Why your page is slow,
and whether your fix helped.
Speed scores move on their own. When the number goes up after you change something, no tool currently tells you whether you earned that or just got lucky.
The same page, twice, nothing changed in between
measured, not an exampleOne measurement did all of it. How much the page jumped about went from 25 of 25 to 15 of 25 on its own. Ship a real fix in between two runs like these and you cannot tell your work from the wobble.
The problem
The report is mostly things that cannot earn you anything.
Open the performance section and you get a long list. Almost none of it is scored, and nothing tells you which parts feed the bits that are.
Items in the performance section that carry any marks. The other 44 are advice, and nothing says which of them will move one of the five.
What one real site was already scoring on the measurement its entire list of image fixes fed into. Doing all of them would have earned close to nothing.
Points between two runs of the same untouched page. Bigger than a lot of fixes are worth, which is what makes the score hard to trust.
Only five things are scored.
Unused JavaScript, render blocking, image delivery. All useful, all worth zero marks on their own. They only count if fixing one happens to speed up one of the five that do, and no report draws that line for you.
Every square here is one item in the performance section. The amber ones are the only ones with marks attached.
5 scored · 44 not
And a slow page earns almost nothing for getting slightly faster.
The scale bends. Going from 7.3 seconds to 6 seconds is a real improvement your visitors would feel, and it is worth about two points. You can work all afternoon and watch the score refuse to move.
Marks out of 25 for how long the biggest thing takes to appear.
22.5 marks at 2.5s · 1.3 marks at 7.3s
What I am building
Measure it enough times to know.
Four things, in order. Whether all four are possible is what I am finding out, and I post the ones that fail as well as the ones that work.
Record the whole page
Every request, every trace event, and the file and line of the script that asked for each one. Not a summary of the page. The whole thing.
Name the actual cause
Not "reduce unused JavaScript" but the file, the line, and the chain of requests that led to it being there.
Say what it is worth first
Some fixes cannot earn points, because the measurement they feed is already at the flat end of its scale. Worth knowing before you spend the afternoon.
Run it until the answer is real
Before and after, repeatedly, until the difference is bigger than the wobble. If it is not bigger, it says so instead of taking the credit.
The tools that already exist
All three are good. None of them close this gap.
Not a pitch against them. I use all three, and this is built on the same browser they use. Here is what each does well and the specific thing it leaves you with.
Lighthouse
Reads your page in a real browser and grades it against fixed scales. It has got much better and it now names files. It will tell you which image is oversized.
What it leaves youIt grades one run. Run it again and the number is different, and it has no opinion on whether that difference means anything.
PageSpeed Insights
Lighthouse on Google's machines, plus real visitor data from Chrome. The visitor half is the part worth having, because it is what search reacts to.
What it leaves youThat half takes 28 days to respond to anything you ship. Neither half answers whether this morning's change worked.
Agentic IDEs
Genuinely good at the fix. Point one at a report and it writes the lazy loading, resizes the image, splits the bundle. Faster than doing it by hand.
What it leaves youIt reads the same report, so it picks the biggest number rather than the winnable one, and it checks its work by running once.
Questions
The obvious objections.
Is this not just Lighthouse with a nicer front end?
It runs the same browser and reads the same trace, so the raw numbers agree. The difference is what happens next. Lighthouse grades one run and stops. This runs the page repeatedly, works out how much the numbers move on their own, and only says a fix worked when the change is bigger than that.
If you only need a grade, Lighthouse is free and already installed. Use it.
Can I not paste the report into Cursor and let it fix things?
For the fix itself, often yes, and quicker than you would. The problem is upstream. The agent reads the same report you do, so it picks whatever number looks worst rather than whatever is winnable, and it confirms its own work by running the test once more.
Running it once more is the part that does not hold up. That is what this is for.
What will it cost?
Not decided, and I am not going to invent a price to look further along than I am. There will be a free version that takes a URL, because that is how anyone will try it in the first place.
When can I use it?
No date. I have built the recording layer and it does not do anything useful yet. Putting a launch date on the internet before the hard part works is how you end up shipping something that does not.
The build is public, so you can watch it get there or watch it fail.
Does it work on any site?
Anything a browser can load without signing in, to start with. Pages behind a login, and anything where the slow part only happens after a click, are harder and come later.
It will not work on everything, and when it cannot measure something it says so rather than guessing.
Who is building it?
One person. I write about it twice a day, including the days it goes badly. The whole thing is on X.
Not a waitlist
Get the scoring guide.
I spent a weekend working out exactly how the performance score is calculated, because every explanation I found was either three sentences or the source code. Which five things carry marks, how a timing turns into points, and why some fixes cannot win you anything.