Keyword density β and why counting words is not the same in every language
Keyword density is the share of a text made up by one word. In English you split on spaces and you are done. In Korean the same word carries a different ending in every sentence, so a word used five times is counted as five different words used once each.
There are two reasons to look at density. One is to check what your own text is actually about: if the top words in a restaurant review are βandβ and βtheβ, the subject got lost. The other is the opposite β to see whether one word has been pushed in too often.
But in some languages the counting itself is the hard part, because words do not arrive in the form a dictionary would list them in.
Particles scatter the count
Korean attaches a particle to the end of a noun to mark its role in the sentence. μμΈ (Seoul) appears as μμΈμ΄, μμΈμ, μμΈμ, μμΈμμ and more, depending on what it is doing. A tool that splits on spaces sees five different words, each used once, and none of them reaches the top of the list.
This tool strips the particles first, longest first β try the short ones first and μμΈμμ becomes μμΈμ rather than μμΈ. The result is one word used five times.
The same class of problem exists in other languages this tool sees: an agglutinative or heavily inflected language will always undercount if it is tokenized as though it were English.
What percentage should you aim for?
There is no target. A figure of two to three percent circulated widely for years, but no search engine has ever published such a threshold. Google's stated position is the opposite one: stuffing a keyword in repeatedly is the problem, and text should read naturally.
So the number is better used as a check than as a goal. Above roughly ten percent a word usually reads as forced; and if the subject of the piece is not in the top ten at all, the writing is probably unfocused.
Short and very common words are ranked out
Function words β the, in, of, and their equivalents β top every text ever written. Leaving them in the ranking pushes out the words worth seeing.
They are therefore excluded from the ranking but still counted in the total. They are genuinely in the text, and removing them from the denominator would inflate every density figure. Only words of two characters or more are ranked.
Japanese and Chinese
Neither language puts spaces between words, so there is no visible place to cut. Doing it properly needs morphological analysis, which needs a dictionary and a model β not something to run in a browser tab.
This tool cuts where the script changes instead, treating a run of kanji, a run of katakana, and a run of hiragana each as one token. ζ±δΊ¬γ«γ‘γ© splits into ζ±δΊ¬ and γ«γ‘γ©. It is not exact, but it finds nouns reliably enough to be useful, and hiragana-only runs β mostly particles and endings β are kept out of the ranking.
Frequently asked questions
Can I see two- and three-word phrases?
Yes, they are counted separately. Real search queries are usually phrases, so the two-word list is often the more useful one.
The CSV export opens as gibberish in Excel.
Not any more. The file now carries a UTF-8 byte order mark, which is what Excel needs in order to read non-Latin text correctly.
Different verb forms are counted separately.
Yes. Particles are stripped, verb endings are not. Removing an ending requires identifying the stem, which goes wrong often without a dictionary, so the words are left as written.
Is my text sent to a server?
No. All the counting happens in your browser.