warmprose Find the sentences that sound machine-made.Rewrite them warm, in your own voice.
warmprose research · 13 November 2026

How long are real sentences?

345,000 words of published English, six genres. Short sentences are rarer than the advice suggests.

Writing advice keeps returning to a number: keep sentences under 20 words. We wanted to know what published English actually does, so we measured the same things across six kinds of writing.

Corpus: 575 articles, 15,428 sentences, 345,000 words. Each article contributes a 600-word window taken from its middle, so the opening paragraph of a piece cannot skew the result.

55%of all 15,428 sentences are 20 words or fewer
22.0median sentence, longest genre (News)
18.0median sentence, shortest genre (Company blog)
575articles · 345,000 words · six genres

Sentence length by genre

WORDS PER SENTENCE · 10th — MEDIAN — 90th PERCENTILE News 22.0 9 39 Opinion 19.0 8 32 Company blog 18.0 9 30 Government 21.5 11 41 Fiction 21.5 8 47 Encyclopedia 21.0 11 33
Median sentence length per article (dot), with the 10th and 90th percentiles across articles. Genre medians run from 18.0 words (Company blog) to 22.0 (News).

The spread between genres is real but smaller than the spread inside any one of them: in every genre the 90th-percentile sentence is more than twice the 10th. A single word-count target has to survive that range.

How much prose sits under 20 words

SHARE OF ALL SENTENCES BY LENGTH News 49% 25% 26% Opinion 60% 23% 17% Company blog 60% 25% 15% Government 50% 25% 26% Fiction 56% 17% 27% Encyclopedia 52% 28% 20% ≤20 words 21–29 30+ words
Every sentence in the corpus, pooled by genre. Company blog leads at 60% of sentences at 20 words or fewer; News trails at 49%.
What this is not

This does not say long sentences are better, or that the 20-word habit is useless when revising. It measures what published, edited English looks like — not what it should look like. Nor is it an authorship check: nothing here estimates whether a text was written by a machine.

All measures

Sentence length, per article

Median and mean of each article’s own sentence lengths, then the median across articles — so a single long article cannot dominate.

MeasureUnitNews
n=100
Opinion
n=75
Company blog
n=100
Government
n=100
Fiction
n=100
Encyclopedia
n=100
Median sentence lengthwords22.019.018.021.521.521.0
Mean sentence lengthwords22.920.020.425.823.921.6
10th percentilewords9.08.09.011.08.011.0
90th percentilewords39.032.030.041.047.033.0
Articles measured10075100100100100
Words measured60,00045,00060,00060,00060,00060,000

Shares and rates, pooled over the whole genre

Every sentence and every word of the genre counted once, so long articles weigh in proportion to their length.

MeasureUnitNews
n=100
Opinion
n=75
Company blog
n=100
Government
n=100
Fiction
n=100
Encyclopedia
n=100
Sentences ≤20 words%49.360.260.449.556.152.2
Sentences ≥30 words%26.016.714.925.727.219.7
Passive constructions (approximate)per 1,000 words7.87.93.014.56.415.0
Dashes (— or –)per 1,000 words3.24.72.30.82.93.9
Three-item listsper 1,000 words5.16.410.57.36.25.2
Connectorsper 1,000 words0.90.90.61.11.21.2

Method

Sources — news: BBC, The Guardian, NPR, Al Jazeera, CBC, Sky News, France 24, Euronews RSS; opinion: The Guardian Comment is Free, The Conversation, Project Syndicate, The Independent, The New York Times opinion, The Washington Post opinions, Los Angeles Times opinion, The Sydney Morning Herald opinion; company blogs: HubSpot, Buffer, Zapier, Shopify, Hootsuite, Moz, Ahrefs, Semrush, Intercom, ConvertKit, Copyblogger; government: GOV.UK publications; fiction: Project Gutenberg English-language books; encyclopedia: Wikipedia Featured articles.

Sampling: articles are taken in source order until 100 per genre; each contributes a randomly placed 600-word window placed at random inside its middle 70% (fixed seed 20261113, offset recorded per article in the CSV as window_start, so any row can be re-cut). Articles with fewer than 700 words are dropped. Navigation, captions, standfirsts and boilerplate are stripped before counting.

Measures: sentences split on terminal punctuation inside a paragraph; words are whitespace tokens. Passive constructions are matched as be + past participle patterns rather than parsed, so that figure is an approximation and should not be used to rank genres closely. Dashes count em (—) and en (–). Three-item lists match X, Y and Z and X, Y, and Z. Connectors are these twelve strings, counted case-insensitively: however, moreover, furthermore, therefore, additionally, in addition, consequently, nevertheless, thus, hence, overall, meanwhile (plus subsequently).

Limitations: one 600-word window per article; six genres, with opinion at 75 articles because only that many qualified from available feeds; English only; fiction is public-domain (pre-1930) and therefore older in style than the rest. The corpus is a snapshot, not a trend.

Reuse this

One row per article: genre, source, URL, window start, word and sentence counts, every measure.

Download the CSV (575 rows)

Archived release with a DOI on Zenodo — the same data plus the method notes and per-genre files, licensed CC BY 4.0.

The same files are also published on Kaggle: English Writing Style Baselines.

Suggested citation: warmprose, “How long are real sentences? 345,000 words measured across six genres”, 13 November 2026, warmprose.com/research/english-style-baselines. DOI: 10.5281/zenodo.24199354