What Readability Scores Actually Measure (It Is Not Comprehension)

Paste text into any readability checker and you get a grade level, usually with the implication that lower is better. What almost none of them tell you is that every one of these formulas measures the same two surface features — how long your sentences are, and how long your words are — and neither of those is comprehension.

That’s not a reason to ignore them. It’s a reason to know exactly what you’re optimizing.

The formulas, and what they actually compute

Six formulas dominate. Here’s what each one is, in full, so you can see how little separates them.

Flesch Reading Ease (1948)

Rudolf Flesch’s original, and the only one that outputs a 0–100 score rather than a grade level. Higher is easier.

206.835 − 1.015 × (words / sentences) − 84.6 × (syllables / words)

Two inputs: average sentence length, average syllables per word. That’s the whole formula. The constants were fit to reading tests from the 1940s and have no deeper meaning.

Score	Conventional label
90–100	Very easy (5th grade)
80–89	Easy
70–79	Fairly easy
60–69	Standard (8th–9th grade)
50–59	Fairly difficult
30–49	Difficult (college)
0–29	Very difficult (graduate)

Note it isn’t bounded. Sufficiently short sentences of one-syllable words score above 100; dense academic prose goes negative.

Flesch-Kincaid Grade Level (1975)

The same two inputs, re-fit to output a US grade level. Developed by Kincaid and colleagues for the US Navy, to check whether technical manuals were readable by enlisted personnel.

0.39 × (words / sentences) + 11.8 × (syllables / words) − 15.59

That origin explains a lot about its behavior. It was validated against military training material — procedural, technical, instructional prose — not marketing copy, fiction, or conversation.

Gunning Fog (1952)

Robert Gunning’s index, aimed at business writing. Introduces the idea of a “complex word”: three or more syllables.

0.4 × [ (words / sentences) + 100 × (complex words / words) ]

Same shape as the others, but it counts how many long words rather than the average syllable count — so a text with many short words and a few very long ones scores differently than under Flesch.

SMOG (1969)

“Simple Measure of Gobbledygook,” by G. Harry McLaughlin. Counts polysyllabic words (3+ syllables) and ignores word count entirely.

1.0430 × √(polysyllables × 30 / sentences) + 3.1291

SMOG is the formula most used in healthcare and public-health material, largely because it was validated against 100% comprehension rather than the 50–75% used by earlier formulas — a stricter target. Its documented caveat: it was designed for samples of at least 30 sentences. Run it on a paragraph and the number is not meaningful, however confidently a tool reports it.

Coleman-Liau (1975)

The interesting outlier. Uses characters instead of syllables, deliberately, because character counting could be automated in 1975 and syllable counting could not.

0.0588 × L − 0.296 × S − 15.8
    L = letters per 100 words
    S = sentences per 100 words

Because it never counts syllables, it never inherits syllable-counting errors — which, as we’ll see, are real. It’s the score to weight most heavily on text with unusual vocabulary.

Automated Readability Index (1967)

Also character-based, developed for the US Air Force.

4.71 × (characters / words) + 0.5 × (words / sentences) − 21.43

ARI tends to run about a grade higher than Flesch-Kincaid on the same text. Neither is more correct; they’re fit to different reference material.

What they all have in common

Look at those six formulas together and one thing stands out: they are all linear combinations of sentence length and word length. The differences are the constants and whether “word length” means syllables or characters.

Which means a readability score cannot see:

• Whether your sentences are in a sensible order. Shuffle every paragraph of this post randomly. The score doesn’t move.
• Whether the words are familiar. “Antidisestablishmentarianism” and “unbelievability” cost about the same. “Cat” and “gnu” score identically. Word difficulty is not word length — Dale-Chall, an older formula not in this set, is the one that uses an actual list of familiar words, which is why it needs a dictionary to run.
• Whether the logic follows. A confident non-sequitur scores wonderfully.
• Whether the reader knows the domain. “The pod will scale to zero” is short, simple, and incomprehensible to most people.
• Jargon, acronyms, or ambiguous pronouns — all of which are short.

