Methodology & Accuracy

Every metric on this site is computed by one small engine of pure functions, and every stage of that computation is documented here — including the places where the output is an estimate and by roughly how much it can be off. The pipeline has three stages: tokenization (finding words), segmentation (finding sentences and paragraphs), and derived metrics (readability, density, frequency, timing) computed from the first two.

This page is the reference: it records what each stage actually does, in the order the code does it, with a worked example per stage that is computed by that same engine when the site is built. The guides are the explanations built on top of it. Each one takes a single stage, walks through it at more length, and shows how to read the output on the tool that exposes it — so if a figure on a tool page surprises you, the guide says why and this page says exactly what was computed. The guides are:

Accepted inputs

Every function in the engine validates its input before doing any work and returns a typed result rather than throwing or producing a NaN, so a tool can show a message instead of a broken number. The bounds are few:

  • Text of up to 2,000,000 UTF-16 code units — most emoji occupy two, so they count twice against this cap. Above that the engine declines with the message "Text longer than 2 million characters is not supported." Empty or whitespace-only text is valid for counting and returns zeros (0 words, 0 sentences, 0 paragraphs), but readability scoring needs at least one word and one sentence and otherwise reports "Enter at least one full sentence to score."
  • Keyword density takes a "top N" between 1 and 100. Anything else — a zero, a fraction, a larger number — returns "Choose between 1 and 100 keywords."
  • The character-limit check takes a limit between 1 and 100,000 and counts Unicode code points against it; an invalid limit returns "Enter a valid character limit."

Nothing else is validated, because nothing else needs to be. There is no language detection, no minimum length, and no attempt to reject text that is "not prose": a list of URLs, a spreadsheet column, or a poem is analyzed by exactly the same rules as an essay. That is the honest choice, and it is also why the sections below spend so much space on edge cases.

Stage 1 — Tokenization: what counts as a word

A word is a Unicode-aware run of letters and digits in which internal hyphens and apostrophes do not break the run. Concretely: "state-of-the-art" and "it's" are one word each, accented and non-Latin scripts count correctly, and numbers count as words. The rule's edges are worth knowing, because they explain most disagreements with other tools:

  • "2,500" is two words ("2" and "500") because a comma ends a run, while "3.14" is also two words — but see below for why the period inside it does not end a sentence.
  • A URL such as https://example.com tokenizes as three words (https, example, com).
  • "e.g." contributes two one-letter words — and, separately, a false sentence break.
  • Emoji are not words; they are counted as characters (one per Unicode code point).

Word count is the base for everything else, so this stage feeds every tool: the Text Analyzer, the Readability Checker, the Speaking Time Calculator, the Keyword Density Checker, and the Word Frequency Counter. The unique-word figure on the Text Analyzer comes from the same token list, compared case-insensitively — "Plan" and "plan" are one unique word — but with any hyphens or apostrophes left attached, so a token wrapped in straight apostrophes is distinct from the bare word. The guide What counts as a word walks through the rule with more cases.

Worked example

"The state-of-the-art scanner can't finish 2,500 pages in 3.14 seconds." — the engine reports 12 words in 1 sentence, 70 characters with spaces and 61 without. The hyphenated compound and the contraction are one word each; each of the two numbers contributes two words rather than one, which is where a hand count and this count part ways.

"See https://example.com for details." — 6 words, because the URL splits at its punctuation. "rock - paper" — 3 words: a hyphen set off by spaces is matched as a run of its own, so a text full of spaced dashes counts slightly high. "Plan the plan. The PLAN holds." — 6 words but only 3 unique, since uniqueness ignores case.

Stage 2 — Segmentation: sentences and paragraphs

Sentences split on runs of ending punctuation (. ! ? …) followed by whitespace or the end of the text, and a fragment must contain at least one letter or digit to count. Two consequences follow. First, abbreviations followed by a space over-count: "Mr. Smith arrived." registers as two sentences, a limitation shared by most automated counters and stated here rather than hidden. Second, decimals like "3.14" do not split, because the period is not followed by whitespace. Headings and list items without ending punctuation merge into the following sentence, which matters for readability scoring (below). Paragraphs are blocks separated by one or more blank lines; single line breaks do not split a paragraph.

Sentence counts appear on the Text Analyzer and the Readability Checker, where words-per-sentence is one of the two ratios in the formula — so a sentence over-count lowers the apparent sentence length and nudges the score upward, and a heading merged into its first sentence does the opposite. If a sentence count looks wrong, the first things to look for are abbreviations followed by a space and headings or bullets without a full stop.

Worked example

"Mr. Smith arrived. He sat down." — 3 sentences for 6 words, one of them the false break after the abbreviation.

