How words are counted, and what the share means
A word is what the browser’s word segmenter for the selected language finds (Unicode UAX #29, with a dictionary for Chinese, Japanese and Thai). Hyphenated words and contractions such as don’t stay one word, and numbers count as words unless you untick Count numbers. With Ignore upper/lower case on, The and the are one entry, lower-cased by the rules of the text language. In Turkish, İstanbul and ISTANBUL become two different words, istanbul and ıstanbul, because a capital I lower-cases to the dotless ı.
In Japanese, 東京都に住んでいます。東京は大きい。 gives 東京 twice and 都, に, 住, んで, い, ます, は and 大きい once each; a counter that splits only at spaces counts that whole sentence as one word. The dictionary splits verb endings too (住んでいます becomes four pieces), so the top of a Japanese list is usually particles such as に and は, and Leave out common words hides them. In our test run, 1,000,000 Japanese characters took about 3.4 seconds and 1,012,000 characters of English 0.23 seconds, because dictionary segmentation is slower.
The share is a term’s count divided by all terms counted with the current settings. Leaving out common words therefore raises the share of the remaining words, and a phrase’s share is measured against all phrases of the same length. We do not give a target keyword density: search engines have not published one, and a share tells you how repetitive a text is, not how it will rank. Phrases are counted inside a sentence only, so the last word of one sentence and the first of the next never form a phrase. With common words left out, a phrase may not start or end with one: quick fox counts, the quick does not.
Doubled words are the same word twice in a row with only spaces or a line break between them, a typo that is easy to miss when one line ends with to and the next starts with to. Numbers are skipped, because 1 1 is often meant. Some doubles are correct (that that, had had), so the page lists each one with its line number before you choose Fix doubled words. The text result can also be every word once, in order of first use, which is what most people mean by removing duplicate words.