Readability Checker scores the text you paste seven ways, and every one of those seven runs in your browser. Nothing is uploaded, there is no request to make after the page has loaded, and your draft never leaves your device. That matters more here than on most tools, because what people paste in is usually unpublished: a specification, a contract clause, an internal memo.
What comes back is a disagreement. Paste 97 words of ordinary technical prose and the seven indices return grades running from 6.4 to 30.6: 24.2 grade levels from the same paragraph. Four land inside a 5.8-point band, one sits far below them, one refuses to answer. None of that is a fault in the tool. It is the finding.
Key Takeaways
- 97 words, seven indices, a 6.4 → 30.6 spread. Flesch–Kincaid (24.8), ARI (26.6), Gunning Fog (30.4) and Coleman–Liau (30.6) sit within 5.8 points of each other.
- Dale–Chall reads 6.4 alone, and on easy prose it inverts: the hardest grade of the seven exactly where ARI returns the easiest.
- Closing that gap would take 139.3% and 164.7% of words flagged difficult. A complete 3,000-word list would not do it either.
- SMOG is the only one of the seven with a published minimum sample, and the only one that declines below it: 30 sentences. The other six return confident numbers from five.
- The word-list method and the counting method do not agree. That is the defensible claim. "Seven random numbers, pick one" is not.
What the checker actually measures
Each index takes two measurements: average sentence length, and one proxy for word difficulty. The proxies disagree, and that disagreement is the subject.
| Index | Second input | Direction | Formula |
|---|---|---|---|
| Flesch Reading Ease | syllables per word | higher is easier | 206.835 − 1.015·ASL − 84.6·ASW |
| Flesch–Kincaid Grade | syllables per word | lower is easier | 0.39·ASL + 11.8·ASW − 15.59 |
| Gunning Fog | % of words with 3+ syllables | lower is easier | 0.4·(ASL + 100·complex%) |
| SMOG | count of 3+ syllable words | lower is easier | 1.0430·√(poly·30/sentences) + 3.1291 |
| Coleman–Liau | letters per 100 words | lower is easier | 0.0588·L − 0.296·S − 15.8 |
| Automated Readability Index | letters and digits per word | lower is easier | 4.71·(chars/words) + 0.5·ASL − 21.43 |
| Dale–Chall (approx.) | % of words off a familiar-word list | lower is easier | 0.1579·%difficult + 0.0496·ASL, +3.6365 when %difficult ≥ 5 |
Two details in that table matter more than the constants. ARI counts letters and digits only, spaces and punctuation excluded, which is what Wikipedia's ARI article means by "letters and numbers". Coleman–Liau used characters deliberately; its article says the index "does not require that the character content of words be analyzed, only their length in characters". Gunning's index came from newspaper and textbook publishing in 1952.
Who built each one, and what they assumed
| Index | Built by | Year | Calibrated against |
|---|---|---|---|
| Flesch Reading Ease | Rudolf Flesch | 1948 | general adult prose |
| Flesch–Kincaid Grade | Kincaid's team, under U.S. Navy contract | 1975 | Navy technical training |
| Gunning Fog | Robert Gunning | 1952 | newspaper and textbook copy |
| SMOG | G. Harry McLaughlin | not printed here | consumer health messages |
| Coleman–Liau | Meri Coleman & T. L. Liau | not stated in my source | hard-copy text |
| Automated Readability Index | not stated in my source | not stated in my source | English character counts |
| Dale–Chall | Edgar Dale & Jeanne Chall | 1948, revised 1995 | fourth-grade American students |
Three cells are blank on purpose: the ARI article names neither an author nor a year, the Coleman–Liau article names authors but no year, and SMOG's year because my sources conflict.
The printed dates came from bibliographic records, not memory. ERIC's reprint collection The Classic Readability Studies lists "A new readability yardstick" by Rudolf Flesch, June 1948, in the Journal of Applied Psychology vol. 32 no. 3. That record anchors Reading Ease's 1948 date, and it is why Reading Ease is never credited to G. Harry McLaughlin, who wrote SMOG. The same collection dates Dale and Chall's two papers to February 1948 in Educational Research Bulletin; the 1944 date that circulates belongs to Irving Lorge, not to a co-author. The 1995 revision is a book, Readability Revisited.
