Improve PageSpeed Insights Scores and Let AI Fix Them

To improve PageSpeed Insights scores, open the Claude or ChatGPT desktop app with your site files connected, run the free report at pagespeed.web.dev, read the Mobile tab first, then paste the prompt below and let it make the changes. Google grades every page on 5 things, twice. Getting the report was not the hard part. Working out which of the dozens of flagged items are worth doing is the hard part, and that is the job the model takes over.

Want the handoff and nothing else? Jump to the 4 steps.

Key Takeaways

• PageSpeed Insights is Google's free page test at pagespeed.web.dev. It grades every page on 5 things, twice, once for mobile and once for desktop, and the numbers are Google's own rather than a third party estimate.
• The 5 are Performance, Accessibility, Best Practices, SEO, and Agentic Browsing, the newest one, which checks whether an AI agent can read your page. Agentic Browsing is a diagnostic, not a Google ranking factor.
• Open the Claude or ChatGPT desktop app first. That is where the prompt gets pasted, and it has to be running on the computer your files are on. A browser tab cannot reach them.
• Then run the report and read the Mobile tab first. It scores lower and it is where the traffic is.
• The model can only make the fix if your website files sit in a folder it can open, or you connect it to your site. Without that it can describe the fix and nothing more.
• Running this on my own site moved Performance from 74 to 94 in one pass, in about 10 minutes.

What Is PageSpeed Insights?

PageSpeed Insights is Google's free page testing tool at pagespeed.web.dev. You paste a URL, it loads the page the way a phone would, and it hands back 5 results plus a list of what is holding each one down.

It runs on Lighthouse, the same auditing engine built into Chrome's developer tools, on Google's machines with the device and network speed fixed every time. Run Lighthouse yourself locally and it uses your laptop and your connection, so the 2 will not agree. That is expected.

Every page gets tested twice, once as mobile and once as desktop, and the 2 results are usually far apart.

Why This Is Worth Running

Start with who is asking. For most sites Google is how people arrive, which makes Google's read on your page the one that decides whether new people reach it. Very little of that read is published. The ranking system is not published in full, which is why so much SEO advice is inference.

This part is not inference. Google built a tool that opens your page, measures it against criteria it has written down, and hands you the result for free with no account. On the piece it is willing to be explicit about, it tells you what it sees.

It tests what you cannot see. Your site is fast on your laptop, on your connection, with everything cached. The report runs it as a first time visitor on a phone, which is the version most of your traffic actually gets.

It is Google's measurement rather than a stand in for it. The score comes from the same engine inside Chrome, run on Google's machines, so it is Google's own read of your page rather than an estimate of it.

Some of it counts toward ranking, and the rest earns its place anyway. Core Web Vitals, Google's name for a small set of loading and stability measurements taken from real visitors, is one of the signals its ranking systems use. Not the largest, and not the score itself. The other categories are not ranking signals and are still worth fixing: a slow page can lose people before your headline has loaded, accessibility failures shut people out, and Agentic Browsing is the same argument arriving early, as more people reach you through an AI rather than a search box.

The 5 Things Google Grades Your Page On

Each of the first 4 categories is scored 0 to 100. The fifth reports a pass ratio instead.

What counts as a good score?

Google colours the 4 scored categories in 3 bands. 90 to 100 is green and means the page is in good shape. 50 to 89 is amber and means there is real work available. 0 to 49 is red. Agentic Browsing sits outside this and reports a pass ratio instead.

Green is the target, not perfection. Chasing the last few points of Performance usually costs more than it returns.

Performance: how fast the page loads

This is usually the score that moves, and the one everything else gets blamed on. Large images, heavy scripts and slow servers all land here, and it is almost always the lowest of the 4. Accessibility and Best Practices, by contrast, tend to already sit close to green and are quick to fix when they do not.

SEO: the mechanical basics, not keywords

This category is not about keywords or content quality. It checks that a page has a title, a meta description, a viewport tag, and links a crawler can follow. Mechanical things a page either has or does not.

Agentic Browsing: whether an AI agent can use your page

The newest category, added to Lighthouse in 2026. It reports a pass ratio rather than a score, and it covers 4 areas:

• The accessibility tree. The structured outline of your page that screen readers use. An AI agent reads the same thing, so a page built for a screen reader is already most of the way there.
• Layout stability. Whether things jump around as the page loads, which breaks an agent trying to click one of them.
• llms.txt. A short file at the root of your domain telling an agent what your site is and where the useful pages are. More on it in the questions below.
• WebMCP tools. The newest of the 4, a way for a page to offer an agent real actions rather than making it click through the screen.

The denominator only counts the audits that apply to your page, so 3/3 on one site and 4/4 on another is normal and does not mean one is better.

It is a diagnostic category, not a Google ranking factor. Chrome's own team calls the category informational and unbenchmarked, so fixing it does not move your rankings today. It is worth doing because it is cheap now and it is unlikely to stay optional.

What You Need Before You Start

3 things, and the first is the one that decides whether any of this works.

