The measure we run on ourselves
This is not a generic similarity score. It is the specific method our own build uses to stop near-duplicate pages shipping: chop both texts into every possible run of eight words, count how many runs appear in both, and divide by the number appearing in either. When two pages in the same family cross a threshold, the build fails and the page does not go out.
That is the whole reason it is worth publishing. Most similarity tools cannot tell you what they compute or why the number is what it is. We can describe this one exactly, because we would notice immediately if it were wrong.
What eight-word runs catch, and what they miss
They catch reuse. A sentence carried from one document into another shares long runs immediately, and the tool shows you the longest shared passages rather than only the number, so you can see what drove it.
They miss meaning entirely. Two documents can make the same argument, cite the same sources and reach the same conclusion in different words, and score near zero. That is not a defect to be apologised for — it is the honest boundary of the method, and it is the same boundary that makes a similarity score in a plagiarism report drop when a passage is reworded. The score fell because the method measures strings. Whether anything was borrowed is a separate question, covered in whether paraphrasing removes plagiarism.
Reasonable uses
Checking whether two of your own pages overlap enough to compete with each other. Comparing a draft against the source you were working from, to see how much survived verbatim. Checking whether a rewrite actually changed anything, which is the use that surprises people most — a rewrite that feels thorough often shares more with the original than expected.
For revisions of a single document, the diff checker is the better instrument. If what you need is whether a quotation matches its source word for word, the quote checker answers that question directly, including where the two diverge.