One caution on Wikipedia dates: its readability navbox lists "Flesch reading ease [1975]", the Flesch–Kincaid grade-level date on the wrong formula. Unusable as a date source.
One paragraph, seven numbers
The passage under test is 97 words of compiler-optimisation prose across five sentences, free of abbreviations, initials and decimals, so nothing about segmentation is arguable.
The implementation persists intermediate representations to an immutable cache, invalidating entries whenever the underlying dependency graph is re-evaluated during incremental compilation. Subsequent compilations therefore observe substantively fewer redundant traversals, which materially reduces wall-clock latency across substantially larger modules. Nevertheless, the amortised benefit diminishes considerably once the graph exceeds several thousand vertices, because invalidation bookkeeping dominates the residual computational budget. Optimisation is additionally constrained by conservative assumptions regarding aliasing, since the compiler cannot prove absence of mutation across opaque boundaries without interprocedural annotations. Consequently, practitioners frequently observe pathological recompilation behaviour that conventional profiling methodology fails to attribute correctly.
| Index | Hard passage, 97 words | Easy passage, 28 words |
|---|---|---|
| Flesch Reading Ease | −48.3 (higher is easier) | 100.7 |
| Flesch–Kincaid Grade | 24.8 | 1.5 |
| Gunning Fog | 30.4 | 3.7 |
| SMOG | declined — 5 sentences, needs 30 | declined |
| Coleman–Liau | 30.6 | 2.7 |
| Automated Readability Index | 26.6 | 0.6 |
| Dale–Chall (approximate) | 6.4 | 5.2 |
Sentence length roughly doubles, 9.3 to 19.4, and syllables per word more than doubles, 1.1 to 2.8. The syllable-based family punishes the second jump; the character-based family punishes the first. That accounts for most of the spread, which is why the seven read as a family with one dissenter.
The outlier inverts on easy prose
Dale–Chall is the outlier on hard prose, 6.4 against a pack between 24.8 and 30.6, and the only index built on a word list rather than syllable or character counts. On the 28-word easy passage the ordering flips: Dale–Chall returns 5.2, the hardest grade of all seven, while ARI puts the same passage at 0.6, the easiest reading anyone produced.
An index that merely disagreed would stay on one side of the pack. This one changes sides: it is not noisy against the character-based indices but mis-scaled relative to them in both directions, and which way it errs depends on the register.
The trimmed word list is not the explanation
The obvious excuse is bundle size, and the source file says so: a trimmed Dale–Chall-style list, because the original holds 3,000 words, "a lot of client bundle to ship for a single percentage". Wikipedia's Dale–Chall article confirms 3,000 familiar words, up from 769 in 1948, built from what fourth-grade students could understand.
On the hard passage the shipped list flags 11 words out of 97: attribute, benefit, budget, conservative, methodology, nevertheless, redundant, several, subsequent, therefore, underlying. It stays silent on implementation, intermediate, representations, compilation, invalidation, amortised, optimisation, interprocedural.
But the trimming is not what produces 6.4. Hold sentence length at 19.4 and walk the difficult-word percentage upward: each point is worth 0.1579 grade levels, so flagging every word in the passage reaches only about 20.4, arithmetically impossible though that is, and still under Flesch–Kincaid's 24.8 and far from Coleman–Liau's 30.6. In reverse: what percentage would reconcile Dale–Chall with ARI's 26.6? 139.3%. With Coleman–Liau's 30.6? 164.7%. Both overshoot 100 by a wide margin.
A complete 3,000-word list could not reconcile Dale–Chall with the character-based indices on technical prose either. The divergence is structural, not a packaging decision.
Three places the code and the documentation disagree
Checking the shipped engine against its own documentation turned up three discrepancies, each checkable in source.
- ARI rounding. Wikipedia's ARI article states that "Non-integer scores are always rounded up to the nearest whole number, so a score of 10.1 or 10.6 would be converted to 11." The shipped engine rounds to one decimal place (
round1, atreadability-tools.ts:387), so 26.6 prints as 26.6, not 27. - The Dale–Chall threshold. Wikipedia: "If the percentage of difficult words is above 5%, then add 3.6365." The code tests
>= 5.aboveand>=disagree at exactly 5.0%, and nowhere else. - The word-list size. The file header and the tool documentation both say 1,071 entries. The file holds 1,071 string literals and 1,032 unique words once
new Set()deduplicates them at line 11.
