What Published Novels Get Wrong About Their Own Words

We scanned 21,956 published books, cover to cover, with the same tool this site sells. That's 1334 million words of finished, professionally edited, printed books. In it we found 483,315 places where a book spells one of its own terms more than one way. Of the 12,602 full length novels among them, sixteen came through clean.

What was measured, and why published books

Every book was scanned on default settings, so a figure here is directly comparable with a figure from a scan of your own manuscript. The corpus is public domain books from Project Gutenberg, which matters for two reasons. It means no customer manuscript was involved, and it means every book had already been through whatever editing and proofreading its publisher applied before it reached print.

That second point carries the whole argument. These are not drafts. They are the finished article, and the error rate below is what survived a professional process.

What we found

Stated as rates rather than totals: 99.9 per cent of novels carry at least one term spelled two ways. 85.5 per cent carry ten or more. 13.4 per cent carry fifty or more. And 2.5 per cent of published novels carry over a hundred.

Sixteen books in 12,602 came through with none at all, and they prove the rule rather than breaking it. The Turn of the Screw is among the shortest full length works in the corpus at 42,239 words and has a few dozen terms in all; Washington Square and The Two Wives are short, with small casts, and so are most of the rest. They are the books with the least to keep track of, and they are the only ones that kept track of all of it.

Capitalisation is the slip, not hyphenation

Nearly seven in ten of the inconsistencies we found are capitalisation: a term written Hive in one chapter and hive in another, or the Order and the order. That ambiguity is rarely a typing error. It's a decision the writer made twice, months apart, and it reads to a reader as two different things.

Hyphenation is the smaller share at 147,928, but it is spread almost as widely: 93.2 per cent of novels carry at least one term written both with a hyphen and without. Blood-magic and bloodmagic and blood magic are three spellings of one word, and spellcheck has no opinion about any of them.

Why it happens to careful writers

The median novel carries 297 invented terms, names and places, and mentions each one about eight times across 73,000 words. That's the shape of the problem. A term you use four hundred times is a term you cannot get wrong. A term you use eight times, first on page nine and again on page two hundred and forty, is one you have to remember, and remembering is what nobody can do reliably across a book.

Front loading makes it harder. Across the corpus, 44.2 per cent of every term a novel will ever use is introduced in its first three tenths. The vocabulary is set early, used sparsely, and then has to stay consistent for another sixty thousand words.

Half of it, 48.5 per cent, appears exactly once in the whole book. A term used once cannot be checked against itself, which is why a list of the terms that appear once is one of the reports a scan produces.

What one novel has to keep straight

Thirty two names, twenty two places and a couple of hundred other terms is not an unusual load. It is the ordinary load, and it is more than a person can hold in their head while also writing the book.

What this means for a manuscript in progress

Every number above comes from a book that was finished, edited by someone paid to edit it, proofread and published. Whatever process those books went through, it left a median of twenty two inconsistent spellings per novel behind.

That is the floor, and it is the part we can prove. A manuscript that has not had any of that done to it cannot reasonably be cleaner.

How much worse is a fair question and we cannot answer it with a measurement, because we never store a manuscript and so have no corpus of drafts to compare against. What we can do is show the arithmetic and let you move the assumption. If a professional edit removes nine of every ten of these, then the twenty two that survived in a 90,000 word novel started as roughly 220. If it removes only half, the same draft started at about forty four. Neither number is a finding, both are what the measured rate implies under a stated assumption, and even the harsher of the two is a great deal more than most writers expect to be carrying.

The useful question is not whether they are there. It's whether they have been found.

What this cannot tell you

Project Gutenberg is overwhelmingly pre 1929, so this corpus measures books edited by the standards of their day rather than by a modern house style guide. Some of what we count as an inconsistency was a live convention at the time, and typesetting practice varied more than it does now.

Two limits, stated plainly. The scanner reads what a book does, not what it meant, so a term deliberately capitalised in one context and not in another is counted as a variant when a human editor might allow it. And the figures are medians across more than a thousand books rather than rules about any one of them; one novel in the corpus carries 686 variants and is not a badly written book.

A practical next step

The reports that produce the figures above are the reports you get on your own manuscript. The capitalisation checker and the hyphenation checker are the exact tests that found the 483,315 spellings, run against your book. The similar names and aliases check is the test behind the confusable names figure, and unexplained terms is the one behind the 1,998,535. Where a spelling sits inside a book, and how large each one is, is a study of its own.

Nothing here involves artificial intelligence and no manuscript is stored. The scan processes your file transiently, produces the reports and never stores it, which is why a study like this uses public domain books and could not have used anybody's work in progress. You can run a scan of your own and read your numbers against these.

Sources checked

All figures come from the Novel Manuscript Benchmark, a public dataset of public domain books that is added to over time. The per book measurements and the corpus manifest are published below as CSV files, free to download and free to quote, and every figure on this page is regenerated from them whenever the corpus grows. The numbers quoted here are from a run of 21,956 books and 12,602 full length novels, measured in September 2026; a later run will move them, and the dataset rather than this page is the authority.

Frequently asked questions