A checklist to pass the Agentic Browsing audit
The explainers tell you what the Agentic Browsing checks are. This is the page that tells you what to do about them, in order of payoff, from a two-minute scan to the one step most sites should skip.
Last updated July 12, 2026To pass Google’s Agentic Browsing audit, work in this order: run it (PageSpeed Insights or the Lighthouse CLI), fix the accessibility tree (name and label your controls), fix layout shift (reserve space so nothing jumps), ship a valid llms.txt, and only then, if agents are a real channel, add WebMCP tools. The first four are cheap and help humans too; the last is optional.
- Step 0: run the audit to get your baseline pass ratio (3/3 in PageSpeed Insights).
- Steps 1-2 (accessibility tree, layout shift) are real UX work that also lifts your Accessibility and Performance scores. Do them first.
- Step 3 (
llms.txt) is a fifteen-minute win that clears its check. - Step 4 (WebMCP) needs an origin trial and helps no human. Skip it unless agents are a genuine channel.
The Lighthouse and PageSpeed Insights guides explain what the Agentic Browsing checks are and why they exist. This is the other page: the ordered, do-this-then-that checklist for actually passing them. The trick is sequence. Two of the checks are real UX work that pays off for humans regardless, one is a fifteen-minute freebie, and one is optional and most sites should skip it. Work them in order of payoff and you clear the audit without wasting effort on the part that helps no one.
00 · BaselineStep 0: run it and read the ratio
Before fixing anything, measure. Two ways to run it:
npm install -g lighthouse@latest
lighthouse https://yoursite.com --viewOr open pagespeed.web.dev, enter your URL, and read the Agentic Browsing section. Remember the denominator: PageSpeed Insights runs three checks (so you are aiming for 3/3), while the Lighthouse CLI on Chrome 150+ with the WebMCP origin trial shows a fourth. Note which specific checks are red; that is your to-do list, and the rest of this page is in the order you should tackle them.
run by default in PageSpeed Insights: the accessibility tree, layout stability, and llms.txt. WebMCP is a fourth, but it needs Chrome 150 and the origin trial, so for most sites “passing” means 3/3.
01 · Highest payoffStep 1: fix the accessibility tree
Do this first, because it is the check that most often fails and the one that most helps everyone else. Agents locate and operate controls through the accessibility tree, the same structured model a screen reader uses, so an unnamed control is invisible to them.
The core fixes:
- Use real semantic elements:
<button>,<a>,<nav>,<main>, real headings, not<div>with a click handler. - Give every interactive element an accessible name: visible text, an associated
<label>for inputs, or anaria-labelon icon-only buttons. - Do not bury real content behind
aria-hidden, and keep the heading order sane.
If your Lighthouse Accessibility score is already strong, you likely pass this for free. The full reasoning and a deeper fix list are in your accessibility tree is now your agent API.
02 · Stop the jumpStep 2: fix layout stability
Next, stop your page moving after it loads. An agent computes where a control is and clicks there; if your layout shifts in between, it clicks the wrong thing. This is the same Cumulative Layout Shift metric from the Performance category, judged for machines.
The standard fixes:
- Set explicit
widthandheight(oraspect-ratio) on images, video, and iframes. - Reserve space for anything that loads late: ads, embeds, banners, cookie notices.
- Preload web fonts and use
font-displaysensibly so text does not reflow on swap. - Never inject content above content that is already on screen.
Aim for CLS under 0.1. The agent-specific reasoning and the full fix set are in why layout shift breaks AI agents.
The accessibility tree and layout-stability fixes improve your site for human visitors, your conversion, and your Lighthouse Accessibility and Performance scores. You would want them even if no agent ever visited. That is why they come first: the payoff does not depend on the agent era arriving on schedule.
03 · The fifteen-minute winStep 3: ship a valid llms.txt
Now clear the cheap one. The audit wants a file at your root with an H1, at least one Markdown-formatted link, and more than 50 characters. The smallest file that passes:
# Acme
> Acme turns messy invoices into clean spreadsheets in one click.
## Docs
- [Getting started](https://acme.com/docs/start)
- [Pricing](https://acme.com/pricing)Use real Markdown links, not bare URLs, or the check fails. Whether llms.txt does anything beyond passing this audit is a separate, honest question covered in do AI crawlers actually read llms.txt; for the checklist, present and well-formed is all you need.
Want your agent-readiness gaps found for you?
Nilkick checks your llms.txt, crawler rules, structured data, and WebMCP tool registration, then lists what is missing in plain language, so you know exactly what to fix before you re-run the audit.
04 · Only if agents are your channelStep 4 (optional): add WebMCP tools
This is the step most sites should skip. WebMCP lets you hand agents explicit tools instead of making them guess, but it needs the origin trial, does nothing for human visitors, and does not even run in PageSpeed Insights. Only reach for it if AI agents are a genuine channel for your product (commerce, an API, a tool). If they are, start with the declarative path: annotate your one or two most important forms (search, add-to-cart, book) and leave the rest.
05 · Re-run and stopWhat "done" looks like
Re-run the audit. In PageSpeed Insights, 3/3 is a full pass and the right place to stop for most sites. Do not chase a single flaky reading: if the ratio wobbles while your code has not changed, it is a borderline layout shift, so widen your reserved space rather than re-running twenty times. And keep the whole thing in proportion. Passing this audit means an agent can use your site; it does not mean an agent will visit, which is the Footprint problem no audit measures.
Common questions
llms.txt), so 3/3 is a clean pass there. The fourth check, WebMCP, only appears in the Lighthouse CLI on Chrome 150+ with the origin trial enabled.llms.txt. It takes about fifteen minutes and needs no code: add a file at your root with an H1, a one-line summary, and at least one Markdown-formatted link. The accessibility-tree and layout-shift checks usually take longer because they touch your actual pages, but they also pay off for human visitors.llms.txt check, is not a Search ranking factor. Do the work because two of the four checks (accessibility, layout stability) genuinely improve your site for humans, and because being usable by agents is worth clearing as a floor, not because it moves you up a results page.llms.txt, and occasionally on its own because the layout-shift check varies run-to-run. If your ratio flickers between 2/3 and 3/3 with no change on your end, it is almost always a borderline CLS, so tighten your reserved space rather than chasing a single reading.Get your free launch-readiness score
See what else is between your product and its first real users. Nilkick scores your readiness and hands you the map. Free, no login.
https:// optional · no account · we don't email you
Lighthouse's new Agentic Browsing category, explained
A field guide to Google's Lighthouse Agentic Browsing category: the four audits, why it scores a pass ratio not 0-100, and the Chrome 150 caveats.
Your accessibility tree is now your agent API
AI agents read your accessibility tree, not your pixels, so an unlabelled button is invisible to them. How the tree works, and the fixes that make you usable.
Why layout shift breaks AI agents (CLS beyond humans)
An AI agent clicks where a control was a moment ago; if your layout shifts, it misses. Why CLS is in the Agentic Browsing audit, and how to hold your page still.