How Much Do You Need to Understand? The 50/70/90/98 Question

Ask a learner how much of a book they understood and you will get a percentage, offered with confidence and meaning almost nothing. “About seventy per cent” is not a measurement. It is an impression of how the reading felt, and impressions of comprehension are known to be generous — readers routinely rate their understanding well above what a test of the same passage produces.

There is a real number underneath, though, and it is worth knowing precisely, because the relationship between it and how much you actually understand is not a straight line. Three percentage points in the wrong place turn a book you can read into a book you cannot.

Two numbers that get confused

Almost every argument about difficulty is really a collision between two different quantities.

Coverage is the proportion of the running words on the page that you already know. It is a property of the text and your vocabulary together, and it is countable: mark the words you do not know on a page of three hundred, and you have your figure to within a point or two.

Comprehension is the proportion of the content you take away. It is a property of the reading, it depends on far more than vocabulary, and it can only be measured by asking you questions afterwards.

Coverage is the input; comprehension is the output; and the useful research question of the last thirty years has been what the curve between them looks like. When people say “the ninety-eight per cent rule”, they mean a coverage figure. When they say “I understood most of it”, they mean a comprehension impression. The two are quoted in the same breath constantly, and they are not the same thing.

The ladder

Take a page of about three hundred words — a normal page of a printed book — and read the ladder as “words you do not know, per page”.

• 80 per cent — one unknown word in five, sixty a page. This is not reading in any sense the word normally carries. Nothing can be inferred, because the material you would infer from is also unknown. Below roughly this line, studies stop finding anything worth measuring.
• 90 per cent — one in ten, thirty a page. The figure that sounds respectable and is not. In the best-known experiment on this, learners reading at ninety per cent coverage almost never reached adequate comprehension unassisted; a handful did, and they were the exception. Three unknown words in every second line is a text you decode.
• 95 per cent — one in twenty, fifteen a page. The threshold often proposed for reading with support. A minority of readers reach adequate comprehension here on their own; most need something — a gloss, a dictionary, a teacher, prior knowledge of the plot. The reading is possible and it is work.
• 98 per cent — one in fifty, five or six a page. The figure for unassisted reading. At this density most readers, though still not all, reach adequate comprehension with nothing but the book. This is where reading stops being a task and becomes an activity.
• 100 per cent. Comfortable, fast, and inert: pleasant reading that teaches you nothing new. Useful for speed, useless for growth.

The two things to take from the ladder are that the interesting range is narrow — everything happens between 90 and 100 — and that the curve inside it is steep. The distance from 95 to 98 is three percentage points and it is the difference between fifteen interruptions a page and five.

Why the curve is so steep

Percentages of running words feel like they should behave gently. They do not, because you do not read words, you read sentences, and the unit that has to survive is the sentence.

Take a sentence of twenty words, which is ordinary prose. At 98 per cent coverage you expect an unknown word in about every second or third sentence, and the sentences around it are clean, so the context needed to place it is intact. At 95 per cent you expect one in every sentence. At 90 per cent you expect two in every sentence — and two unknowns in one clause do not cost twice as much as one. They cost far more, because the first is now to be inferred from a context that contains the second.

That interaction is the whole explanation of the shape. Guessing an unfamiliar word from its surroundings is the engine of acquisition through reading, and the engine requires the surroundings to be known. Coverage does not degrade comprehension gently; it degrades the mechanism that was holding comprehension up.

What “adequate comprehension” meant

There is a second threshold hidden inside the first, and it is the reason the figures sometimes look arbitrary. The studies that produced 95 and 98 had to decide what counted as understanding a text, and they generally set the bar at a comprehension score of around 70 per cent on questions about the passage — with 80 used when a stricter criterion was wanted.

So the famous rule, spelled out in full, reads: to reach about 70 per cent comprehension of a text without help, you need to know about 98 per cent of its running words. Both halves are needed. Move the comprehension criterion and the coverage figure moves with it, which is why 95 and 98 both appear in the literature and neither is wrong.

It also explains why the ratio feels so unfair. Knowing forty-nine words in every fifty buys you seven answers out of ten. Language does not offer a better exchange rate than that, and a method that promises one is measuring something else.

Why “I got the gist” is not a measurement

Gist is reconstructed, not received. A reader who knows that stories have plots, that this one is about a duel, and that the man in the second paragraph is angry will produce a plausible account of a page in which the actual sentences went past in a blur. The account may even be right. It was not, however, obtained from the text, and nothing was acquired in the process, because acquisition needs the sentence to have been understood at the moment it was read.

This is why self-assessment runs high and why it runs high most in exactly the situation where it matters — a text slightly too hard, where world knowledge is doing the work quietly. The tell is specificity. Ask what happened on the page and a comprehending reader gives you a who and a what; a gist reader gives you a theme.

The useful instrument is the countable one. Coverage takes ninety seconds to measure and does not lie to you, which is why the practical version of this article is a counting exercise performed on the page rather than an estimate performed on your feelings. That procedure belongs to choosing a text, where it is set out in full.

The thresholds are not fixed

A coverage figure describes a text and a reader. Several things move the line, and they are worth knowing because they explain why a book at 95 can be entirely readable in one situation and impossible in another.

• Support moves it down. A gloss at the foot of the page removes the unknown word from the arithmetic, at a cost of about two seconds instead of thirty. A footnoted text at 95 per cent behaves for the reader roughly like an unfootnoted text several points higher — which is the entire design principle behind graded and annotated editions.
• Prior knowledge moves it down. If you know the story, the subject, or the writer’s habits, you are supplying context that the text would otherwise have to supply, and the effective coverage rises without a word changing.
• Purpose moves it. Reading for pleasure at speed needs the high figure. Studying one paragraph closely needs no threshold at all; you can work at 70 per cent coverage for a page, and it is a different activity with a different name.
• The language moves it. The figures come mostly from English. In a heavily inflected language the cost of an unknown word is higher — the form on the page may not resemble the form you would look up — so the same coverage buys less comprehension.
• Word families cut both ways. Coverage is normally counted in word families, so govern, government and governing count once. That assumes you can see through the morphology, which a beginner often cannot.

What the figures mean for a vocabulary size

Converted into words known, the ladder gives the shape of the road. Something like the commonest 3,000 word families of a language will give you about 90 per cent coverage of general text; 5,000 gets you into the region of 95; and unsimplified adult fiction generally wants 8,000 to 9,000 for 98. Those are approximations, they vary by language and by text, and the gaps between them are years.

Which is the real argument for graded material, stated as arithmetic rather than as advice: it is a way of having the 98 per cent experience at 3,000 words instead of waiting until you have 9,000.

This article deliberately stops at the diagnosis. If you have counted a page and the book in your hands is well under the line, there is a series of things to do about it — and the behaviour that keeps the unknown words cheap in the first place is a skill of its own.