Agentic Browsing in Google PageSpeed Insights Your service is currently being assessed in terms of AI agents
Google has added a fifth category to PageSpeed Insights, which checks whether your website can be accessed not by a human, but by an AI agent. We explain exactly what it measures, why it’s important, and how to optimise a TYPO3 website — from llms.txt to layout accessibility and stability.
In brief
What happened:
In May 2026, Google added a new category called ‘Agentic Browsing’ to Lighthouse (the PageSpeed Insights engine). This checks how well an AI agent can read and process your website — alongside the familiar categories of Performance, Accessibility, Best Practices and SEO.
The three default tests: a
clean accessibility tree, a stable page layout (low CLS) and the presence of a valid llms.txt file in the domain’s root directory. The result is not a score from 0 to 100, but a ratio (e.g. 3/3).
Is this a ranking factor?
No. Google clearly states that this is not a ranking factor. It is a diagnostic tool, labelled as experimental and ‘under development’. Don’t panic — but do take note.
Why this matters: An
increasing proportion of web traffic comes from AI agents (OpenAI’s Operator, Anthropic’s Computer Use, Google’s Project Mariner, Perplexity), which act on behalf of users. Google has now made their needs measurable.
Good news for TYPO3:
Most of the requirements align with what we’re already doing well: semantic HTML, WCAG accessibility, a stable layout and llms.txt. A well-built TYPO3 site has a head start here.
For years, PageSpeed Insights was a tool that answered one question: how quickly does your website load for a human user? In 2026, Google added a completely new type of question to it: can an AI agent — software acting on behalf of the user — read, understand and interact with your website? This new category is called Agentic Browsing and has appeared alongside the familiar ones: Performance, Accessibility, Best Practices and SEO.
This may seem like a technical development, but it signals a profound shift in how we think about websites. The web is no longer browsed exclusively by humans — increasingly, it is being read by machines acting on our behalf. In this article, we explain what Agentic Browsing is, exactly what it checks for, what it does NOT mean, and what specific steps you should take on your TYPO3 website to ensure it is ready for a world where your visitors may include AI agents.
What is ‘agentic browsing’?
Agentic browsing is a situation in which a website is visited not by a human scrolling and clicking, but by an autonomous AI agent — which programmatically reads the page, understands its structure, locates the relevant elements and carries out a task. Instead of ‘click, scroll, read’, the agent asks: where is the ‘Add to basket’ button, how do I fill in this form, how do I compare these prices?
This is no longer just a theory. By 2026, a whole range of agents operating on behalf of users will be navigating the web: Operator from OpenAI, Computer Use from Anthropic, Project Mariner from Google, Perplexity, and ChatGPT’s browsing mode. When a user asks an AI assistant to “find the best deal” or “book an appointment”, it is an agent — not a human — that visits your website and attempts to carry out the task.
The difference compared to a traditional search engine bot is fundamental. Google’s crawler indexes content in order to rank it. An AI agent performs actions – it fills in a contact form, compares your price list with that of your competitors, extracts an offer and presents it to the user as ready-made options. A website can score 100/100 in traditional SEO and still be unusable for the agent.
Lighthouse 13.3 and PageSpeed Insights Where did it come from?
PageSpeed Insights runs on an engine called Lighthouse — the same one that powers the Lighthouse panel in Chrome DevTools and dozens of monitoring tools. Lighthouse is updated every few weeks, usually without users noticing. This update was different.
On 7 May 2026, Lighthouse version 13.3 moved the ‘Agentic Browsing’ category from an experimental feature (hidden behind a flag) to the default configuration. This is the moment when a niche experiment becomes a standard part of the report. Around 21 May 2026, PageSpeed Insights began displaying results for this category to all users. If you ran a speed test on your site earlier and then ran it again, you’ll see a new section that wasn’t there before.
An important reassuring note: if your standard performance score now looks different, it is almost certainly not due to Agentic Browsing. This category does not affect the 0–100 score. Changes in the performance score are due to periodic adjustments to Google’s methodology or normal data variability — not to the new category.
What exactly does Agentic Browsing check?
In PageSpeed Insights, the default Agentic Browsing score consists of three tests. Important: there is a lot of incorrect information circulating on this subject, including made-up audits that do not exist. Here is the version in line with the official Lighthouse documentation.
