A readability score belongs to the sentence, not to the phrase

Why “this rewrite gains you +8 reading ease” is not a well-formed claim, measured on twenty real excerpts. 9 August 2026.

Editing tools like to attach a number to a suggestion. Replace due to the fact that with because and you are told the change is worth, say, eight points of Flesch reading ease. The number looks like a measurement. It is not one, and the reason is structural rather than a matter of calibration: the Flesch formulas take a whole text as their argument, and a phrase is not a text. The effect of a substitution is a function of the sentence you make it in. There is no context-free value to publish.

This note works that out three ways — by measuring phrases in isolation, by measuring the same swap inside real host sentences, and by deriving the shape of the dependence algebraically — and then shows that even the whole-text score has a free parameter most implementations never disclose. Every figure below was computed from the corpus reproduced in full at the end, so all of it is re-checkable.

The two formulas, exactly

Flesch reading ease and Flesch–Kincaid grade level are both functions of two ratios: words per sentence, and syllables per word.

FRE  = 206.835 - 1.015 * (words / sentences) - 84.6 * (syllables / words)
FKGL =   0.39  * (words / sentences) + 11.8  * (syllables / words) - 15.59

Three things follow immediately, and they drive everything else here.

1. Sentence count is in the denominator. Any edit that changes word count changes words-per-sentence, and the size of that change depends on how many words were already there. Removing four words from an eight-word sentence is a 50% change; removing four from a forty-word sentence is 10%.
2. The two terms pull against each other. Shortening a sentence lowers the first penalty and raises the second, because the filler words you delete are usually short. Deleting due to the fact that removes five words carrying 5 syllables, so the words you took out averaged one syllable each. The words left behind are longer on average than the ones removed, and syllables-per-word goes up.
3. Neither formula has a term for meaning. Reordering a sentence into nonsense while preserving word and syllable counts leaves both scores identical. The score is a proxy for effort, not for comprehension, and it is a proxy that can be optimised directly.

The corpus

Twenty short prose excerpts spanning academic, technical, casual, formal and mixed registers — 446 words in 41 sentences. Recomputed here with the syllable counter described below. Mean syllables per word across the corpus is 1.693; mean words per sentence is 10.88.

The range is the first useful result: -55.8 to 95.2 on a scale usually described as running from 0 to 100. Reading ease is not bounded. The lowest-scoring excerpt is a single 18-word sentence at 2.889 syllables per word, and it lands at -55.8; the highest is an early-reader excerpt of 4 short sentences at 1.263 syllables per word, which lands at 95.2. Any product that renders this number as a percentage bar is lying to its user at both ends.

Measurement 1: the phrase on its own

The obvious way to get a per-phrase delta is to score the phrase as though it were a one-sentence document, score the replacement the same way, and subtract. Here is what that produces for six standard plain-English swaps.

The results run from -80.6 to +169.2, and three of the six are negative — the formula says these standard improvements make the text harder. The mechanism is plain once you look at it. due to the fact that is five one-syllable words, so its syllables-per-word is 1.0, the best score the second term can produce; because is one word of two syllables, ratio 2.0. Scored in isolation, filler beats concision every time, because filler is made of short words. Meanwhile utilize to use gains +169.2 points, which is not a measurement of anything either — it is one three-syllable word against one one-syllable word with the words-per-sentence term held at 1.

These numbers are arithmetically correct and semantically worthless. They are what you get when you evaluate a function outside its domain.

Measurement 2: the same swap inside real sentences

The honest version of the question is: what happens to a real sentence when you make the swap in it? Below, each phrase is placed as a trailing clause on five host sentences drawn from the corpus, with host length as the variable. The swap is the only difference between the two variants of each sentence.

due to the fact that → because — spread 15.0 points

a large number of → many — spread 10.3 points

utilize → use — spread 6.8 points

Three things to take from this.

The spread is large relative to the effect. For due to the fact that the delta ranges from -18.1 to -3.1 across five host sentences — a 15.0-point spread on an effect that a tool would report as a single constant. Whatever constant you pick is wrong for most sentences.

The sign is not stable across metrics. Still on due to the fact that, the grade-level delta runs from -0.6 to +1.5. In the same corpus, with the same swap, one formula says the text got a year and a half harder and another says it got easier. They disagree because 1.015 and 84.6 weight the two ratios differently from 0.39 and 11.8, so when the ratios move in opposite directions the two formulas can resolve the conflict differently.

