Image CDN

Generating every size of every image at build time, which is slow and produces files most visitors never request.

Deciding which formats to support, since the service negotiates on what each client accepts.

Re-processing an entire library when a new format becomes worth adopting.

Handling the originals editors upload, which vary in size, shape and format in ways no build step anticipates.

Which is why it is usually the cheaper answer than converting files by hand, especially on a content-managed site.

Parameters in the URL — a width, a quality, a format — or a path convention encoding the same.

Which means the requested variant is visible in the markup and auditable by reading the page.

It also means anyone can request any variant, so services generally restrict the parameter space or sign the URLs.

An unrestricted service is a generation endpoint anyone can drive, which is a cost question before it is a security one.

Each distinct parameter combination is its own cache entry, so a page requesting slightly different widths per breakpoint multiplies the variants.

A cold variant is generated on demand, which is slower than serving a stored file — and the first visitor to each combination pays that.

Which argues for a small fixed set of widths reused across the site rather than a width computed per layout.

Negotiated formats multiply the count again, though that multiplication is one the service is built for.

Choosing the right rendered size, which is the srcset and sizes decision and stays in your markup.

Lazy loading and fetch priority, which are attributes on the element rather than properties of delivery.

Alt text, which is content.

So a site can use an image service and still serve images at three times the needed width, because the service delivered exactly what was asked for.

Whether images are delivered through a dedicated image service is detected from the hosts and URL shapes in the markup.

The format and weight really delivered are measured regardless, since that is what a visitor receives.

Oversized files are reported the same way whether or not a service produced them.

Which transformations are available, and how the service is configured, are not observable from the responses a scan receives.