Pure accessibility treeAI agents do not read a page visually, but via what is known as an ‘accessibility tree’ — a simplified map of the roles, names and states of interactive elements (the same one used by screen readers). The test checks whether buttons, links and form fields have meaningful labels, correct ARIA roles and readable names. Will the agent recognise ‘Submit enquiry’ as a button, rather than an unnamed
with click handling?
Stable page layout (Cumulative Layout Shift)This is the well-known CLS metric — which measures how much page elements shift around whilst the page is loading. For a human, a jittery layout is irritating. For an AI agent, which operates faster than a human and often ‘views’ a page via a screenshot, a shifting layout means clicking in the wrong place. The recommended CLS is below 0.1.
A valid llms.txt file in the root directoryllms.txt is a proposed file standard (in Markdown format) that concisely describes to AI agents what the website is about and where its key subpages are located. The test checks whether the file exists at /llms.txt, whether it has an H1 heading, whether it is not too short, and whether it contains links.
The fourth test outside the default scoring system: WebMCP
There is also a fourth, more advanced audit — WebMCP — but it is deliberately excluded from the default scope. WebMCP is a proposed standard that allows a website to explicitly declare the actions an agent can perform (submit this form, add to basket, start a booking) — rather than having to guess from the page layout. However, it requires Chrome version 150 and registration for the official Origin Trial, and is currently supported by only a tiny fraction of websites worldwide — even large, well-built ones. That is why it is not included in the default score, and for the time being there is no need to worry about it.
Scoring Why is it a ratio, rather than 0–100?
Unlike a traditional performance score, Agentic Browsing does not provide a single figure between 0 and 100. It shows a ratio — for example, 3/3 — indicating how many readiness tests your website passes. This is a deliberate design choice: in an area where standards, agent behaviour and protocols are constantly changing, a single weighted score would be misleading.
Key point: this is a list of opportunities, not a penalty table. You won’t ‘fail’ this category simply because you don’t have any AI features. Even a minimalist website (such as example.com) can score full marks because it meets the basics: a clean structure and a stable layout. Treat the result as a checklist for developers, not a verdict.
What ‘Agentic Browsing’ does NOT mean Important disclaimers
✗ This is not a Google ranking factor. Google’s Search Team has explicitly stated that llms.txt is not used for ranking or for AI Overviews. The category originates from the Chrome team and relates to web crawlers, not search engine visibility.
✗ This is not a mature standard. The category is explicitly labelled as ‘experimental’ and ‘under development’. The tests and scoring methods may still change.
✗ A poor score is not a cause for alarm. Most websites currently fail the most difficult tests (WebMCP, agents.json) — and that’s perfectly fine.
✗ This does not require the implementation of “AI features”. Most of the requirements are solid foundations that are worth having anyway.
It is worth noting, however, a certain internal contradiction within Google itself: the Search team says that llms.txt is not necessary, whilst the Chrome team has just introduced an audit that checks precisely this file. This clearly illustrates that we are at an early, uncertain stage — it’s worth keeping an eye on, but there’s no need to overturn your entire strategy because of it.
Why does it matter anyway? The direction is clear
It is easy to dismiss this category: experimental, small-scale, still in its infancy. However, that would be a mistake. The trajectory is clear. AI agents are becoming full-fledged web users alongside humans and search engine bots. Google has just made their needs measurable — and that is usually the first step towards something becoming the norm.
There is another compelling argument: PageSpeed Insights has a greater impact than its name suggests. PSI scores end up in client audits, sprint tickets, boardroom questions and tenders. A new category within this report doesn’t stay solely in the hands of engineers for long — it quickly finds its way into client meetings, bid comparisons and SEO roadmaps. A company founder can run this test. A client can run it before a sales meeting. A procurement team can use it to compare suppliers.
It’s also worth remembering the history: llms.txt went from being a niche feature to a Lighthouse test in about a year. The ‘experimental, Origin Trial only’ label very quickly turns into ‘default’. Anyone who prepares their website now will be ready when the bar is raised.
Our own test drblitz-weblab.com scored 3/3
Rather than just discussing this in theoretical terms, we have launched Agentic Browsing on our own website. 7 July 2026, mobile test on PageSpeed Insights (Lighthouse engine 13.4.0), at drblitz-weblab.com. Score in the Agentic Browsing category: 3/3 — full marks. The site passed all three default tests: clean accessibility structure, stable layout and a valid llms.txt file.
