Readability Checker

A readability score on its own is not much use. Knowing a page reads at grade 13 tells you it is too hard for a general audience, but not which sentences did it or what to do about them. This checker scores your text against eight published formulas, judges the result against a reading level you choose, and then points at the specific sentences, words and phrases behind the number.

Start by saying who the text is for. General web content, plain language, children, teenagers, technical documentation, academic — each preset sets a target grade, and every score is then coloured against that target rather than against an abstract scale. A single consensus figure sits at the top with the gap between where the text reads and where you wanted it, so the answer to "is this right for my readers?" arrives before any of the individual formulas.

Underneath, the eight scores are reported separately because they disagree in informative ways. Flesch Reading Ease and Flesch-Kincaid are built on sentence length and syllables per word. Gunning Fog and SMOG count words of three or more syllables. Coleman-Liau and the Automated Readability Index use letters per word instead of syllables, which makes them a useful cross-check on the syllable estimate. LIX adds the Scandinavian measure, and Linsear Write weights long words heavily for technical prose. Where the grades spread more than five years apart, the tool says so — that spread means the text is unusual, not that the reading is precise.

Then the diagnosis. The sentences that run long are shaded where they sit in your text and listed with their word counts. The words of three or more syllables are tabled by how often each appears, so the top rows are the ones worth replacing. Possible passive constructions, -ly adverbs and wordy phrases each get their own list, with a shorter alternative suggested for every phrase. Edit in the box and the grade moves as you type.

These formulas were developed in the middle of the twentieth century and they measure surface features only. They cannot tell whether the argument follows, whether jargon is explained, whether the structure helps a reader find what they need, or whether the writing is any good. Text can score well and still be incomprehensible, and a passage of short choppy sentences will always score better than a well-built longer one.

That is why the score is not the product here. The lists underneath it are: the sentences that are doing too much, the long words that recur often enough to be worth replacing, the passive constructions where naming who acts would be shorter as well as clearer, and the phrases that have a two-word equivalent. Those findings are heuristics too and say so — the passive check flags "was tired", which is an adjective, alongside "was written", which is not — but they are things a writer can act on, which a grade is not.

Treat the number as a signal rather than a target. Chasing a specific figure by mechanically chopping sentences produces worse writing. The interventions that genuinely help — shortening a sentence that was carrying two ideas, using a plain word where jargon crept in, naming the actor, defining a term on first use — happen to move the score as a side effect, which is the right way round.

Six of the eight formulas divide by the number of sentences, which makes sentence detection the single most consequential thing a readability tool does. Get it wrong and you do not get one wrong number, you get six. A naive splitter that breaks on every full stop reads "Dr. Smith went to the U.S. yesterday to meet Mr. Jones." as five sentences rather than one, dropping the average sentence length from eleven words to two and moving Flesch Reading Ease by roughly ninety points.

So the splitter here knows about titles and abbreviations, about initials such as J. R. R. Tolkien, about the full stop inside 3.50, and about ellipses and closing quotation marks. It also treats a line break as a sentence boundary by default, which matters more than it sounds: a heading or a bullet point carries no full stop, so without that rule it gets glued onto the paragraph below it and inflates the average sentence length of the whole document.

The same care applies to what counts as a character. Coleman-Liau and the Automated Readability Index are defined over letters, not over every character in the string. Feeding them the raw character count — spaces and punctuation included — inflates both by five or six grade levels, which is easy to do and impossible to notice from the output.

Every formula here is a regression fitted to passages of a hundred words or more. Below that the numbers are noise, and below about twenty-five words there is no signal at all — so nothing is scored, and the reason is stated instead of a number being invented. Between twenty-five and a hundred words every score is marked provisional on its own card.

SMOG deserves particular mention because its definition is explicit: it is calculated on a sample of thirty sentences. Applied to a five-sentence paragraph it will still produce a figure, and that figure means nothing. Here it carries a caveat below thirty sentences and is left out of the consensus grade until the sample is large enough, so one unreliable formula cannot drag the headline number.

Syllable counting is the remaining soft spot, and it is a heuristic in every tool that does this without a dictionary. English orthography defeats simple rules — science is two syllables, being is two, wanted is two, hoped is one. This implementation handles the silent e, the consonant-plus-le ending, the -ed and -es inflections and the vowel pairs that are genuinely two sounds, and it scores 128 out of 128 on the test battery kept alongside the code. Expect a wider margin on proper nouns and invented technical terms, and expect nothing useful at all from text that is not English: these formulas and their grade mappings do not transfer between languages, and the tool warns you when the text looks like it might not be.