Google's New Test Asks If AI Agents Can Use Your Website. Here's How Mine Scored 3 of 3.

When I finished rebuilding my portfolio website, I ran it through Google's PageSpeed Insights like I always do. Performance 97. Accessibility 100. Best Practices 100. SEO 100. I was happy.

Then I noticed a fifth row I had never seen before: Agentic Browsing, 3 of 3.

Curious, I tested the two client websites I had built earlier, an oil and gas company and an online French school. The oil and gas site showed 2 of 2. Not a failure, but not the same either. When I dug into why, the difference turned out to be one tiny text file sitting at the root of my website, called llms.txt. Neither client site had one.

Then I found something that surprised me even more. Google Search says plainly that this same file does nothing for your rankings. So one Google tool checks for it, and another Google team says you can ignore it.

Both statements are true, and once you understand why, you will understand where the web is going. In this post, I will explain what AI agents are, what this new test actually checks, what llms.txt really does and does not do, and the simple steps to make your own website ready for a new kind of visitor.

Your website now has two kinds of visitors

For as long as the web has existed, websites were built for one kind of visitor: a human with eyes, a mouse or a finger, and some patience.

That is changing. A growing number of visitors are now AI agents: software that goes to websites and does tasks on behalf of a person. Google's own guide to AI search describes them as autonomous systems that can do things like booking a reservation or comparing product specifications. It also explains how they see your website. Some look at screenshots of the page. Some read the page's underlying structure. And many rely on something called the accessibility tree, which I will explain in a moment.

Picture a busy professional in Lagos who types into an AI assistant: "Find me an online French tutor who prepares students for the TCF exam, and tell me her prices and class times." The assistant does not just give a list of links. It visits websites, reads them, compares them and comes back with an answer.

Now ask yourself an uncomfortable question. If that agent landed on your website, could it find what it needs? Could it understand your services, read your prices and find how to contact you? Or would it get lost, give up, and recommend your competitor instead?

This is exactly what Google's new test tries to measure.

What the Agentic Browsing score actually checks

The test that appears in PageSpeed Insights comes from Lighthouse, the free testing tool built by Google's Chrome team. Its new Agentic Browsing category looks at how easily a machine can read your page and act on it.

The first thing to know is that it is not scored out of 100 like the other categories. According to the Lighthouse documentation, it shows a ratio of how many checks your site passes, such as 2 of 2 or 3 of 3. Google also labels the whole category as experimental and based on proposed standards. In other words, the rules of this game are still being written.

Here is what it looks at, in plain English:

• Accessibility for agents. Agents rely heavily on the accessibility tree, which is a simplified map of your page that browsers build for screen readers. It lists every button, link, form field and heading, and what each one is called. If a button has no name, or a form field has no label, an agent cannot tell what it does.

• Layout stability. If things jump around while your page loads, an agent may try to click a button just as it moves away. This is the same Cumulative Layout Shift we talked about in my earlier posts.

• llms.txt. Whether your website has a short summary file written for AI systems at its root.

• WebMCP. A brand new, proposed way for websites to tell agents exactly what actions they offer, like "book a class" or "request a quote". It is still being tested and most websites won't show this check at all yet.

So why did my oil and gas client's site show 2 of 2 while mine showed 3 of 3? The llms.txt check is marked "not applicable" when a website has no llms.txt file, because the file is optional for now. Not applicable checks are left out of the count. So the client site passed everything that applied to it. My site simply had one more thing that applied, and it passed that too.

That is an important point, because I don't want anyone to panic. A score of 2 of 2 is not a failing grade.

What llms.txt is, and what it is not

llms.txt is a small text file that sits at the root of your website, at yourwebsite.com/llms.txt. It was proposed in September 2024 by Jeremy Howard of Answer.AI, as a simple way to give AI systems a short, clean summary of a website and point them to the pages that matter most.

