The Agentic Browsing checks in PageSpeed Insights (and how to pass them)

If you have run PageSpeed Insights lately and found an unfamiliar fifth score sitting beside Performance and SEO, that is Agentic Browsing, and it is not measuring anything about your human visitors. It grades how ready your page is for an AI agent to read and operate. PSI inherited it from Lighthouse in May 2026, it runs a specific three checks (not the full four), and it can hand you a different number each time you run it. This guide is the PSI-specific version: what the three checks are, how to pass each, and why your ratio moves.

01 · The new scoreWhat you are actually looking at

Around 21 May 2026, PageSpeed Insights updated to Lighthouse 13.3 and the Agentic Browsing category came with it. Two things surprise people. First, it is not a score out of 100; it is a pass ratio, usually shown as 3/3, because the underlying standards are new and Google is gathering data rather than grading you. Second, it is explicitly not a ranking factor. Nothing here moves you in Google Search. It is a readiness read-out, not an SEO lever, and treating it as the latter will only frustrate you.

For the full history and the scoring rationale, the Lighthouse category guide is the reference. This page is about the version you actually see in PSI.

02 · What it measuresThe three checks PSI runs

Lighthouse defines four agentic audits. PSI runs three of them:

The missing fourth is WebMCP validation. It needs the WebMCP origin trial, which PSI has no way to enable for your site, so it simply does not run here.

03 · The fixesHow to pass each check

Accessibility tree. Agents locate controls by their accessible name and role, the same way a screen reader does. Give every interactive element a name: visible text inside a real <button> or <a>, an associated <label> for inputs, or an aria-label where there is no visible text (icon buttons are the usual culprit). Prefer semantic elements over <div onclick>. If your Lighthouse Accessibility score is already high, this check mostly passes for free.

Layout stability. This is Cumulative Layout Shift, judged for agents rather than humans. The fixes are the standard ones: set explicit width and height (or aspect-ratio) on images and video, reserve space for ads and embeds so nothing reflows when they arrive, preload web fonts to avoid a late swap, and never insert content above what is already on screen. Aim for CLS under 0.1.

llms.txt. The audit wants a file at your root with three properties: 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)

Note the links are real Markdown links, not bare URLs; a plain https://acme.com will fail the check. Whether llms.txt does anything beyond passing this audit is a genuinely open question, covered honestly in do AI crawlers actually read llms.txt. For PSI, “present and well-formed” is all you need.

04 · The flickerWhy your ratio keeps moving

A pass ratio that reads 3/3 on one run and 2/3 on the next is not a glitch in your site so much as the nature of the checks.

• CLS is variable by design. It measures what shifts as the page loads, and that depends on network timing, which ad filled the slot, whether a font swapped. A borderline layout will tip over the threshold on some runs and not others.
• The category is experimental. Google is still tuning thresholds and audits between Lighthouse releases, and PSI adopts them on its own schedule, so the same page can score differently across a version bump.
• The llms.txt check has had false flags. On some builds it has failed valid files. If that check flickers while your file is unchanged, read its exact message before touching anything.

The practical response: fix the structural causes (dimensions, reserved space) so you are comfortably inside the threshold, not balanced on it, and do not chase a single flaky run.

05 · Keeping it in proportionWhat it means for your launch

Two of the three checks (accessibility tree, layout stability) are real quality work that pays off regardless of whether a single agent ever visits: they help your human visitors, your Performance and Accessibility scores, and your conversion. Do those for their own sake. The llms.txt check is cheap enough to just clear. But keep the whole score in proportion: it is a readiness signal, it does not touch your ranking, and being usable by an agent is not the same as being found by one. Passing 3/3 gets you over the technical floor; getting an agent to actually arrive is the Footprint problem, which no PSI check measures.