Keyword Extractor

Extract significant keywords and phrases from text content using TF-IDF and statistical analysis algorithms.

Tested tool guide Text and writing tools Checked August 16, 2026

What Keyword Extractor does and how it behaves

Paste a document and this tool returns the words and phrases that carry its content. Each term receives two opposing measures: how often it occurs (term frequency), and how concentrated it is, since a word appearing in only one or two sentences counts as more distinctive than one spread everywhere (inverse document frequency). The two multiply into one score and the highest-scoring terms rank first. The surprise: frequency alone does not win. A topic word present in nearly every sentence is deliberately demoted, so the top of the list often excludes the word you expected. Every result is a literal word from your text, never a synonym.

How the result is produced

Scoring pipeline

The text is split into tokens, lowercased, and common stop words (the, and, of) are removed. Each surviving term receives a term frequency, its count of occurrences, and an inverse document frequency, the log of the total number of documents divided by the number containing it, where a document is usually one sentence or paragraph of your input. The product of the two is the term's score.

Phrases and normalization

Multi-word results are typically assembled from adjacent terms: when two or three individually strong words sit side by side, the tool reports the combination as a phrase. After scoring, values are usually normalized to a common scale, so numbers are easy to scan but every score is relative to the strongest result in your text. The computation runs in the browser, so scoring uses only what you pasted; nothing is uploaded.

Good uses

Check a draft before publishing: compare the top-scoring terms with the topic you intended. A page meant to be about solar panels whose top keywords are 'installation' and 'cost' shows where the emphasis actually landed.
Triage long source material: after an interview or a pile of research notes, extract keywords to recall what each text was about and to group or file them without rereading everything.
Generate candidate metadata: use the top-scoring terms and phrases as tags, index terms, or internal labels, reviewing the suggestions first since the scoring has no idea what the document is about in any human sense.

Limits and checks

The expected topic word often ranks low or vanishes entirely. A subject repeated in nearly every sentence has a document frequency close to the corpus size, which pushes its inverse document frequency toward zero, so the most obvious keyword can finish below one-off terms.
The list is lexical, not semantic. 'Run', 'runs', and 'running' score as three separate terms, synonyms only appear if your text contains them, and no result reflects meaning. Words the stop-word list removes, such as 'well' or 'south', never appear no matter how important.
Scores are relative, not comparable. A 0.2 does not mean half as important as a 0.4, and the same text scored against a different corpus definition or stop-word list returns different numbers, sometimes a different order. In short inputs, many terms occur exactly once and tie at the same score.

Common questions

Why is the word I expected at the top missing from the results?

Three usual causes. It may be on the stop-word list and removed before scoring. It may appear in so many sentences that its inverse document frequency is near zero, pushing it below the cutoff. Or the text uses a different form of the word, since 'jaguar' and 'jaguars' are scored separately. If the tool shows what it removed, that usually settles it.

Can I compare scores across two different texts?

Not meaningfully. Scores are normalized within each run, so a 0.3 in one document and a 0.3 in another do not measure the same thing. Use the ranking to judge emphasis inside one document, and for comparing documents look at how much the ranked lists overlap rather than at the numbers.

References and verification

The behavioral notes were checked against the browser implementation. Standards and primary references below define the relevant format, formula, or platform behavior.