What is agentic browsing?

For twenty years a website had two kinds of visitor: a person, and a crawler pretending to care what the person sees. In 2026 a third kind arrived in numbers, an AI agent that opens your site not to read it but to use it, clicking and typing its way through a task for someone who may never look at your interface at all. That behaviour has a name now, agentic browsing, and in May 2026 Google made it measurable by adding an Agentic Browsing category to Lighthouse. This is the plain-English version: what the term means, how it differs from crawling, and why it suddenly has a score attached.

01 · The shiftWhy the word exists now

The trigger was a wave of AI browsers and agent modes reaching real users. Google put Gemini directly inside Chrome and folded its Project Mariner research into a browsing agent. OpenAI shipped the Atlas browser and a ChatGPT agent mode. Perplexity shipped Comet. What these share is the important part: they do not just answer from a trained model, they open a live browser and operate real sites to get something done.

That breaks a twenty-year assumption. Your page was built for a human with eyes and a cursor, and search crawlers were tolerated as a necessary second reader. Now a third reader can navigate, decide, and act, and it experiences your site through a completely different channel than a person does. “Can a human use this?” and “can an agent use this?” turn out to be different questions, and most sites have only ever answered the first.

02 · The definitionA plain definition

The load-bearing word is act. A search crawler and an AI agent might fetch the exact same HTML, but the crawler’s job ends at “understand this page,” while the agent’s job is “use this page to accomplish something.” An agent that reads your pricing perfectly but cannot find, focus, and submit your signup form has failed, in a way a crawler never could.

03 · The distinctionBrowsing is not crawling

The two get blurred constantly, usually by someone selling you something. Keeping them apart is the whole point of the term.

The bottom two rows are where agentic browsing adds genuinely new requirements. A crawler does not care whether your “Buy” button drops 200 pixels half a second after load, or whether it is a real <button>Buy</button> or a <div> with a click handler. An agent cares enormously about both, because it locates and operates controls the way an assistive screen reader does, through structure and labels, and it clicks based on where things are.

04 · The scoreWhat Google decided to measure

Because these agents browse through Chrome, Google can measure how well a site holds up, and in May 2026 it did. Lighthouse 13.3 added an Agentic Browsing category (now in PageSpeed Insights too) that checks four things an agent leans on:

1. A well-formed accessibility tree, so the agent can find and name your controls.
2. A stable layout (low Cumulative Layout Shift), so it does not click where a button used to be.
3. An llms.txt file, so it can map your site without crawling every page.
4. WebMCP tools, an emerging way to hand an agent explicit actions instead of making it guess from the DOM.

The unusual part: it is not a 0-to-100 score. It reports a pass ratio, like 3/3, because the standards are still young and Google is gathering data rather than ranking anyone. The mechanics of each check, and the Chrome-version caveats, are their own guide: Lighthouse’s Agentic Browsing category, explained. WebMCP is new enough to deserve its own definition too.

05 · Necessary, not sufficientThe catch: readiness is not distribution

Here is the part the audits will not tell you. Passing every agentic-browsing check makes your site usable by an agent. It does not make an agent show up.

An agent lands on your site the same way a person does: it was searching, comparing, or following a recommendation, and something pointed it at you. If nothing references your product, a flawless accessibility tree changes nothing, because no agent has a reason to be there. In Nilkick’s model this is the split between Readiness and Footprint: agentic browsing lives almost entirely inside Readiness, the half you fully control, and it is worth little without the Footprint half that gets you found in the first place.

So treat agentic browsing as a new floor, not a new game. It is a real, cheap-to-clear floor, and most of clearing it (real server-rendered content, labelled controls, a stable layout) is work you owed your human visitors anyway. But a site an agent can use is still not a site an agent will use. That second half is the harder, older problem, and no audit scores it for you.