The reason is easy to understand. A normal web page is built for human eyes. The real information is wrapped in menus, buttons, images, scripts and footers. An AI system has to dig through all of that to find the useful part, and that costs time and money. llms.txt hands it the useful part directly.

You may already know two other files that live in the same place:

• robots.txt is like the security guard at the gate. It tells crawlers which areas they may visit.

• sitemap.xml is like the full directory of every room in the building.

• llms.txt is like the receptionist's one page brochure: who we are, what we do, and where to find the important things.

The file is written in Markdown, the simple formatting style that uses a hash sign for headings and dashes for lists. The format is simple. It starts with a heading carrying the name of the website, which is the only required part. Then a short summary. Then sections of links to the most important pages, each with a short note explaining what is there. Thousands of websites now publish one, including the developer documentation of OpenAI, Anthropic and Google's Gemini.

Now the honest part, the one many people selling "AI optimisation" leave out.

llms.txt does not help you rank on Google. Google's own guide to optimising for AI search lists llms.txt among the things you can ignore, because Google Search does not use these files. It says creating one will neither harm nor help your visibility in Google Search, including AI Overviews and AI Mode. Google's John Mueller has even compared llms.txt to the old keywords meta tag. If you read my post on technical SEO, you will remember that one: a tag Google stopped using long ago.

So if anyone offers to sell you "llms.txt for better Google rankings", you now know better.

Then why bother? Because ranking on Google is only one way people find you now. AI agents and assistants also visit websites to answer questions and complete tasks, and Chrome's own testing tool treats this file as part of being agent ready. Google's AI search guide itself says it is completely fine to create one for other services that use it. It takes about ten minutes, costs nothing, and gives any machine that does read it a clean, accurate summary of who you are, written by you.

What I put in my own llms.txt

When I added the file to my portfolio, I gave myself three rules.

One: describe the site, don't sell it. This file is read by a machine, not a customer. Words like "world class", "amazing" and "best in Nigeria" carry no information, so they don't belong. Plain facts do.

Two: answer the questions an assistant is likely to be asked. Who is this person? What do they build? What work have they done? How do you contact them?

Three: link to the pages that hold the details. The file is the map, not the whole territory.

Here is a simplified version of the structure I used, so you can see the shape of it:

# Chukwuma Ukaha

> Full stack developer based in Nigeria. Builds fast business websites,
> web applications and AI features using Next.js, React, Node.js,
> TypeScript and PostgreSQL.

This site contains case studies of client and personal projects, services for businesses and real estate companies, and articles about websites, performance and AI.

## Work

- [Case studies](https://chukwumaukaha.com/work): Four projects: a company website for an oil and gas services firm, a website for an online French school, a multi tenant lead capture platform, and an offline AI advisor for cassava farmers.

## Services

- [Websites for businesses](https://chukwumaukaha.com/business): What is included, how projects run and how to start.
- [Websites for real estate](https://chukwumaukaha.com/real-estate): Property websites with listings the client can update.

## About and contact

- [About](https://chukwumaukaha.com/about): Background and how I work.
- [Contact](https://chukwumaukaha.com/contact): Email and WhatsApp.

Notice what is not in there. No adjectives. No promises. No keyword stuffing. Just a clear summary and a short guided tour.

One more thing I learned. The llms.txt proposal suggests a simple test: give an AI assistant only your llms.txt file, then ask it questions about your business. If it can answer correctly, your file is doing its job. If it gets confused, your file is probably too vague.

The file is the easy part. Here is what matters more.

If you only add llms.txt and stop there, you have missed most of the point. The Chrome team's guide to building agent friendly websites explains that agents see your website in three main ways: screenshots, the raw HTML, and the accessibility tree. Here is what helps on each.

Use real buttons and real links. Some websites build buttons out of plain boxes styled to look like buttons. A human can't tell the difference, but an agent can't tell it is clickable. Chrome's guide recommends using proper button and link elements, because agents recognise them as things you can act on.