Here is a sentence that scores at roughly a 4th-grade level: “The GC pause hit p99 and the pods got OOM killed.” Nine words, mostly one syllable, one short sentence. Flesch-Kincaid says a nine-year-old can read it.

And here’s the inverse failure — you can improve any score mechanically by chopping sentences at their commas and swapping long words for short ones, without changing meaning at all. The score is gameable, which means the score is not the goal.

The syllable-counting problem

There’s a technical wrinkle worth knowing about, because it makes some scores less reliable than they look.

Four of the six formulas need a syllable count, and English syllable counting has no closed-form rule. Software approximates by counting vowel groups with an exception list. That approximation gets business (2, not 3), created (3, not 2), every (2, not 3), and hundreds of other words wrong in either direction.

The errors mostly wash out over a long document, because the formulas use averages — that’s the saving grace, and it’s why syllable-based scores are fine on an article and shaky on a paragraph. But it’s a genuine reason to look at Coleman-Liau and ARI as a cross-check: if the character-based scores and the syllable-based scores disagree sharply, your text probably has unusual vocabulary that the syllable estimator is mishandling.

Sentence detection has the same issue in miniature. A naive splitter on . breaks at “Dr. Smith” and “e.g.” — inflating your sentence count and making the text look easier than it is. Any decent implementation carries an abbreviation list. Our readability analyzer handles this and reports the underlying sentence, word, and syllable counts alongside the scores, so you can sanity-check the inputs rather than trusting a single number.

The “8th grade” target

You’ve heard the advice: write at an 8th-grade level. It’s real, but its provenance is narrower than the way it’s repeated.

The strong version of it comes from health literacy, where organizations including the NIH and AMA have long recommended that patient-facing material target roughly a 6th-to-8th-grade reading level. The rationale is specific and well-supported: patient materials must be understood by everyone, including people who are stressed, unwell, or reading in a second language, and the cost of misunderstanding is medical harm.

Generalizing that to all writing is a stretch. A reference page for senior engineers is not patient discharge instructions. What survives the generalization is the principle — write for the audience you actually have, under the conditions they’re actually in — not the number 8.

The more defensible use of these scores in general writing is relative, not absolute: run your draft, then run a piece of writing you admire in the same genre, and compare. A grade level 4 points above your reference is a signal worth investigating. A grade level of exactly 8.0 is not an achievement.

Where the scores are genuinely useful

Despite everything above, they earn their place:

• Catching drift. You cannot feel your own sentence length while writing. A score that jumped from 9 to 14 between drafts tells you something real happened.
• Finding the worst paragraph. Scoring section by section locates the dense one faster than rereading.
• Enforcing a floor across a team. For documentation and support content written by many people, a threshold in CI prevents the worst outliers, even if it can’t produce good writing.
• Legal and regulatory compliance. Several US states mandate specific reading levels for insurance policy documents. There, the number is the requirement, gameable or not.
• Arguing with a stakeholder. “This paragraph reads at a graduate level” is a more actionable note than “this feels dense.”

The actual advice

If you want to write more readably, the formulas tell you exactly two levers, and they’re both real:

1. Cut sentence length — but by removing subordinate clauses and restructuring, not by replacing every comma with a period. Choppy prose scores well and reads badly.
2. Prefer shorter words where a shorter word means the same thing — “use” over “utilize,” “before” over “prior to.” But don’t swap a precise long word for an imprecise short one; that trades comprehension for score.

Then do the three things no formula can measure: put the conclusion first, cut what the reader doesn’t need, and define the jargon you keep.

Try it

• Readability Analyzer — all six formulas with the underlying sentence, word, syllable, and polysyllable counts, plus a plain-English verdict. Runs entirely in your browser.
• Word Counter — counts, reading time, and frequency.
• Text Diff — compare drafts once you’ve made the cuts.