1. Your website files somewhere the model can reach. Either a folder on your own machine you can point at, or your site connected to the app directly. If you are not sure, the test is whether you can open a folder and see files like index.html, or a project you or a developer built. A site you edit only through a website builder's admin screen does not count, and the model will not be able to change it.
2. The Claude or ChatGPT desktop app. The desktop version specifically. A browser tab has no access to your hard drive, so the web and mobile versions cannot do this part.
3. Browsing turned on, and not just the ability to fetch a URL. It needs to open the results page the way you would, because that page builds itself as it loads and the scores are not there until it does.

Improve PageSpeed Insights in 4 Steps

Step 1: Open the Claude or ChatGPT desktop app

Start here, before you touch the report. This is the window the prompt gets pasted into, and it is the one that can see your site files. In both apps the connection lives in settings, under file or folder access: point it at the folder your website lives in, and the app will ask what it is allowed to do with it. Approve that and you are set up.

Step 2: Run the report at pagespeed.web.dev

In your browser, go to pagespeed.web.dev, put in the page you want graded, and press Analyze. It comes back in well under a minute.

Step 3: Read the Mobile tab first

There are 2 tabs, Mobile and Desktop, and the tool opens on Mobile. Stay there. Mobile is where most of your traffic is, and it is the harsher test, so it is where your worst number lives. A page can sit at 95 on desktop and 60 on mobile.

Any of the 4 scores sitting amber or red is a candidate, and so is an Agentic Browsing ratio that is not full. That is the whole reading job. You do not need to understand the individual audit names, because the next step hands the sorting to something that reads faster than you do.

Step 4: Paste the prompt into the desktop app

Back in Claude or ChatGPT, fill in your URL and send it. The folder line is only needed if you have not already connected the project.

Act as a technical SEO and web performance specialist.

Website URL: [YOUR URL]
Local project folder: [FULL PATH TO YOUR SITE FILES, or skip this line if
the project is already connected]

Open pagespeed.web.dev in your browser, run a report on the URL above, and
read both the Mobile and the Desktop tab. Report the 5 category results for
each tab. If you cannot open the page, say so rather than estimating.

Then read my local folder and list every failing item you could fix in these
files, ordered by how much score it moves for the effort it costs.

Do not edit anything yet, and do not tell me a fix will move my Google
rankings. Show me the list and wait for me to pick.

You are not pasting the report or sending screenshots. It gets the scores itself, which is why Steps 2 and 3 are you looking rather than copying.

3 things in there are doing more work than they look.

• "Open pagespeed.web.dev in your browser" is the load bearing one. That page builds itself as it loads, so a tool that only fetches a URL gets an empty shell and is one step away from filling in scores it never read. A tool that opens a real browser tab reads your actual numbers. That is what separates a real report from an invented one.
• "Say so rather than estimating" is the backstop for when it cannot.
• The hold at the end keeps it from editing before you have seen the list.

Rules like these belong in a project file rather than being retyped, which is what project instructions are for.

What comes back, and what to do with it

You get a list of concrete changes rather than audit names. Pick the ones you want, say so, and let it edit the files. Then have it run the report again and show you the categories, so the score move is measured rather than assumed.

Worth keeping the 2 apart: this works because the model has your files, not because it drives your screen. Astra computer use is the other one, where the job lives on a screen and there is no code to hand over.

What This Looked Like on My Own Site

I ran this on withpt.ai.

Mobile Performance came back at 74. The other 3 scores were already green and Agentic Browsing was passing, so there was one number to move.

I gave it the URL and told it to fix what it found, which skips the list and pick step above. With one number to move and my own site to break, that was a fair trade. On a client site, or with more than one category in the red, I would want the list first. Performance went from 74 to 94 in that pass, in about 10 minutes. What moved was page weight: the images and scripts the page was pulling before it showed anyone anything.

What surprised me was how little I had to supply. No report, no screenshots, no explaining which items mattered. It pulled the scores, ran its own audit against the pages to see what was slow, made the changes in the files, ran the test again to confirm the number had moved, checked the fixes held, then pushed.

That is the full setup guide: the click path for connecting a folder in each app, an 8 step order to work through, and a longer prompt that asks for both tabs as one table. Enter your email in the card below to get it immediately.

Where This Stops Working

The model cannot fix a site it cannot reach

Diagnosing the problem works from the URL alone. Changing anything needs access to the code, which is a setup step rather than a prompt.

A green score is not a ranking

The lab score is not the thing Google ranks, and a 100 on a page nobody wants to read is still a page nobody wants to read.

Field data takes time to catch up

The score you move immediately is a single lab run. Above it sits real Chrome user data, collected over a 28 day window, so that half catches up roughly a month later. It only appears once Google has enough visitors to your page, so a quieter site may see no field section at all, and nothing is wrong.

How Often to Run It

Once a month, plus any time you change something structural: a redesign, a new page template, a new tracking script. The check itself costs you under a minute. It is skipping it for a year that costs you.

If you want a hand building this kind of check into how your business actually runs, that is what the workshop is for.