Give every form field a label. A box that only says "Type here" tells a human what to do from context. An agent needs the field to be labelled "Email address" or "Your message" in the code. The same goes for icon buttons, like a WhatsApp icon with no text. Give it a name, such as "Chat with us on WhatsApp", even if it is only visible to machines and screen readers.

Keep your layout still. Chrome's guide warns that a shifting layout confuses agents that work from screenshots. My oil and gas client's site and my own portfolio both have a layout shift score of 0, because every image declares its size before it loads. I explained how in my post on image optimisation.

Make sure your content is in the HTML. As I explained in my technical SEO post, a website that only builds its content with JavaScript arrives almost empty to some readers, and not every bot runs JavaScript. My website and my client sites are built in Next.js, so every page arrives as finished HTML.

Put important facts in text, not pictures. A price list saved as an image looks fine to a human but is much harder for a machine to read reliably. Your prices, hours, services and contact details should be real text on the page.

Don't hide things behind overlays. Big transparent layers over buttons, or popups that cover the page, can stop an agent from seeing what is underneath.

Here is the beautiful part. Every single one of these also makes your website better for humans, especially people who use screen readers or slow phones. As the Chrome team puts it, getting agent ready is really a reminder to build websites that are well structured, accessible and clear. If your Accessibility score in PageSpeed is already 100, you are most of the way there.

And looking ahead, a proposed standard called WebMCP will let websites tell agents directly what actions they offer, so an agent could "request a quote" or "book a class" without guessing which button to press. It is still experimental, but it shows where all of this is heading.

Step by step: make your website agent ready

1. Check where you stand. Run your home page through PageSpeed Insights. Look at the Accessibility score and, if it appears, the Agentic Browsing row. Write both down.
2. Fix the accessibility issues first. Open the Accessibility section of the report. Anything about buttons without names, form fields without labels or poor structure helps humans and agents at the same time. This is where most of the real value is.
3. Write your llms.txt. Open any plain text editor and follow the structure I showed above: your name or business name as the heading, a short factual summary, then a few sections of links with a short note on each. Leave out sales language.
4. Put it at the root of your website. If your site is built with Next.js, place the file in the public folder and it will be served at yourwebsite.com/llms.txt. If you use WordPress, popular SEO plugins like Yoast and All in One SEO can generate one for you. Wix generates one automatically for its websites.
5. Check that it loads. Visit yourwebsite.com/llms.txt in your browser. You should see your plain text, not an error page.
6. Test it with an AI assistant. Paste your llms.txt into an AI assistant and ask it what your business does, what you offer and how to contact you. If the answers are wrong or vague, rewrite the file.
7. Keep your layout stable. Make sure every image on your site has its width and height set, so nothing jumps while the page loads.
8. Move key facts out of images. Prices, opening hours and contact details should be real text.
9. Run PageSpeed Insights again and compare.
10. Update the file when your business changes. A new service, new prices, a new page. An outdated summary is worse than none.

Build for both visitors

Let me bring it all together.

If you want to rank on Google, including its AI answers, do good SEO: helpful content, a fast website, clean technical foundations. llms.txt will not change your Google rankings, and anyone who tells you otherwise is selling you something.

If you want AI agents to understand your business and act on it, make your website clear, accessible and stable, and add an llms.txt as a short, honest introduction. It is ten minutes of work, and it is now part of what Google's own testing tool looks for.

The businesses that will win the next few years are the ones whose websites work well for both kinds of visitors: the human scrolling on a phone, and the agent working on that human's behalf. The good news is that both want the same thing. A website that is fast, clear and honest.

My own website now scores 3 of 3, and adding an llms.txt to both of my client sites is next on my list.

If you would like me to check how ready your website is for AI agents, or build you a website that is ready for both kinds of visitors from day one, send me an email at hello@chukwumaukaha.com or a message on WhatsApp.