Before and after
One score is a reading. Two are a measurement. Put a page in each field and see what the edit actually changed — including the tells it introduced along the way.
Loading the comparison tool…
Which way the number moves
The delta is always the second URL minus the first. So a negative delta means the second page looks less vibe-coded, and a positive delta means it looks more so. That is the opposite of how most progress bars read, which is why the direction is written out in words beside every number on the page rather than left to an arrow.
When both URLs share a hostname the two sides are labelled Before and After, because that is almost always what you are doing: the same page, a week apart. When the hostnames differ they are labelled A and B, because “after” would be a claim about a page somebody else owns.
A worked example
Take a generated landing page that scores 97. It ships Tailwind from the CDN, a scaffold title, no viewport tag, and three placeholder images. You spend an afternoon on it: real title, real description, viewport tag, images replaced, the CDN script swapped for a build step.
Re-run it here and the comparison does not just say “97 → 34.” It lists Tailwind from a CDN, scaffold title and placeholder images under Fixed, and — the part worth having — it lists whatever the rewrite introduced. A hero rebuilt in one pass to replace the placeholder images very often arrives with a violet gradient and a pill badge that were not there before. The score fell by sixty points and two new tells appeared in the same edit. Only a diff shows you that.
Craft moves the same way in both directions. Adding a skip link and a canonical URL shows up under Craft gained. Losing your hand-written stylesheet to a bundler shows up under Craft lost, and it is worth knowing even when the total went the way you wanted.
What a diff of two heuristics is worth
Both readings come from the same deterministic pattern matcher over one HTML document each. No JavaScript is executed on either side, neither page is stored, and the comparison is not given a URL of its own — close the tab and it is gone. A lower score is evidence that fewer patterns fired, not proof that the page got better, and neither number says anything about who wrote it.
If a result looks wrong to you, it might be: submit it for human review. If you want the pre-launch list instead of the diff, the launch checklist covers the things a single score never will.