None disturbs the thesis. All three argue for reading a tool's source: documentation is a claim about code.
SMOG is the only one that refuses to answer
Wikipedia's SMOG article gives the procedure: "Take three ten-sentence-long samples from the text in question. In those sentences, count the polysyllables (words of 3 or more syllables)." Three times ten is 30, and the same article says why the floor is not negotiable: "tables for texts of fewer than 30 sentences are statistically invalid, because the formula was normed on 30-sentence samples."
No published minimum was found for the other six. That is a negative result from the sources I read, not proof that none exists.
On the 97-word passage the checker renders that refusal as visible text inside the tile, not a greyed-out control: SMOG is calibrated on 30 sentences — this has 5. The other six hand back confident numbers from those same five sentences. A 6.4 and a 30.6 arrive with identical authority, and exactly one of the seven has published grounds for saying I need more text first.
What a readability score does not tell you
A readability score is not a writing quality score, and a grade level is not a reader. Wikipedia's Flesch–Kincaid article states the limitation in one sentence:
"As readability formulas were developed for school books, they demonstrate weaknesses compared to directly testing usability with typical readers. They neglect between-reader differences and effects of content, layout and retrieval aids."
Content, layout, retrieval aids and the differences between readers sit outside all seven formulas. Typography is not outside readability in general: Wikipedia's article on readability itself counts it in, listing "font size, line height, character spacing, and line length". Every one of the seven ignores something readability properly includes.
The scale is also unbounded: the same article notes the score "does not have an upper bound", and puts the lowest grade in theory at −3.40, "belonging to the passage 'Go. See. Stop. Rest.' for example". A low number is not a failure state and a high one is not a wall. Grade 12 names a calibration point extrapolated out of a school-book corpus, not a person.
No index here is the correct one. Read the seven as a fan, not a committee: four agreeing in a narrow band means the count-based methods broadly agree.
The best-known formula in this family was once asked whether any of it worked. IEEE Transactions on Professional Communication ran an interview with J. Peter Kincaid in 1987 titled "Readability formulas: Useful or useless?". Putting that question to the formula's own inventor is not a verdict. It is the right one to leave open.
How the checker reads a sentence
The seven document numbers are only half the page. A sentence-level diagnosis is what these formulas support, and where the interface spends its effort.
- It segments once and shares the answer. Boundaries come from a shared segmenter the word counter also uses, so abbreviations, initials and decimals land the same on both pages.
- It computes all seven indices in one pass, exposing the formula and every named input behind each tile.
- It bands each sentence on a three-step rail driven by Flesch ease: 70 and above is easy, 30 to 69 is medium, below 30 is hard. Colour, rail width and a written label travel together, so no band depends on seeing colour.
- It names the reason, not only the verdict. Open a row and you get the excess over your word limit and the count of three-or-more-syllable words.
- It declines rather than guesses. When an index cannot be computed, the tile prints the reason.
Two style detectors sit behind switches that default to off: passive voice as a be-verb plus past participle, with adjectival forms such as named and born suppressed, and adverbs from a curated list plus a conservative -ly rule. Past 30% non-Latin letters the counts stay valid and the grades are labelled uncalibrated.
Scoring is synchronous: 5.1 ms for 650 words, 7.8 ms for 2,418, 78.9 ms for 23,699. Past roughly 100,000 characters the report moves into an idle callback.
No search-volume data stands behind this page: the keyword tool is unavailable in this pipeline, /tools/readability-checker/ returns zero rows in 90-day Search Console, and /tools/word-counter/ earns 5,555 impressions.
Related Tools
- The word counter renders Flesch, Flesch–Kincaid and Gunning Fog for a whole document on the same shared segmenter, and cross-links here.
- The online thesaurus is the next step: the checker names the difficult word, the thesaurus replaces it.
- The text summarizer covers the next job after a rewrite, when the question becomes what to cut.
- The text diff checker compares two drafts of one document, the way to see whether an edit helped.