Give the address of a page. This fetches it, finds its images, and re-encodes each one to AVIF and WebP with the same encoder and settings a ZawCloud node uses — so the savings below are bytes that were actually produced, not an estimate.
The encoder and settings a node uses, applied to the bytes the page actually sends.
Each image is downloaded the way a current browser asks for it, so a site that already serves AVIF or WebP is compared against what it actually sends, not against a fallback JPEG. It is then decoded and encoded to AVIF and WebP at quality 80 with its metadata removed — the defaults a ZawCloud node applies — and the smaller of the two is reported.
When the page displays an image at a width it declares, and the file is more than twice that wide, it is also encoded at twice the displayed width. Twice covers a high-density screen; every pixel past that is downloaded and thrown away.
A result never claims a saving that was not produced: an image whose re-encoded version came out larger is reported as already as small as it gets.
It reads the HTML the server sends, without running it.
srcset candidate a given screen picks. The largest is audited — what a high-density desktop downloads.width. Without a browser there is no layout to measure.