For twenty years the web shipped broken structure and got away with it. The human visitor paid the repair bill. The agent will not.
web.dev published a line that is easy to read past.
Your website has a new type of visitor.
They meant the AI agent. The system a person hands a goal to. Book the table. Compare the two plans. Find the part that fits. The agent goes to the site and acts.
The line is true. But the interesting part is not that the visitor is new.
It is what the new visitor refuses to do.
For twenty years we built websites for a reader who repaired them while using them.
A button that is a div, styled to look clickable. The reader waits half a second, senses it will work, clicks it. A form field with no label. The reader looks at the field beside it and infers. A layout that shifts while the page loads. The reader waits, then finds the thing again.
None of that was free. It was paid. The reader paid it, in small amounts of effort, on every visit, and never sent an invoice.
So the web learned a quiet lesson. Structure could be sloppy. Markup could be wrong. A page could lie about what its own parts were. A human would stand in the gap and make it work anyway.
We called that good enough, because nothing ever made us call it anything else.
The agent stands in no gaps.
It does not see a website the way you do. web.dev is precise about this. An agent works from three machine-readable views of a page. A screenshot, read by a vision model. The raw HTML and the structure of the DOM. And the accessibility tree, the browser’s own summary of what each element is and does.
Read that list again. Two of the three are things most teams have never once looked at.
The agent reads all three. Then it acts on what it found.
If a button is a div, the agent sees a div. If a field has no label, the field has no meaning. If the layout moved, the screenshot no longer matches the page.
Picture the checkout. A person reaches the last step, sees a box that reads Place order, and clicks it. The agent reaches the same step and asks the page for a button. It finds none. The box is a div with a click handler bolted on. There is nothing to act on. The order does not get placed. Not because the agent is weak, but because the page never declared the one thing that mattered.
The agent can still guess. It reads the picture, the color and size and position of things. But the picture is a fallback, not a source. Guessing from pixels is slow, costly, and easy to get wrong. The agent does it only when the code failed to tell it the truth.
The human guessed for free, in half a second, and never minded. The agent guesses last, and at a price.
That is the gap the human used to fill. The agent hands it back to you.
So the agent is not a new audience to optimize for.
It is the first visitor that audits you.
It arrives, reads what the page says about itself in code, and either finishes the job or cannot. No patience. No benefit of the doubt. The page works as built, or it does not.
That changes the reflex the industry is about to have.
The reflex will be to treat the agent as a new surface. A new file to add. A new schema. A layer of markup written for machines.
Google answered that in the same guidance that named the agent. The optimization guide it published on May 15, 2026 spends a full section on what site owners do not need. No llms.txt. No special AI markup. No content sliced into machine-sized pieces.
The instruction is the opposite of a new checklist.
web.dev puts it in one sentence. Everything they suggest to make a site agent-ready also makes it better for humans. Semantic HTML. A button that is a button. A label tied to its field. A layout that holds still. An accessibility tree treated as something you ship, not something you forgot.
None of it is new.
That is the uncomfortable part.
This is not new advice. It is old web advice, ignored because the cost stayed hidden.
Accessibility specialists asked for exactly this, for people using screen readers, long before any agent existed. SEO asked for clean structure and honest markup since before the smartphone. Both were heard. Both were filed under later. Every time for the same reason. It was never punished. The site ranked anyway. The user converted anyway. The human visitor was generous enough to cover the difference.
The agent ends that subsidy.
It does not ask a page to do more than it ever claimed. It asks the page to finally be what it always said it was. A button that is a button. A field that states what it is. A structure that matches the surface.
The work was always due. We just had a visitor kind enough never to collect.
My last piece was about a number. The click. The Receipt, Not the Transaction traced how that metric slowly stopped describing the business it was built to measure.
This one is about the thing under the number. The page itself.
Now a visitor has arrived that forgives nothing and reads everything.
The question worth sitting with is not how to optimize for it.
It is this. How much of what we built only ever worked because a person was quietly willing to fix it for us. And what happens to everything we ship next, now that the visitor has stopped doing us that favor.
