Agentic Browsing for AI Text

A New Audience: AI Agents

Websites have historically been optimized for two audiences: people and search engine crawlers. Recently, a third has emerged – AI agents that act on behalf of users, independently navigating websites, reading information, and carrying out actions such as filling out a form or booking an appointment.
Google is now accounting for this development. With Lighthouse 13.3 (early May 2026), a new category called "Agentic Browsing" appeared. Lighthouse is the audit engine behind the well-known PageSpeed Insights tool. The new category sits conceptually alongside Performance, Accessibility, Best Practices and SEO – and checks something none of the others cover: how well can an AI agent understand and operate your page?
Important context: the category is explicitly still "under development", i.e. experimental. It doesn't produce a 0–100 score but a simple pass ratio (e.g. "2 of 2"). And as of today, it's not a confirmed ranking factor. Rollout into the hosted PageSpeed Insights interface is happening gradually; the checks themselves, however, can already be run today (e.g. via the current Lighthouse version or third-party tools).

What Exactly Is Checked

The category combines four signals:

1. The accessibility tree – the machine-readable structural representation of your page. AI agents process this simplified structure far more efficiently than the full HTML code or a screenshot. A clean, correctly marked-up structure therefore helps screen readers and agents alike.
2. Layout stability (the well-known CLS metric). If page content jumps around while loading, an agent – which often acts faster than a human – can easily click the wrong element.
3. WebMCP: a new interface through which a website can offer agents specifically defined actions (for example via annotated forms). This is the genuinely new building block that goes beyond classic optimization.
4. An llms.txt: a proposed file at the website root that gives AI systems structured pointers to your content. The check verifies whether the file exists and is sensibly structured (heading, scope, links). Note: llms.txt is so far just a proposal and isn't yet widely used by AI tools.

How Does the Watermark Work Technically?

The method is called SynthID-Text and originally comes from Google DeepMind (published in 2024 in the journal Nature); the theoretical foundation traces back to a 2022 proposal by computer scientist Scott Aaronson. The trick: while writing, Claude repeatedly chooses between several equally valid phrasings. Normally, this choice is random. In the watermarking process, that randomness is instead steered by a cryptographic key that, together with the preceding text, creates a pattern – invisible to humans, but detectable later with the right key.

Detecting it requires three things: the cryptographic key (which only Anthropic holds), a sufficiently long passage of text, and the unmodified original text. The watermark doesn't reveal which user or organization generated the text – only that a Claude model was involved. Anthropic has yet to release a public verification interface for third parties.

The Good News: You're Already Doing a Lot Right

If you already prioritize accessibility, fast load times and structured content, you automatically meet most of the new requirements. Three of the four checks overlap with quality traits a well-built website should have anyway – and that we've been implementing at WACON for years:

• Our work on low-barrier TYPO3 websites ensures a clean accessibility tree.
• Our pagespeed and Core Web Vitals optimization keeps the layout stable. And our work on SEO and GEO (visibility in search engines and AI systems, e.g. via structured data with schema.org) lays the groundwork for machines to correctly capture your content.

What's new and exciting above all is WebMCP – a standard is currently emerging that lets websites offer specific functions to agents. We track exactly these kinds of developments for our clients and implement them as soon as they're production-ready.