Words are counted the way a word processor counts them: runs of characters separated by whitespace. A hyphenated compound counts as one. A number counts as one. Punctuation attached to a word does not add to the count.
The deduction is taken from the text you submit, not from the output. A rewrite that comes back slightly longer than the original does not cost more than what you pasted.
Only text you actually process counts. Typing into the editor costs nothing; the deduction happens when you run a rewrite or a scan. Clearing the field before running costs nothing at all.
Running the same passage twice costs twice. If you are experimenting with tones on a long document, test on a representative section rather than the whole piece — a 300-word sample tells you most of what a tone will do to 3,000 words.
The detector counts separately and identically, drawn from the detection allowance rather than the humanizer one.
If a run fails because of an error on our side, it should not be deducted. If you see an allowance drop against a run that returned nothing, tell us and we will correct it — that is a fault, not a usage question.
Markdown, HTML tags and code inside your submission count as words too, because they are whitespace-separated text. If you are processing a document with a lot of markup, stripping it first will make your allowance go noticeably further — and the output is usually better without markup in the way.