Your next visitor is an agent, and Chrome just taught it to read

We are moving quickly into a world where most of the visits to your website will not be made by a person. A human will want something, and their agent will go get it for them. So I want to share some new information about how agents actually see your site, because for the first time we have a clear picture of it.

Dan Petravich recently published a genuinely interesting teardown of how Google's Chromium understands the content on a page. Chromium ships with something called a page content agent, and his experiments extracted exactly what it produces when it looks at your website. All of this happens inside the Chrome browser itself. Other browsers are surely working on the same thing, but I doubt any of them are close to Chrome here, and it matters less than you would think: most browsers run on Chromium anyway. If you use Microsoft Edge, you are effectively using Chrome. Same for most headless browsers. Playwright and Puppeteer? Chromium. And if you run Gemini in Chrome, the browser hands this data straight to Gemini, though it could be handed to any large language model. Since essentially every AI can drive some version of Chromium, it is safe to assume every AI has access to what Chromium extracts.

Markdown is harder than it looks

What the teardown shows is that Chrome's page content agent has gotten remarkably good at taking a web page and stripping it down to Markdown for the agent. That sounds trivial. Grab the HTML, strip the tags, done. In practice it is much messier. Plenty of what renders on a page never appears in the raw HTML. Some elements arrive via JavaScript. Some content is pulled from your specific account after login. Some of it is text you typed into a form field two minutes ago. The page content agent captures all of it.

Until a few months ago, the only way an AI could get at that information was screenshots. Take one, scroll, click, take another, wait a couple of seconds for each to process. Slow, clumsy, and expensive in tokens. The page content agent runs entirely on your own hardware, so extraction is close to instant and essentially free. Because it is local, it can also see sensitive things. The teardown found it will read what is sitting in your bank dashboard or a password field, and Google then strips the sensitive parts before anything leaves your device.

Reading was step one. Clicking is step two.

The part that should really get your attention is that this is not just about reading. Every element the agent documents comes with coordinates. The AI knows exactly where the buttons are and can press them without ever rendering the page for a human eye. Agents just got dramatically better at operating websites, not only understanding them. We have already covered what happens when an agent reads the wrong thing on a page, so I will just note that both edges of this sword got sharper this week.

What can you do about it today? Open your website in Chrome, open Gemini in the side panel, and ask it for the raw output of its page content agent for the page you are on. Then actually read it. Check that your products, your prices and your calls to action survive the translation to Markdown. In my experience it does a surprisingly good job even on sites that never prepared for agents. But surprising is not the same as guaranteed, and this output is now the version of your website that matters.

Opinion columns reflect the personal views of the author. Our reporting on the stories referenced here lives on the linked pages.