Audit text metrics before publishing or comparing drafts

Problem

A request such as “check the keyword density,” “make this a grade-eight article,” or “tell me whether this review is negative” omits the choices that create the number. Characters may mean bytes, code points, or grapheme clusters; words depend on segmentation and case; keyword percentages use incompatible denominators; English readability formulas do not transfer to Korean or Japanese; a lexicon match is not emotion or intent; and normalization can erase a difference that matters. Without a written contract, two correct tools can disagree and an attractive score can replace the question that prompted it.

When to use this

• An editor wants to reduce repetition or compare terminology across two bounded drafts.
• A content review asks for keyword density but has not named token rules, phrase length, overlap, or denominator.
• A team wants an English readability baseline while understanding that formulas do not test comprehension or factual quality.
• Support or research staff need a transparent first pass over obvious English positive and negative words without claiming an AI emotion model.
• Two Unicode strings look alike or different and the decision depends on whether normalization, case, grapheme edits, or word-set overlap is the relevant identity.

Steps

1. Step 1

   Write the decision before choosing the metric

   State what action could follow the result: revise a repeated term, verify a target phrase, compare two English versions, queue comments for human review, or detect whether a normalization policy explains a mismatch. Name the audience and the cost of a false conclusion. If no decision would change, avoid manufacturing a score merely because a tool is available.

2. Step 2

   Preserve and sanitize the source

   Keep an untouched source, a stable version label, and enough context to reproduce the sample. Remove secrets and personal data from the working copy. Local browser processing prevents this tool from uploading text, but copied reports, screenshots, downloads, browser extensions, and later sharing can still expose content. Use Zero-Width Character Remover only after inspecting whether invisible characters are evidence rather than noise.

3. Step 3

   Record the counting contract

   For frequency or density, record the segmentation locale, NFC, NFKC, or no normalization, case sensitivity, number inclusion, minimum token length, whitespace policy, row cap, and exact phrase. Define a character as an Intl.Segmenter grapheme cluster. Define density as overlapping occurrences divided by eligible same-length windows that do not cross filtered tokens, sentence punctuation, or line breaks.

4. Step 4

   Apply language-specific boundaries

   Use the five readability values only for predominantly English passages and retain the sentence, word, letter, estimated-syllable, and complex-word counts beside them. Treat samples below 100 representative words as unstable comparisons. Use lexicon sentiment only for English AFINN evidence, and label its token coverage as coverage rather than confidence. Do not translate or infer missing labels.

5. Step 5

   Run the bounded analysis and inspect diagnostics

   Load the sanitized source in Text Analysis Workbench, set the written options, and run the dedicated Worker. Review every exclusion and transformation notice before the first score. For similarity, verify whether exact match means original text or transformed text, and confirm the grapheme-cell preflight stayed below 12,000,000. For a changed draft, open Diff Checker as well so location and meaning are not replaced by one percentage.

6. Step 6

   Report evidence, limits, and the human decision separately

   Save the source version, options, aggregate counts, complete report, browser date, and interpretation in separate fields. Quote the denominator with density, the formulas with readability, matched terms and coverage with lexicon tendency, and transformation policy with similarity. Then let an editor or domain reviewer decide whether wording, correctness, safety, accessibility, or tone should change. Re-run the same contract after revision rather than moving the target.

Example

Decision: reduce accidental repetition in an English help article. Contract: English segmentation, NFKC, case-insensitive, numbers excluded, minimum one grapheme, target “local text tools,” overlapping three-token windows, sentence and line boundaries preserved.

Report the target occurrence count and eligible three-token window count together, list any normalization or exclusion notices, inspect each occurrence in the source, and revise only when repetition harms clarity rather than chasing a universal percentage.

Common mistakes

FAQ