"Wait... what?! Really." — 3 sentences: a run of dots is one break, and "?!" is one break, because the splitter consumes the whole run of ending punctuation at once.

"First line." on one line, "Still the same paragraph." on the next, then a blank line and "A new paragraph." — 2 paragraphs holding 3 sentences. The single line break does not open a new paragraph; the blank line does.

"Quarterly summary" as an unpunctuated heading above "Revenue rose. Costs fell." — 2 sentences averaging 3.0 words each, because the heading's words are folded into the sentence that follows it.

Stage 3a — Readability: the Flesch formulas and their sensitivity

Reading Ease = 206.835 − 1.015 × (words ÷ sentences) − 84.6 × (syllables ÷ words). Flesch–Kincaid Grade = 0.39 × (words ÷ sentences) + 11.8 × (syllables ÷ words) − 15.59. These are the standard published coefficients — Rudolf Flesch's 1948 Reading Ease and the 1975 grade-level recalibration produced for the US Navy — applied without modification. The coefficients tell you what moves a score: each additional word of average sentence length costs about 1 Reading Ease point, while a rise of just 0.01 in syllables-per-word costs about 0.85 points. Word length, not sentence length, dominates — which is why swapping latinate vocabulary for plain equivalents moves scores faster than splitting sentences. Note the scale is conventionally read as 0–100, but the formula itself is unbounded: extremely dense prose scores negative, and the checker reports that honestly rather than clamping it.

Both scores run on the Readability Checker and the Text Analyzer. The two guides, Flesch Reading Ease, explained and Flesch–Kincaid grade level, take each formula apart in turn. The checker attaches a plain-language band to the Reading Ease score; the band is a conventional label for a range of the scale, not a measurement of any particular reader, and the examples below quote the band the engine actually assigns.

Worked examples

Four short original texts, scored by the engine at build time. Only the two ratios in the middle columns reach the formulas; everything to their right follows from them.

Sample Words Sentences Syllables Words ÷ sentences Syllables ÷ words Reading Ease FK grade Band
Plain 16 3 18 5.33 1.13 106.2 −0.2 very easy — about 5th-grade level
Everyday 41 3 65 13.67 1.59 58.8 8.4 fairly difficult — high-school level
Dense 8 1 34 8.00 4.25 −160.8 37.7 very difficult — graduate level
Rewrite 12 1 18 12.00 1.50 67.8 6.8 plain English — 8th–9th grade

The texts, in table order. Plain: "We met at noon. The rain had stopped. We walked to the shop and bought bread." Everyday: "The garden needed water every morning, so we set a timer and stopped forgetting. Most of the plants recovered within a week, although the tomatoes stayed stubborn. By August the beds were full again and the neighbours started asking for cuttings." Dense: "Institutional reorganization necessitates comprehensive stakeholder consultation before implementation." Rewrite: "We must ask everyone involved before we change how the office works."

The dense sample shows the arithmetic at its most lopsided. Its Reading Ease is 206.835 − 8.12 − 359.55 = −160.8: the first subtracted term comes from sentence length and the second from word length, and it is the second, not the first, that pushes the score below zero. The rewrite says the same thing in everyday words and scores 206.835 − 12.18 − 126.90 = 67.8, a gap of 229 points — almost all of it from the syllables-per-word ratio falling from 4.25 to 1.50. The grade formula consumes the same two ratios and rescales them: for the rewrite, 0.39 × 12.00 + 11.8 × 1.50 − 15.59 = 6.8. Both formulas consume the same two ratios, but they weight them differently, so two texts can rank in a different order on the two scales; the grade is a second formula, not a relabelling of the ease score.

Common misreadings

  • Scoring one sentence. The ratios are whole-text averages. With a single sentence there is nothing to average, so one long or one polysyllabic sentence is the entire result — the dense sample above is one sentence, and so is the rewrite. Score a passage.
  • Scoring an outline. Headings and bullets without ending punctuation merge into the sentence after them (Stage 2), inflating words-per-sentence. Score running prose.
  • Reading the grade as an age. The grade number is a rescaling of the same two ratios into a US school-grade convention; it is not a claim about any reader's age or ability, and a negative grade, or one far above any school year, simply means the two ratios produced a number the convention has no label for.
  • Reading the band as a verdict. The formula counts syllables, not meaning. Short words used technically score "easy" while being hard to follow; a long but familiar word costs as much as an obscure one. The score measures the two ratios and nothing else.

