What published novels actually do, measured rather than assumed.
21,956 public domain books scanned on the tool's default settings. The datasets below are free to download and quote under CC BY 4.0.
Why this exists. Almost everything written about how many characters a novel should have, how long a chapter runs or how much invented vocabulary is too much is repeated rather than measured. These are measurements. The corpus is public domain fiction, so no customer manuscript is or ever could be part of it: nothing in this product's storage holds manuscript text or extracted terms, and its own test suite enforces that.
Every figure is generated, never typed. Each study is recomputed from the corpus whenever it grows, so a number on an article and the file behind it cannot drift apart. Three of them read the tables below; the series and drift tables name the invented words themselves and are withheld. Where a page and a dataset disagree, the dataset is right.
What it cannot tell you. The corpus is weighted towards books first published before 1930, because that is what is legally usable end to end. The comparison treats the rates it rests on as stable across eras, and places length alone against a second population of novels published since 1930. The derived benchmarks are our own estimates. The methodology states that assumption and every limit on one page.
The corpus cut five ways. genre, length, British or American spelling, readership, English or translated. The benchmark pages are the reading of those slices, with the number of novels behind every figure and no figure published from a group too small to support one.
The five studies
Each one is the reading of the corpus, and each names the run its figures came from.
| Study | The question | The headline |
|---|---|---|
| What Published Novels Get Wrong About Their Own Words | How much of a finished book's own vocabulary is spelled two ways? | 99.9 per cent of novels spell at least one of their own terms more than one way; the median novel does it 22 times. |
| What a Thousand Novels Do With Their Characters | How many people does a novel name, and when do they arrive? | Across 12,602 public-domain novels, the median novel names 32 characters (middle half 22 to 45). |
| What Happens to a Series Between Books | Do two books in one series agree with each other? | 6.9 per cent of the 8,000 pairs of novels that share a cast spell at least one of their shared invented words differently from each other. |
| Where a Novel Loses Track of Its Own Words | How big is an inconsistency, and where in the book does it sit? | Odd spellings are spread evenly through a book: 49.1 per cent fall in its second half and the median one sits at the 49.3 per cent mark. There is no sloppy second half in this corpus, which is the opposite of what the folklore says. |
| Which English Do Published Novels Actually Keep? | British or American, -ise or -ize, `Mr.` or `Mr`? | -ise against -ize is a separate question and behaves like one: 72.5 per cent of 11,817 novels are mostly -ize, including 38.7 per cent of the novels that keep British spelling everywhere else. |
The tables
Plain CSV and JSON Lines, openable in a spreadsheet. Row counts and sizes are read off the files themselves.
| File | One row is | Rows | Size | Read on |
|---|---|---|---|---|
| measurements.csv | one scanned book | 22,025 | 5.4 MB | What Published Novels Get Wrong About Their Own Words |
| chapters.jsonl | one book | 16,378 | 8.0 MB | What a Thousand Novels Do With Their Characters |
| facets.csv | one measured book | 21,956 | 2.5 MB | What Published Fiction Does, by Genre |
| modern_lengths.csv | one published work | 14,779 | 3.7 MB | What Published Novels Get Wrong About Their Own Words |
| style.csv | one novel | 12,602 | 1.9 MB | Which English Do Published Novels Actually Keep? |
Citing this. CC BY 4.0 asks for attribution and nothing else. "The Novel Kit, The Novel Manuscript Benchmark, 2026-10-02" with a link to this page is enough, and a link to the specific file is better.
Press and research enquiries are welcome, including requests for a cut of the data that is not published here. The corpus is larger than the tables, and the published rows are de-identified, so a cut that is not published here is a request rather than a download.
Measure your own manuscript against these
The same scan, on the same settings. Free to preview, no sign-up required.