Insight · July 20, 2026
Is your website
agent-ready?
More of the web is machines than people now. Most sites still answer only to the eye.
01 · The shift
Your newest visitor does not have eyes.
In June 2026, Cloudflare reported that automated traffic reached 57.5 percent of requests for HTML content, against 42.5 percent from humans. For the first time in the history of the web, machines outnumbered people. The company had not expected the crossover until 2027.
The fastest growing slice was not training crawlers. It was agents, software acting for a person, browsing pages, filling forms, comparing options, and finishing tasks. Cloudflare put that category near two percent of automated traffic at the start of 2025 and roughly eighty times larger by the end of it. A visitor that used to be rare is now the majority, and it does not see your site the way you designed it.
02 · What an agent sees
An agent reads the tree, not the picture.
When a person opens a page, the browser paints pixels. When an agent opens the same page, it works from the accessibility tree, the internal model a browser builds from the DOM so that machines can understand content. Each element in that tree carries four things: a name, a role, a state, and a description. A button has the role button and a name like Submit. A checkbox carries its checked or unchecked state. The tree also records what can be done, that a link can be followed and a text field can be typed into.
This is not a fringe detail. The open source browser tools built for agents, like Vercel's agent-browser, navigate exactly this way. They take a snapshot of the accessibility tree, read the roles and accessible names, and act on stable references rather than fragile visual selectors. The design you labored over is not what the agent consumes. The semantics underneath it are.
The one line to keep
“A human reads your website. An agent reads your markup. Only one of them is now the majority.”
03 · What breaks
Five things that quietly lock an agent out.
01
Content that only appears after scripts run
If the words a visitor needs show up only after client side code executes, an agent reading the first response sees an empty shell. Rendering the full page is a separate, slower step, and not every agent takes it.
02
Interstitials that sit on top of the page
Consent prompts, cookie walls, and modal overlays are built for a person to click through. To an agent they are a blank wall between the request and the content, and the thing it came for never enters the tree.
03
Controls with no accessible name
An icon button with no label is, to the machine, a control named nothing. The agent can tell a button exists. It cannot tell what the button does. Every unlabeled control is a dead end.
04
Meaning carried by layout alone
A price that is large and bold is obvious to the eye and invisible to the parser. If the only signal that a block is the price, the stock, or the primary action is its visual weight, the agent has to guess.
05
Pages too heavy to read
Language models work inside a fixed context window. A page buried in navigation, advertising, and script is mostly noise. The signal has to be findable, or it gets cut off before the agent reaches it.
04 · The fix
Agent-ready is mostly good engineering you already know.
None of the fixes are exotic. Render the content that matters on the server, so it is present in the first response and not conditional on a script. Use real semantic elements, headings, lists, buttons, and labels, so the accessibility tree describes the page accurately. Mark up the facts with schema.org, the vocabulary that already sits under more than one hundred million websites as a semantic layer. Keep the critical path light.
A new layer is also forming on purpose. The llms.txt proposal puts a plain markdown file at the root of a site, in the spirit of robots.txt and sitemap.xml, because context windows are too small to hold most sites and raw HTML is full of navigation and advertising the model has to wade through. It hands a machine a clean, concise version of what matters. Microsoft's NLWeb goes further, turning a site into an endpoint an agent can query in natural language and get schema.org answers back over the same protocol that serves people. Its authors put it in one line: NLWeb is to MCP and A2A what HTML is to HTTP.
05 · The same page, read two ways
A photograph to a person. A fact to a machine.
Picture a product page. A person sees a photograph, a bold number, and a confident button. The agent sees whatever the markup admits. If the number is a styled span with no role and the button is a bare div, the agent gets a shape it cannot name. If the number is marked as a price and the button carries an accessible name like Add to cart, the agent gets a fact and an action it can take.
The difference is not more work. It is the same page, described honestly. A small llms.txt at the root makes that honest version explicit, a machine readable table of contents that points at the pages worth reading.
# Acme Studio > Independent product studio. Design, engineering, and > brand for teams shipping software. ## Services - [Product design](https://acme.example/design): scope, research, and interface design - [Engineering](https://acme.example/build): production software with agents in the loop ## Optional - [Field notes](https://acme.example/notes): essays, lower priority for short context
A root level llms.txt · the machine readable index for your site
06 · Why it matters
Agents are a buying audience now, not a curiosity.
The reason this matters is not tidiness. Agents are starting to do the work that ends in a decision. They gather options, they compare, and increasingly they act. A site an agent cannot read is a site left off the shortlist before a human ever weighs in.
The same discipline that makes you visible inside AI answer engines, clean structure and honest markup, is what makes you usable to the agents those engines are becoming. This is the through line from search visibility to the agentic web. First the model had to be able to cite you. Now it has to be able to use you.
07 · The caution
Do not build a second site for the machines.
The wrong response is a hidden, machine only version of your site that says something different from what people see. That is the old cloaking mistake in new clothes, and it breaks trust the moment the two versions disagree. The goal is one honest page that serves both readers.
Be deliberate about the walls too. Consent and privacy obligations do not disappear because the visitor is software. An overlay that blocks an agent is also a sign that you have not decided how automated visitors should be treated. Decide it on purpose, rather than leaving a modal to make the call for you.
Closing
Half the web cannot see. Make sure it can still read you.
Open your most important page and look at its source, not its design. Strip the styling in your head. If what remains does not name your product, your price, and your primary action in plain markup, that is the version most of your visitors already get. Fix that page first.
Cloudflare Radar traffic figures · June 2026 · accessibility tree behavior per MDN Web Docs and Vercel's agent-browser · llms.txt proposal, AnswerDotAI · NLWeb, Microsoft · verified July 2026
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.