The average is the least useful number on the page
Mean sentence length is the figure every readability formula leans on, and it is the one that hides the most. Take a page averaging twenty words. That could be forty sentences all sitting between eighteen and twenty-two, or it could be twenty sentences of eight words alternating with twenty of thirty-two. The mean is identical. The reading experience is not remotely the same, and only one of them has a problem worth fixing.
So the panel leads with the mean and then immediately qualifies it three ways. The median tells you where the middle sentence actually sits; when it drops well below the mean, a small number of very long sentences are doing the pulling. The standard deviation puts a number on the spread. The coefficient of variation — the standard deviation divided by the mean — makes that spread comparable between a technical document averaging twenty-eight words and a children's story averaging nine.
Reading the coefficient of variation
| CV | What it usually means | What to do |
|---|---|---|
| Under 35% | Sentences are close to uniform in length | Look for two adjacent sentences to merge, and one to cut to a fragment |
| 35% to 75% | The range most edited prose falls in | Nothing structural; read the long ones on their own merits |
| Over 75% | Strong alternation, or one huge outlier | Check whether the outlier is deliberate or a run-on |
These bands are descriptive rather than prescriptive. They come from what edited prose tends to look like, not from a study establishing that readers prefer a particular figure. A tightly-specified procedure or a set of numbered steps will legitimately sit under 35% and should stay there, because uniformity is the point in a checklist. Narrative and argument are where a low CV is a symptom.
The flat-run detector, and why it is separate from the CV
A document can have a healthy coefficient of variation overall and still contain a stretch of eleven consecutive sentences all landing between seventeen and twenty words. The variation is elsewhere, in the introduction and the conclusion, and the middle of the piece plods. That is what the flat-run search is for: it slides a window across the sentences in document order and finds the longest unbroken stretch where every sentence sits within your tolerance of every other in the stretch.
Three or four sentences in a row of similar length is normal and often deliberate — parallel construction relies on it. Once a run passes five, readers start to notice a rhythm without being able to name it, and past eight it becomes the dominant impression of the passage. The fix is small: merge two of the run into one longer sentence with a semicolon or a subordinate clause, or cut one down to a four-word fragment. Either breaks the pattern and neither costs you any content.
What counts as a sentence here
Sentence boundaries come from the same splitter used across these writing tools. A period ends a sentence only when whitespace or the end of the text follows it and the next word does not start lowercase; a lone period is also ignored after a single letter, which handles initials, e.g., i.e., a.m. and U.S., and after a list of known abbreviations covering titles (Mr, Mrs, Dr, Prof), company forms (Inc, Ltd, Co, Corp), reference terms (Fig, Vol, vs, etc, al) and abbreviated months and weekdays. Decimals such as 3.5 are safe automatically, because no whitespace follows the point. Ellipses split only when the following word is capitalised.
Line breaks always end a sentence. That is the right call for prose but it means a bulleted list registers as a series of short sentences, and a heading registers as a very short one. If your draft is heavily structured, the histogram will show a spike in the one-to-five-word bucket that reflects the formatting rather than the writing. Strip the headings and bullets out before analysing if you want a clean read on the paragraphs.
Length is not the same thing as difficulty
A thirty-five word sentence built from three parallel clauses with clear signposting can be easier to follow than an eighteen-word sentence with a nested relative clause and two abstract nouns. The threshold on this page exists because length correlates with difficulty on average, not because thirty words is a defect. When the tool lists your longest sentences in full, read each one and ask a single question: does it hold one idea, or has it accumulated three? The ones holding one idea can stay at any length. The ones that accumulated are what you split.
For the formula-based view of the same text, the readability score checker converts this distribution into Flesch and grade-level numbers. For counts of words, characters and bytes, see the text statistics tool, and if the issue turns out to be repeated wording rather than rhythm, word frequency will find it.
Questions people ask
What is a good average sentence length?
There is no defensible universal answer, which is why the tool does not offer one. Contemporary general-audience nonfiction tends to land somewhere between fifteen and twenty-five words; instructions and safety material run shorter by design; legal and academic writing runs longer for reasons that are partly necessary and partly habit. The number worth watching is your own across drafts, and the shape of the distribution rather than its centre. A draft that moved from a CV of 30% to 55% got more readable in a way no average would show.
Why does the median differ so much from the mean?
Because sentence length has a long right tail. Most sentences cluster in a fairly narrow band and a handful run very long, and the mean is dragged toward those few while the median stays with the crowd. A large gap between the two is a reliable sign that a small number of sentences are doing something unusual, and those are exactly the ones printed in full at the bottom of the page. If the median sits well above the mean instead, which is rare, you have a scattering of very short fragments pulling the other way.
Does the flat-run finder count paragraphs?
No, it works purely on consecutive sentences in document order, and it does not care whether a paragraph break falls inside the run. That is deliberate: the monotony a reader feels comes from the sequence of sentences, and a paragraph break does not reset it. If you want the paragraph-level view of how a piece is joined together, the transition and connective checker is the tool for that.
My bulleted list is skewing the histogram.
It will. Every line break is treated as a sentence boundary, so each bullet becomes its own short sentence and a list of ten items adds ten entries to the one-to-five-word bucket. That is correct behaviour for counting but unhelpful for judging prose rhythm. Delete the lists and headings from the pasted copy and run it again; the difference between the two runs also tells you how much of your document is structure rather than prose.
Is anything sent to a server?
No. The splitting, counting and histogram all happen in JavaScript in this page, on your machine, with no request going anywhere. You can disconnect the network and it behaves identically, which is the simplest way to confirm it for yourself.