How much of this text is already inside your first 5,000 words?

Coverage is the share of running words in a text that you already know. It's the number behind every "learn 1,000 words and you can read a newspaper" claim, and it's worth measuring on a text you actually want to read rather than on somebody's corpus.

So paste one in. Everything runs in your browser and nothing is sent anywhere.

It counts forms, and that matters

The checker on this page counts written forms. Hablo and hablas are 2 separate items in the list and 2 separate lookups here.

The famous coverage targets are not counted by written form. Paul Nation's 2006 paper, the source of "8,000 words to read a novel," counts word families: a headword plus every inflection and derivation you could decode from it. His own lists average 4.33 forms per family, so 8,000 families works out at 34,660 distinct written forms.

The unit difference means a percentage from this page and a percentage from Nation's paper are different measurements wearing the same sign. If you hit 90% here at 3,000 forms, you have not hit Nation's 3,000-family mark. Romance verbs make the gap worse: an English verb has 4 inflected forms, and hablar has around 44 before a derivation like hablador joins it.

Lemmatising would fix the unit, and I didn't do it. A tagger that guesses wrong on porto, which is a Portuguese city and a form of portare, corrupts 2 ranks at once, and I'd rather show you a number whose flaw is printed than a number whose flaw is hidden inside a model. So: forms.

Do proper nouns count toward coverage?

Yes, as coverage you get for free, because Madrid is Madrid in every language you'll read it in. They are also additive to the published figures rather than part of them. Nation's conclusion says the first 1,000 words plus proper nouns cover 78% to 81% of written text, and the clause falls off almost every time somebody quotes the number.

The proper-noun block is not small and it is not stable. Across the 14 texts Nation profiled it runs from 0.50% of The Turn of the Screw, a ghost story with 4 characters in it, to 6.12% of the Brown newspaper corpus. His 5 novels average 1.45% and his 5 newspaper corpora 5.41%, and the 2 groups don't overlap at all.

So a checker that quietly folded names into one percentage would move your score by up to 5 points depending on whether you pasted a novel or a news story. This one counts them, prints them, and leaves them out of the band figures.

The list, and its licence

The ranks come from hermitdave/FrequencyWords, built from the OpenSubtitles 2018 corpus through OPUS. The top 5,000 forms per language, lowercased, with duplicates removed after lowercasing. Downloaded 4 September 2026.

The FrequencyWords licence is CC BY-SA 4.0 for the data. The repository's README splits that from the MIT licence covering its code, and the LICENSE file holds only the MIT text, which is how "hermitdave is MIT" got started. Share-alike is a real obligation and this page honours it by attribution.

FrequencyWords is the list I could ship. Of the free options covering all 4 of these languages, it's the only one whose licence permits commercial use. SUBTLEX-ESP and SUBTLEX-IT are CC BY-NC-SA, SUBTLEX-PT-BR adds NoDerivatives on top, SUBTLEX-PT carries no licence file at all, and KELLY Italian is CC BY-NC-SA 2.0. I went through every list and read each licence at its own source, and several of the most-recommended ones would have been the wrong answer here.

Being subtitles, the OpenSubtitles corpus is dialogue. These ranks describe how people speak on screen, so a legal contract or a chemistry paper will score badly and that is the corpus talking, not you.

What this will get wrong

The proper-noun test is a capital letter that isn't at the start of a sentence. It's a heuristic, and it has 5 failures worth knowing.

• A name in the first position of a sentence gets counted as vocabulary.
• A text pasted in all lower case gets no names at all.
• A text pasted in ALL CAPS gets almost every word binned as a name.
• A capitalised common noun mid-sentence gets wrongly binned.
• An abbreviation ends a sentence as far as the checker is concerned, so in Sr. Garcia fue a casa the period after Sr makes Garcia look sentence-initial and it slips through as vocabulary.

The count of names is printed so you can eyeball whether it looks right for what you pasted.

Anything past rank 5,000 shows as outside. The list itself runs to 50,000, and I cut it at 5,000 because that's the range a learner works in and 4 full lists would be a 12 MB page.

Numbers, and tokens with no letters in them, are skipped entirely rather than counted as unknown. Apostrophes and hyphens follow the list rather than a general rule, because the list treats them differently. An elided clitic keeps its mark and stands alone, so l'homme is l' plus homme, and l' is rank 16 in French. A hyphenated compound stays whole, so est-ce is one word at rank 70 and peut-être one at 130. Portuguese carries 206 hyphenated forms like diz-me on the same rule. Splitting both marks, which is what this did first, put 202 French and 215 Portuguese entries permanently out of reach and sent aujourd'hui to the unknown pile from rank 271.

And the whole checker measures recognition of shapes on a page, which is the easiest thing vocabulary does. Reading desarrollo and knowing it means development is not the same skill as producing it in a sentence while somebody waits. The frequency curve explains why the first thousand ranks buy so much and the tenth buys so little.

Questions people actually ask