What reading level should I write at?

What the score is actually counting

Readability formulas have a reputation for insight they do not deserve, and understanding their inputs is the fastest way to use them well.

Flesch Reading Ease, published by Rudolf Flesch in 1948, takes exactly two measurements: average words per sentence, and average syllables per word. That is the whole formula. It returns a number normally between 0 and 100 where higher is easier, and the Flesch-Kincaid Grade Level published in 1975 rearranges the same two inputs to report a United States school grade instead.

Every other common formula works the same way, differing only in which proxy it uses for word difficulty. The Gunning Fog index counts words of three or more syllables. SMOG counts polysyllables per thirty sentences. The Coleman-Liau index uses characters per word instead of syllables, which makes it easier to compute mechanically. The Dale-Chall formula is the interesting outlier: rather than counting syllables it checks words against a list of about 3,000 that fourth-graders were found to know, which is closer to measuring familiarity than measuring length.

The consequence of counting only length is that the formulas are blind to almost everything that makes text hard. They cannot see whether a sentence is ambiguous, whether a pronoun has a clear referent, whether paragraphs follow in a sensible order, or whether a term has been defined before it is used. A passage of fluent nonsense scores well. A clear explanation using necessary technical vocabulary scores badly.

This is easy to demonstrate. "The cat sat on the mat that the man that the dog that the boy owned bit made" is monosyllabic, short-sentenced, and scores as very easy. It is close to unparseable. Conversely "photosynthesis converts light into chemical energy" is four words of plain fact that any competent reader follows, and its syllable count drags the score down hard. The formulas are useful as a smoke alarm for bloated sentences, not as a judge of writing.

The number to aim for, and when to ignore it

Flesch scores map to audiences roughly as follows. Above 90 is very easy, at about a fifth-grade level. The 60 to 70 band is described as plain English and corresponds to eighth or ninth grade. The 30 to 50 band is difficult and typical of academic writing, and below 30 is very difficult, which is where insurance policies and legal contracts live.

For general web writing, 60 to 70 is the sensible target. Reader-survey work has long put average adult reading ability in developed countries at around a ninth-grade level, and a substantial minority of adults read comfortably well below that. Plain-language guidance in government and healthcare frequently specifies sixth to eighth grade for material intended for everyone, and some United States states set a grade ceiling in statute for insurance documents.

Aiming lower is not automatically better. Pushing a score above 80 usually means chopping sentences that belonged together, which removes the connective words — because, however, therefore, although — that carry the logic. The result reads as a list of assertions with the reasoning deleted, and it is harder to follow despite scoring easier. Short sentences uniformly applied also produce a monotonous rhythm, and rhythm is part of how a reader keeps their place.

The score should also change with the audience. Documentation for engineers is allowed to use the vocabulary engineers use; substituting a paraphrase for a precise term to please a formula makes the document worse and longer. What does not change with audience is sentence length. Nobody, at any expertise level, benefits from a forty-word sentence with three subordinate clauses.

What actually makes writing hard

If the formulas cannot see the real problems, it helps to name them, because these are the edits that improve a page whether or not the score moves.

Nominalisation is the most common. It is the habit of turning verbs into nouns and then needing a weaker verb to carry them: "we performed an evaluation of" instead of "we evaluated", "there is a requirement for" instead of "we need". Each one adds words, removes the actor, and makes the sentence flatter. Hunting for words ending in -tion, -ment and -ance finds most of them.

Stacked prepositions are the second. "The review of the report of the findings of the committee" forces a reader to hold four nested relationships at once. Three prepositional phrases in a row is a reliable signal that a sentence needs restructuring rather than shortening.

Undefined terms are the third and the most damaging, because they are invisible to the writer. A term you use daily reads as ordinary to you and as an obstacle to a reader meeting it for the first time. A formula counting syllables cannot tell whether a word is known, which is precisely why the Dale-Chall approach of checking against a familiarity list catches problems the others miss.

Sentence length variance is the fourth, and it is the one that formulas actively mislead on. A page of uniformly twelve-word sentences scores well and reads as a drone. Good prose alternates: a long sentence that develops an idea, then a short one that lands it. Because the formulas use the average, they cannot distinguish healthy variation from uniformity, and optimising for the average tends to destroy the variation.

Finally, structure sits above all of it. Headings that describe their content, one idea per paragraph, and the conclusion before the reasoning rather than after it do more for comprehension than any sentence-level edit. Readers scan first and read second, and a page that cannot be scanned is not read at all.

Keyword density is not what it was

Word-frequency tools sit next to readability tools and are frequently misused, so it is worth being direct about what they are good for now.

Keyword density — the share of words on a page that match a target term — was a genuine ranking input in the earliest generation of search engines, which largely counted term frequency. It stopped being one a long time ago. Modern retrieval works on embeddings and on how a document relates to a query semantically, and search-engine representatives have said plainly for over a decade that there is no target density to hit. Writing to a percentage produces the stilted repetition that quality systems are specifically trained to recognise.

What frequency analysis is still useful for is diagnosis rather than optimisation. It tells you what a page is actually about, which is occasionally not what you intended. If a page written to explain one thing has a different term as its most frequent, the emphasis has drifted. It also exposes unintentional repetition — the same phrase opening six paragraphs — which is a real prose problem that reads as carelessness.

The other honest use is coverage. If a topic has obvious subtopics and none of their vocabulary appears anywhere on the page, the page is probably thinner than it feels to write. That is a question about completeness, not about hitting a percentage, and the fix is to say more rather than to repeat more.

Questions