The most interesting part is the comparison with the other results. In the same test, the site scored 100/100 for Accessibility, 100/100 for SEO, 96 for Best Practices — and, crucially for agents, a Cumulative Layout Shift of 0 (ideal). The Performance score, however, was 67, with Largest Contentful Paint at 6.8 s — so there is room for improvement in terms of raw speed.
And that is precisely the crux of the matter. Our website achieved full marks in Agentic Browsing not because it is the fastest — as we do indeed have work to do in terms of performance — but because it has a clean accessibility structure (100/100), zero CLS and a valid llms.txt file. This confirms the thesis of this article in black and white: readiness for AI agents depends on semantics, accessibility and layout stability, not on raw loading speed. A page can have an average performance score and still achieve a full 3/3 in Agentic Browsing — because these are two different things.
For us as an agency, this is a practical lesson that we apply to our clients’ projects: the foundation of accessibility that we build in line with WCAG and Polish law simultaneously makes websites ready for AI agents — even before the majority of the market had begun to think about it. This is exactly the same approach we described in our article on combining TYPO3 SEO with accessibility and WCAG — accessibility and visibility (for people, search engines and now AI agents) are one joint endeavour, not three separate ones.
What does this mean for a TYPO3 website? Why do you have the advantage?
Here’s the good news. Most of what Agentic Browsing rewards isn’t ‘new AI work’, but the fundamentals of good web engineering — precisely those that a well-built TYPO3 site already has. Let’s take a look at the three tests from a TYPO3 perspective.
Availability the foundation we are already building
A clean accessibility structure is simply semantic, accessible HTML that complies with WCAG. TYPO3 gives you full control over every element of a template (Fluid), so creating correct labels, ARIA roles and a sensible heading structure is well within our reach. This is the same foundation we use for compliance with WCAG 2.1/2.2 AA — now required by the Polish Accessibility Act and the BFSG. In other words: working on accessibility for people with disabilities automatically improves readiness for AI agents. One investment, two benefits. This is also where the most important – and at the same time the most ‘boring’ – piece of advice comes in: stop rendering text and tables as images. An image of a table is invisible to the accessibility tree. A genuine HTML table (or the component that renders it) is readable by screen readers, bots and agents — straight away. This is the most common pitfall of complex page builders, but one that is easily avoided in a well-maintained TYPO3 installation.
System stability The issue of effective implementation
A low CLS is the result of careful implementation: fixed image and embed dimensions, no elements that load late and cause content to shift, and well-planned font loading. This is the same work we do for Core Web Vitals and search engine optimisation. By improving CLS for users and SEO, you’re also improving it for search engine crawlers.
llms.txt Cheap, quick, worth it
The llms.txt file is the simplest of the three elements. It is a standard Markdown file in the domain’s root directory, which describes what the company does and links to key subpages — written in such a way that the AI model can read it quickly. Preparing it takes a matter of hours, not a development sprint. Even if Google’s Search Team claims it does not affect rankings — the risk is virtually zero, and the file is already checked by Lighthouse and read by some crawlers. In TYPO3, it can be implemented statically or generated dynamically from the site tree.
Action plan Where to start
If you want to prepare your TYPO3 site for the world of AI agents, here’s a sensible order to follow — from low-cost, high-impact measures to more advanced ones.
- Run a test. Enter your URL into PageSpeed Insights and check the ‘Agentic Browsing’ section — you’ll see which of the three tests you pass.
- Add llms.txt. The cheapest step with the best return on effort. An H1 heading, a short description, links to key subpages.
- Improve accessibility. Labels for buttons and links, correct ARIA roles, no ‘empty’ clickable elements. This work is required by PAD and BFSG anyway.
- Replace text-as-image elements with proper HTML. Tables, price lists, diagrams rendered as images — rewrite them as semantic HTML.
- Stabilise the layout. Image and embed dimensions, font loading control, elimination of content shifts — CLS below 0.1.
- Strengthen structured data. Schema.org (Organisation, Service, FAQPage, Article) provides search engines with context they can trust.
