How this test works
This test shows you a bit over a hundred words and hands back a number in the tens of thousands. Here is how it gets from one to the other, and how much of it to believe.
You cannot ask about every word
English has somewhere north of a hundred thousand words in ordinary circulation, and no one is going to sit through a hundred thousand checkboxes. So the test does what a pollster does: it asks about a small sample and reasons outward from it. Show someone a handful of words drawn fairly from a slice of the language, count how many they know, and you can say something about the whole slice.
The trick is drawing fairly. Words are not interchangeable — water and hypotenuse and obtund are three completely different questions — so the sample has to be spread across the range on purpose, not scooped from wherever is convenient.
Rarity is the axis that matters
Every word in the test carries a frequency rank: roughly, where it lands if you order the language from commonest to rarest. That single number predicts more about whether you know a word than length, spelling, or subject does. Knowledge falls off as words get rarer. The shape of that fall is consistent enough to build on.
This is why the test comes in three rounds that get harder. Each round samples a different stretch of that curve: everyday words first, then the middle, then the long tail. Your answers in each round pin down one point, and together they place the drop-off.
Why the last round is the long one
Most of a language is rare. The words you meet all day are a rounding error against the mass of words that turn up once a year, and the sampling has to respect that.
Each round's answers get scaled by how much of the language that round stands for. Know half the words in a round, and you are credited with roughly half of everything that round represents. Add the three together and that is the estimate. The arithmetic is simple. The judgment sits earlier, in how the words were picked.
The awkward part: people say yes too easily
Ask anyone whether they know a word and a certain number will say yes about words they have merely seen before, or that look like something they know, or that they feel they ought to know. Almost nobody is lying. Recognition is a vague feeling, and a checkbox cannot tell I could define this apart from that looks familiar.
Uncorrected, this wrecks the arithmetic: every over-eager tick gets multiplied by the size of the band it sits in.
So some of the words in the test do not exist. They were built to look and sound like English — pronounceable, plausibly spelled, sitting at a believable level of obscurity — and then checked against several dictionaries to make sure they are not real, not misspellings of anything real, and not a real word with a prefix bolted on. They are scattered through all three rounds with nothing to mark them out.
Nobody can know them, so every one you tick is a measurement of how freely you tick. That rate is then used to discount the rest of your answers. Tick none of them and your score is untouched.
The numbers behind this chart
| guessing rate | corrected | uncorrected |
|---|---|---|
| 0% | 27,000 | 27,000 |
| 5% | 26,470 | 27,720 |
| 10% | 24,920 | 28,430 |
| 15% | 22,550 | 29,150 |
| 20% | 19,610 | 29,870 |
| 30% | 13,140 | 31,300 |
The practical consequence: guessing cannot inflate your result. An occasional slip costs almost nothing, but every further made-up word costs more than the one before, so answers given very freely end up with a low score.
What the number is worth
The number comes from a hundred-odd questions about a language with six figures of vocabulary, so the last digits are noise. Run the same person through twice and the two results will differ, by more at the low end than the high end. What holds up is the rough magnitude and the comparison between people.
A few other things it is not. It counts word forms in a dictionary sense, not word families, so it will not line up with figures quoted from studies that count families and typically land lower. It only asks about nouns and verbs. It cannot tell whether you know a word well or only well enough to recognise it. That is a real distinction, and no checkbox will find it. And it measures recognition, not use. Most people recognise a good deal more than they would ever say out loud.
The words themselves come from a frequency-ranked list of English built from large-scale usage data, filtered down to single words of reasonable length. Which words appear, how many of each round are invented, and where the cut-offs sit are not published here, for the obvious reason.
Prior art
None of the underlying ideas are ours. The yes/no checklist format goes back to Meara and Buxton in the late eighties, along with the finding that it needs invented words to mean anything. Nation's Vocabulary Size Test established sampling by frequency band and scaling each band by what it represents. The particular correction used here — discounting your answers by how often you claim words that do not exist — is the one set out by Huibregtse, Admiraal and Meara in 2002. On top of it we add a penalty of our own that grows with every made-up word you tick, so that ticking almost everything cannot pass for a large vocabulary. Anyone curious about how many words people actually know should read Brysbaert and colleagues, who put the figure for an adult native speaker around forty thousand and explain at length why that number is so hard to pin down.
Take the test