prove it on yourself first. Our own machine-readable front door went up before this post did. The first agent that buys from you will not introduce itself. It will simply expect you to be ready.