Stage 3b — The syllable heuristic, measured honestly

Syllables are estimated, not looked up: the engine counts vowel groups (a, e, i, o, u, y), subtracts one for a silent final "e" (except "-le" endings, so "table" keeps its two), and never returns less than one for a word that still has letters once everything outside a–z is stripped; a token with no a–z letters at all — a bare number, or a stray hyphen or apostrophe — returns zero. Spot-checking the heuristic against dictionary syllabification shows both error directions: "created" comes back as 2 (actually 3, because "ea" merges into one vowel group), while "business" comes back as 3 (actually 2, because the silent "i" is counted). "Beautiful" (3), "readability" (5), and "strengths" (1) come back exactly right. Because a text's score averages over many words, individual miscounts partially cancel, and totals typically land within a few points of hand computation — the same trade-off virtually every automated checker makes. Treat scores as strong signals, not laboratory measurements, and expect small differences between tools that use different heuristics.

The syllable count has no tool of its own; it appears only inside the readability panel on the Readability Checker and the Text Analyzer, where it feeds the syllables-per-word ratio. The guide How syllables are counted steps through the rule letter by letter. Two further mechanics matter. The heuristic strips every character outside a–z before counting — hyphens, apostrophes, digits, and accented letters all vanish — so a hyphenated compound is scored as one long word, and a number contributes no syllables at all while still counting as a word in the denominator, which pulls the ratio down for number-heavy text. And because the stripping removes anything that is not an unaccented Latin letter, the readability stage is English-only by construction.

What the heuristic returns

Each count below is produced by the engine at build time. The first five words are the ones checked against a dictionary above; the rest show the rules in action.

Word Syllables returned
created 2
business 3
beautiful 3
readability 5
strengths 1
table 2
cake 1
hello 2
little 2
state-of-the-art 3
2500 0

Walking the rule: "beautiful" has the vowel groups "eau", "i", and "u", and no final "e", so the group count stands. "Cake" has "a" and "e", ends in "e" but not "-le", and has more than one group, so one is subtracted. "Table" also has two groups, but the "-le" exception keeps both. "Created" has "ea" and "e" — two groups — and ends in "d", so nothing is subtracted and the third syllable a dictionary would give it is lost. "Business" has "u", "i", and "e", and the silent "i" is counted like any other vowel. "State-of-the-art" is scored as the single string of letters left after the hyphens are removed, and "2500" has no letters at all.

Stage 3c — Keyword density and word frequency

Both tools rank words case-insensitively, with two exclusions from the ranking: a fixed list of 74 common English function words (the, and, of, that…) and single-character tokens. Two design decisions matter when comparing against other tools. Density percentages divide by all words — stop words included in the base — so a word appearing twice in 40 words is 5% here even though a tool that strips stop words first would report a higher share for the same text. And there is no stemming: "run", "runs", and "running" rank separately, deliberately, because collapsing word forms would hide exactly the repetition patterns the frequency counter exists to reveal.

This stage drives the Keyword Density Checker and the Word Frequency Counter, and the density table on the Text Analyzer. The guide Keyword density, explained covers what the percentage can and cannot tell you. Three smaller rules complete the picture: tokens are lowercased before counting, so capitalized and lowercase forms of a word merge into one entry; the single-character exclusion is measured in code points, so a lone letter, a lone digit, or a stray hyphen or apostrophe standing alone is skipped (emoji never become tokens in the first place, as Stage 1 notes); and ties in count are broken alphabetically, so the order of two equally frequent words carries no information.

Worked example

"Runners run because running clears the mind. A runner who runs daily runs farther, and the run becomes the reward. Run early, run often, and keep running: the running takes care of itself."

The text has 33 words in total, stop words included. Asking for the top 5 keywords returns:

Word Count Density
run 4 12.1%
running 3 9.1%
runs 2 6.1%
because 1 3.0%
becomes 1 3.0%

The top entry is "run": 4 occurrences ÷ 33 words × 100 = 12.1%. Its other forms rank as separate entries rather than being folded in, the capitalized "Run" that opens the last sentence counts toward the lowercase entry, "the" is absent because it is a stop word (though every one of its occurrences is still in the denominator), and "a" is absent on both grounds — it is a stop word and a single character.

Stage 3d — Timing estimates

Reading time divides the word count by 238 words per minute, the adult silent-reading average from Brysbaert's 2019 meta-analysis of reading-rate studies. Speaking time uses 150 wpm, a typical prepared-speech pace. Both are averages over populations and genres: skimmers read faster, technical material reads slower, and spoken delivery of number-heavy text runs long because "2,500" counts as two words yet is voiced as four. The figures are estimates for planning, not stopwatches.

Reading time appears on the Text Analyzer and beside the speaking estimate on the Speaking Time Calculator; speaking time on both. The guides Reading time and words per minute and Speaking time and speech length explain where the two averages come from and how far a real reader or speaker can drift from them. Mechanically the stage is a division and a formatter: the division produces minutes as a decimal, and the formatter rounds to the nearest whole second, drops a zero seconds figure, reports anything that rounds to between one and four seconds as "under 5 sec", and reports anything that rounds to zero — empty text, or a single word — as "0 sec". No adjustment is made for content — a table of figures and a paragraph of prose with the same word count get the same estimate, which is why the number-voicing caveat above matters.

