Word frequency and keyword density: use counts to review a draft

If a draft feels repetitive, a frequency list can help you find the words and phrases to inspect. Paste the text into the word frequency tool, look at the highest counts, then return to those passages. A frequent word might be a writing habit, necessary terminology or part of a quotation.

The count identifies something to review. It does not decide whether to rewrite it, and its percentage is not a search-ranking score.

Count and share answer different questions

Count tells you how many times a term occurs. Share expresses that count as a percentage of the terms counted under the current settings.

Try this short example with the default single-word settings:

The cat and the cat sleep.

There are six word tokens. With case ignored, cat occurs twice, giving it a share of 33.33%. the also occurs twice. and and sleep each occur once.

Turn on Exclude common English stop words and only cat, cat and sleep remain in the counted stream. The count for cat is still two, but its share becomes 66.67%. The text did not change; the denominator did.

Know what the tool calls a word

The current tokenizer matches runs of Unicode letters and numbers, with internal straight or curly apostrophes and hyphens. This means:

• don't and well-known each form one token;
• Cat and cat combine when Ignore case is enabled;
• standalone numeric tokens such as 2026 are omitted from the frequency table by default, while a mixed token such as A1 remains;
• café and cafe are different spellings, not automatically combined.

The summary’s Total words counts the raw tokens before these filters. It is not always the denominator for the Share column. Unique words is also a raw, case-folded summary and does not follow every table setting.

There are language limits. The tokenizer does not perform dictionary-based segmentation for unspaced Chinese or Japanese text, and it does not normalise Unicode combining marks. Visually similar accented text can therefore split or count differently depending on how it is encoded. Unicode-aware character matching is not the same as accurate linguistic word counting in every language.

Filters do not all affect percentages in the same way

Removing numbers, stop words or custom ignored words changes the stream being counted. Minimum word length also removes terms before their shares are calculated.

Minimum count works later: it hides low-frequency rows without removing their occurrences from the denominator. In the six-word example, setting minimum count to two leaves only cat and the visible, still at 33.33% each. The visible shares do not have to add to 100%.

Searching within results hides table rows too. It does not recalculate their shares, and copy/CSV/JSON exports use the full result list rather than only the rows matching that search. Record the settings when comparing two drafts or sharing a table.

Stop words can carry meaning

The built-in stop-word list is English. It removes words such as the, and, of and also not. That can make a topic list easier to scan, but it can also hide a meaningful distinction in a quotation or response.

For an initial writing review, compare the list with stop words included and excluded. Use Extra words to ignore for known boilerplate, such as a repeated brand name, only when removing it suits the question you are asking. These entries are individual words, not multiword phrases.

Phrase counts need a check against the original text

The tool can count two- and three-word sequences. This can reveal repeated phrasing that a single-word list does not show.

It builds those sequences after removing stop words and other token filters. For cup of tea, excluding stop words creates the two-word result cup tea, even though that phrase was not written consecutively. In the cat example, filtering first also creates cat cat.

Sequences can cross punctuation and sentence boundaries. With cup of tea. tea time and stop words excluded, the results include cup tea, tea tea and tea time. Those are adjacent entries in the filtered stream, not necessarily phrases a person wrote.

In phrase mode, Share is a proportion of counted phrase windows, not of the document’s raw words. Leave filtering off and inspect the source when your question is about literal repeated wording; sentence-boundary crossings still need checking.

Keyword density does not supply an editing target

Google’s spam policies describe keyword stuffing as repetition intended to manipulate rankings, including wording that becomes unnatural. They do not provide an ideal keyword-density percentage to aim for.

A local frequency table cannot reveal the weight of a term in a search engine’s ranking system or predict a ranking change. Claims that a particular percentage guarantees improvement—or that frequency can never matter in retrieval—go beyond what this tool can establish.

Use the words that identify the subject clearly and answer the reader’s question. Do not add or delete a necessary term merely to reach a target share. Consistent repeated terminology can be helpful in instructions.

Turn a result into a useful edit

Choose a few unexpectedly frequent terms and read the passages containing them. Remove repetition where it adds nothing, keep it where it prevents ambiguity, and check whether quotations or pasted boilerplate explain a high count.

For before-and-after comparisons, use the same text boundaries and settings. The table can document a change in wording or emphasis; judging accuracy, clarity and usefulness still requires reading the draft.

Sources and further reading

• Unicode Standard Annex #29 — text segmentation; broader than this tool’s regex tokenizer
• Google Search Central — helpful, reliable, people-first content
• Manning, Raghavan & Schütze, Introduction to Information Retrieval — term frequency and stop words

Last reviewed 2 October 2026.