- Keep an eye on WebMCP and agents.json. These are experimental for now — don’t force implementation, but do keep track, as ‘experimental’ quickly becomes ‘default’.
Note that steps 3–6 are efforts that pay off for other reasons anyway: accessibility (legal compliance), Core Web Vitals (SEO), and structured data (search engine visibility and AI). Agentic Browsing doesn’t create a new to-do list — rather, it rewards the things that a good agency does anyway.
The bigger picture Findable, citable, usable
Agentic Browsing is part of a wider shift taking place online. In the past, there was just one question that mattered: is my content findable — in other words, SEO. Today, two further questions have emerged: is it citable by AI models — this falls under AEO/GEO (Answer Engine / Generative Engine Optimisation) — and is it usable by an agent — this is precisely what agentic browsing is all about.
The websites that will succeed in this new reality will be findable, citable and usable all at once. The good news is that the foundations of all three largely overlap — clean structure, semantic HTML, structured data, a stable layout, and a clear llms.txt file. A page that a machine can easily read is a page that a machine is more likely to cite and on which it operates more efficiently.
Frequently Asked Questions
No. Agentic Browsing does not affect the standard 0–100 performance score. If your speed score has changed, it is almost certainly due to periodic adjustments to the PageSpeed Insights methodology or data fluctuations — not because of the new category. There is no cause for concern here.
Not directly. Google has clearly stated that llms.txt is not a ranking factor and does not affect AI Overviews. The new Lighthouse audit checks this file for a different reason — for AI agents crawling the web via Chrome. It is worth adding it for the sake of being prepared for these agents and due to the negligible risk, rather than for search engine rankings.
It doesn’t depend on the CMS, but on the quality of the implementation. Agentic Browsing rewards semantic HTML, accessibility, a stable layout and llms.txt — and this is a matter of how the site is built. TYPO3 gives you full control over templates (Fluid), which makes it easier to meet these requirements. Elaborate page builders (common in WordPress) are more likely to generate messy code — button icons without labels, empty links, images instead of text — which lowers the score here.
Not yet. WebMCP is experimental; it requires Chrome 150 and registration with Origin Trial, and almost every website — including the major ones — fails it at the moment. It isn’t included in the default scoring. It’s worth keeping an eye on, but don’t force its implementation. Focus on the three default tests first.
The simplest way is to enter the URL into PageSpeed Insights (pagespeed.web.dev) and look for the ‘Agentic Browsing’ section. Alternatively, run Lighthouse in Chrome DevTools with the ‘Agentic Browsing’ category selected, or use Lighthouse from the command line (requires the latest version). Please note: different tools may run different versions of Lighthouse, so the results may vary slightly.
About the author
Dr Blitz-Weblab is an agency specialising exclusively in TYPO3, with TYPO3 Silver Member status and an office in Kraków. For the past eight years, we have been working exclusively with TYPO3 — creating, updating and maintaining websites for clients in Poland and the DACH countries. We specialise in accessibility (WCAG 2.1/2.2 AA, the Polish Accessibility Act, BFSG), security and compliance (NIS2, GDPR, CRA), and combining SEO with accessibility within a single process. Our portfolio includes, amongst others, the Universitätsmedizin Mannheim portal (25,000 subpages in multiple languages), the website of the Town of Mszana Dolna, RED NOSES International and the International Youth Meeting Centre in Oświęcim. This article is based on the official Lighthouse documentation and our own tests on live websites — including our own site.
Would you like to find out how your TYPO3 website performs in Agentic Browsing and what could be improved? At drblitz-weblab.com/kontakt, you can select a convenient time for a 60-minute consultation from our calendar.
How can we help?
Preparing a TYPO3 website for the world of AI agents is largely the same work we do for accessibility, performance and SEO — just from a new perspective.
We offer:
✓ AI agent readiness audit — Agentic Browsing test, accessibility, CLS, llms.txt, structured data, with a report and priority
list✓ llms.txt implementation — statically or dynamically from the TYPO3 site
tree✓ Accessibility improvements (WCAG 2.1/2.2 AA) — compliance with PAD and BFSG plus readiness for AI agents in one✓
Optimisation of Core Web Vitals and layout stability (CLS)
✓ Implementation of Schema.org structured data for better visibility in AI and search engines
. Telephone: 12 333 44 01. Email: [email protected].