Worked example

The sentence "Check the valve and close the hatch." repeated until it reaches 476 words: reading time is 476 ÷ 238 = 2.00 minutes, displayed as 2 min; speaking time is 476 ÷ 150 = 3.17 minutes, displayed as 3 min 10 sec.

The 12-word tokenization sample above reads in "under 5 sec" and is spoken in "5 sec" — the rounding and the under-five floor both visible. And "2,500" on its own is 2 words to the engine, which is the mechanism behind the number-heavy caveat.

Stage 3e — SMS segmentation

The engine also implements SMS segmentation, following GSM 03.38 — the rules that decide how many text-message segments a string occupies. No tool on this site currently exposes it; it is documented here because the shared library computes it and because the guide SMS length and character counting explains it in depth. The rules are mechanical and all-or-nothing:

  • If every character in the message belongs to the GSM-7 alphabet, the message is encoded as GSM-7. Characters in the basic set cost one septet each; the extended characters — ^ { } \ [ ~ ] | € and the form feed — cost two septets each.
  • A GSM-7 message fits in one segment up to 160 septets. Beyond that it is concatenated, and each segment carries 153 septets.
  • If any character falls outside both GSM-7 sets, the whole message switches to UCS-2. Length is then counted in UTF-16 code units, with 70 units in a single segment or 67 per segment when concatenated. Most emoji occupy two code units, so each costs two.

The switch is the part that surprises people. Encoding is decided for the entire message, so one character outside the GSM-7 alphabet does not add a unit or two — it more than halves the capacity of every segment. One easy way to trigger it is a curly apostrophe or quotation mark in place of the straight form: the straight forms are in the GSM-7 set and the curly forms are not. The engine reports the encoding it chose, the units consumed, the segment count, the per-segment capacity, and the units remaining before another segment would be needed.

Worked example

Message Characters Encoding Units Segments Per segment Remaining
Confirmation 146 GSM-7 146 1 160 14
Confirmation, one sentence longer 179 GSM-7 179 2 153 127
Extended characters 25 GSM-7 28 1 160 132
Straight apostrophe 13 GSM-7 13 1 160 147
Curly apostrophe 13 UCS-2 13 1 70 57
Emoji 11 UCS-2 12 1 70 58

The confirmation reads: "Your table for two is confirmed for Friday at 7pm. Reply YES to keep it or NO to release it. Doors open at 6.30pm; parking is behind the building." It uses 146 septets and fits one segment with 14 to spare. Adding "Please arrive ten minutes early." takes it to 179 septets and 2 segments, at which point the capacity per segment drops to 153.

"Total due: €45 [inc. VAT]" is 25 characters but 28 septets: the euro sign and both square brackets cost two each. "Don't be late" and "Don’t be late" are the same 13 characters, but the straight apostrophe stays GSM-7 with 147 units to spare while the curly one forces UCS-2 with 57. "Meet at 6 🙂" is 11 characters but 12 UTF-16 units under UCS-2, the emoji alone accounting for two.

Why tools disagree, and which one is right

Word processors, browsers, and online counters differ at the margins because the margins are genuinely ambiguous — hyphenated compounds, headings, trailing punctuation, URLs. Differences of a few units on long texts are normal and no cause for alarm. When an exact figure is contractual (a strict submission limit), the only counter that matters is the recipient's; ours tells you whether you are close enough to care.

How the engine is tested

The analysis library is pure functions with no DOM dependencies, pinned by automated tests: classic pangram counts, hyphen and apostrophe handling, Unicode and emoji cases, sentence and paragraph edge cases, readability arithmetic against hand-computed values, stop-word and density behavior, and the timing math. The same library runs on our sister sites WordCntr.com and CharLimit.net, so a correction verified once applies everywhere. If you believe a result is wrong, the contact page explains what to include — confirmed issues are fixed in the engine and locked in with a new test.

The worked examples on this page are held to the same standard. None of their figures is typed in: each is computed by the engine when the site is built, from sample texts written for this page, and a sample the engine rejects fails the build. A number here therefore cannot drift from what the tools report, and if the engine changes, this page changes with it.

Privacy as a design constraint

Everything you type or paste is processed by JavaScript running locally in your browser. Nothing is transmitted, logged, or stored — there is no server that ever sees your text, which you can verify in your browser's network tab: analysis generates zero requests. For how each stage above is read on the tool that exposes it, the guides take them one at a time. To see the whole pipeline applied to a text of your own, start with the Text Analyzer or the Readability Checker.