The classic swap scores as a regression. Cutting due to the fact that to because lowers reading ease in all five hosts. This is the clearest possible demonstration that the formula is not measuring what the advice is about. The advice is right: the short form is better prose. The formula prefers the long form because the long form is made of one-syllable words. If you tune prose against the score you will be pushed towards padding.

By contrast utilize to use gains between +5.3 and +12.1, because that swap removes syllables without removing words — it moves only the second term, so the two terms do not fight. Substitutions that change word count are the unstable ones. Substitutions that change only syllable count are comparatively well behaved. That is a usable rule, and it falls straight out of the algebra.

Measurement 3: the dependence, derived

You do not need the experiment to know the shape of the result. Take a one-sentence text of W words and S syllables. A substitution removing w₋ words and s₋ syllables and adding w⁺ and s⁺ gives a new score:

FRE(W, S)  = 206.835 - 1.015*W - 84.6*(S/W)          # one sentence, so w/sent = W
ΔFRE       = FRE(W - w₋ + w⁺, S - s₋ + s⁺) - FRE(W, S)

The S/W term makes ΔFRE a rational function of W, not a constant. It has no limit-free value: as W grows the delta decays towards the contribution of the words-per-sentence term alone. Holding syllables per word at the corpus mean of 1.693 and applying the due to the fact that -> because swap (5 words and 5 syllables out, 1 word and 2 syllables in):

From -80.5 to -1.7 over a plausible range of host lengths, and it does not stop there: as W grows without bound the syllables-per-word term returns to the corpus mean and the whole delta converges on -1.015 * (w⁺ - w₋) = 4.06, so the sign eventually flips. One substitution, one metric, and the answer runs from strongly negative through zero to mildly positive depending only on how long the surrounding sentence is. This is the whole argument in one table: the quantity a per-phrase delta claims to be does not exist, because the function it would have to be constant over is not constant.

The free parameter nobody prints: syllable counting

English syllable counts are not derivable from spelling. Every implementation either ships a pronunciation dictionary or guesses with a heuristic, and the heuristics disagree. The counter used throughout this note is:

1. lowercase the word and strip everything that is not a-z
2. count maximal runs of [aeiouy]           -> n
3. if the word ends in 'e', is not '-le', '-ee' or '-ye', and n > 1: n = n - 1
4. return max(n, 1)

Step 3 is the silent-e rule and it is the only judgement call in there. Dropping it — which is what the simplest published heuristics do — changes nothing about the text and yet moves the score:

Mean absolute difference across the corpus is 6.8 points; the worst case is 17.6. That is larger than most of the per-phrase deltas anyone quotes. Two tools can score the same paragraph, both be correctly implementing Flesch, and differ by more than the edit you were considering. It follows that reading-ease figures are only comparable within one implementation, and that a target such as “keep it above 60” is meaningful only once you name the tool.

What the number is actually good for

None of this makes readability scoring useless. It makes it a different instrument from the one it is usually sold as.

• Score documents, not edits. The formulas are stable over a few hundred words because the ratios stop swinging. Over one clause they are noise.
• Use it as a detector, not an objective. A section scoring far below the rest of a document is worth reading again; that is a real signal. Optimising the number directly rewards short words and short sentences regardless of whether the writing improved, and, as shown above, will actively push you to keep filler.
• Track direction within one tool, never absolute values across tools. A ten-point move measured twice by the same counter means something. The same ten points compared between two products may be entirely the silent-e rule.
• Keep the phrase list; drop the phrase deltas. Advice like prefer because to due to the fact that, or use to utilize, is sound editing independent of any formula. It is sound for reasons the formula cannot represent. Attaching a fake number to it does not strengthen it, and in the cases measured here the number actively contradicts it.

The general failure this is an instance of: a metric defined over one unit of analysis gets reported at a smaller unit, and it acquires a spurious precision on the way down. Reading ease is defined over a text. Report it over a text.

I keep this kind of thing in mind because it is the substance of the tool I work on — rewriting and clarity feedback for non-native English writers at belikenative.com — where the temptation to put a confident number next to every suggestion is constant and mostly worth resisting.

Appendix: the corpus in full

Reproduced so every figure above can be recomputed. Twenty excerpts, 446 words.