Comprehensible Input in French: The 15-Page Test

Comprehensible input works, and 98% is the right target. What nobody tells you is that 98% is an average over a whole book, and you do not read books in averages. You read them one page at a time, and the pages vary far more than the number suggests.

We measured that variation. Eleven French classics, 928,000 words, scored against Lexique 3.83, the reference frequency database of French. Every word in every text was tagged as known or unknown for a reader with a given vocabulary size, in reading order, so we could look at the spacing rather than just the total.

What does 98% coverage feel like on the page?

It feels like an unknown word every forty words or so, which is roughly one every page-and-a-half lines of a paperback.

Here is the actual spacing for a reader who knows the 11,338 most frequent lemmas, the vocabulary that delivers 98% of general written French.

The right-hand column is the one worth staring at. The longest unbroken stretch without a single unknown word, in the easiest book here, is 771 words. That is roughly two and a half pages of uninterrupted reading. In Le Tour du monde en 80 jours, the best run in the entire novel is 127 words, less than half a page. Same reader, same vocabulary, completely different experience of what reading is.

Are unknown words clustered or evenly spread?

Close to evenly, and this surprised us.

The intuitive story is that hard words come in bursts: a paragraph of nautical terms, then a stretch of plain narration. If that were true, the typical gap between unknown words would be much shorter than chance predicts, because most gaps would be tiny and a few would be huge.

It is barely true. Across the eleven books, the median gap between unknown words runs from 0.71 to 1.01 times what pure random placement would give, with a median of 0.87. In other words the typical gap is about 13% shorter than chance, and the long gaps are about 3% longer. That is a real effect but a small one.

This matters because it kills a piece of common advice. If unknown words genuinely arrived in bursts, you could push through a hard patch and expect relief. Within a page, you cannot. The spacing is close to random, so a dense paragraph does not predict an easy one behind it.

Why is one page harder than the next?

Because the unevenness lives at a bigger scale than the paragraph.

We split each book into 200-word slices, about half a paperback page, and compared how much coverage varied between slices with how much it would vary if words were placed at random. Random placement already produces variation, which is exactly the trap: some pages fall below any threshold you pick purely by chance.

The observed variation is larger. In every book, at every vocabulary size we tested, the page-to-page spread exceeds the random baseline by a factor of 1.44 to 2.19, median 1.64. Thirty-three measurements, no exceptions. French novels are systematically about 60% more uneven than random word placement would make them.

The consequence is concrete. Les Misérables averages 95.9% coverage, comfortably in supported-reading territory, and still contains half-pages at 78%. Madame Bovary averages 95.5% and drops to 80%. Those stretches are where a new setting arrives with its own furniture: the Waterloo chapters, the agricultural show, the convent. Le comice agricole occupait toute la place - “the agricultural show filled the whole square” - and with it come thirty nouns you have never needed before.

That is what the wall feels like. It is not that your French collapsed on page 140. It is that page 140 is a different book from page 139.

How many pages does it take to judge a book?

Fifteen, and almost nobody reads fifteen before deciding.

The overdispersion above has a direct arithmetic consequence. To estimate a book’s true average coverage to within one percentage point with 95% confidence, you need a median of 4,600 words of sample, about 15 paperback pages. If French prose were as even as random placement, 9 pages would do. The unevenness costs you the difference.

Sample less and the estimate is loose in a way that makes the test useless. A single 300-word page fixes the book’s coverage only to within 4.2 points either way, a spread of 8.4 points. For a book whose real figure is 95%, one page can easily read like 91% or 99%. Those are not fine distinctions. They are the difference between abandoning a book and enjoying it.

What is the corrected page test?

We published a one-page test in our own guide to French books for beginners: read a page, count how many times you stop, more than twice and the book is too hard. Having now measured it, both halves of that rule are wrong, and the honest thing is to say so.

The threshold was far too strict. On a 300-word page, at the 98% level, here is what you actually meet:

Six unknown words on the easiest page of the easiest book. A rule that rejects anything above two stops rejects the entire canon, including books that are genuinely at a strong B2’s level.

The sample size was also too small, as the 4.2-point interval shows. So the corrected test is this. Read three pages from different parts of the book, early, middle and late, and take the median number of words you could not have guessed. Six or fewer means you are near 98% and can read the book for pleasure. Around ten to fifteen means it is study material, workable with a dictionary. Above twenty, put it back and come to it in six months.

Three pages narrows the interval from 4.2 points to about 2.4, which is enough to make the decision. Taking them from different parts of the book is the part that does the work, because the variation is between scenes, not within them.

What should you do when a page turns hard?

Keep going, and lower your standard for that page only.

Since the difficulty is bunched by scene, a hard stretch is genuinely temporary in a way that a hard book is not. The nouns of a battlefield or a courtroom arrive together and leave together. Reading through a 78% page inside a 96% book costs you that page and buys you the next twenty.

The failure mode is to conclude from one bad page that the book is beyond you and quit. The measurement says that inference is unsound: one page tells you almost nothing about a book, in either direction.

What does not work is choosing a main book below 95%. At that level you meet an unknown word every twenty words, several per paragraph, and the lookups eat the reading speed that extensive reading depends on. There is no amount of persistence that fixes a text that is genuinely too hard, which is a different problem from a hard page.

This is the whole argument for graded material at the stages where native prose does not yet fit. A graded reader engineers the 98% line instead of leaving it to a book written in 1862 for adults who already spoke the language. Our own 61 levels from A1 to C1 exist for that reason: the coverage target moves up with you, rather than you guessing which classic happens to sit at your level this month.

Two limits on the numbers. Everything here is public domain, so the modern titles that dominate beginner lists are absent, and we will not estimate figures we could not compute. And coverage counts words, not understanding. Knowing 98% of the words in a sentence of Flaubert is a necessary condition for reading Flaubert, not a sufficient one.