A 1,000-word article with 98 percent vocabulary coverage still contains about 20 unknown running words. Some repeat. Others carry the argument. For a B1 reader on the Banqiao Line, that difference decides whether a 17-minute ride holds a continuous story or a sequence of dictionary stops. Teachers should measure the lexical fit of a text before calling it easy, then check the reader and the task instead of treating 98 percent as a universal pass mark.
One word in fifty sounds small. On the page, it is busy.
Coverage counts encounters
Vocabulary coverage is the percentage of running words a reader knows. If the word “permit” appears six times, its six appearances all enter the calculation. That makes coverage different from a count of distinct dictionary entries. Repetition can help after the first encounter, though the opening occurrence may still block a sentence.
The arithmetic gives a useful preview. At 95 percent coverage, a 600-word text contains roughly 30 unfamiliar word tokens. At 98 percent, the figure falls to 12. A learner who reads slowly enough to check each one will feel those two versions as different tasks, even when the topic, syntax, and length stay fixed.
Marcella Hu and Paul Nation tested this relation with 66 learners in a pre-university English course. Participants read one of four versions of a 633-word narrative. Pseudowords replaced ordinary words to create 80, 90, 95, or 100 percent coverage. Comprehension generally rose as unknown-word density fell. The authors inferred from their regression that about 98 percent would be needed for most learners to reach their chosen standard of adequate comprehension.
That famous figure came from an inference. The experiment had no 98 percent condition, used one narrative, and defined success through particular multiple-choice and cued-recall scores. It offers a planning estimate, not a law of reading.
The slope matters more than a cliff
Norbert Schmitt, Xiangying Jiang, and William Grabe later studied 661 participants from eight countries. Learners took vocabulary measures based on words in two texts, read those texts, and answered comprehension questions. More known vocabulary was associated with better comprehension, yet the results showed no sudden threshold where understanding jumped at one percentage.
A 2023 replication made the boundary even clearer. Kremmel and colleagues assigned 104 Sri Lankan learners to 80, 90, 95, 98, or 100 percent versions of a narrative and an informational text. Scores mostly improved as coverage rose. The study failed to confirm that 98 percent was sufficient for adequate comprehension, and results varied with genre and question format. A narrative multiple-choice test behaved differently from short answers about an informational passage.
This is the useful conclusion: every reduction in unknown-word density can make reading easier, while no percentage settles comprehension by itself. Background knowledge, syntax, genre, and the location of each unfamiliar word remain in the room.
“Stoat” may be harmless in a paragraph that describes woodland animals. “Unless” can reverse the condition in an exam instruction. Both count as one unknown token. Their costs differ.
Match the number to the purpose
Independent reading for pleasure needs a generous lexical fit because repeated interruption defeats the activity. A short lesson built around inference can carry more unfamiliar vocabulary, provided the teacher has chosen which words deserve attention and which can be ignored. A GEPT reading practice sits somewhere else: the learner needs enough difficulty to rehearse decisions under test conditions without turning the passage into a vocabulary quiz.
I would rather assign a slightly easier article and ask a demanding question than assign a dense article and spend 24 minutes translating its nouns. The first choice leaves attention for argument, reference, tone, and evidence.
Coverage also explains why a text labelled B1 can fail a particular B1 reader. Frequency lists describe populations. A software engineer may know “repository,” “deployment,” and “latency” while missing common words in a health article. A nurse may show the reverse pattern. Proper nouns, compounds, and familiar word families can also distort an automatic count.
Use a profiler as a first pass. Then mark the words the named learner actually knows. For a 700-word article, sample 70 words from across the text rather than only the opening. Ask the learner to mark each as known, partly known, or unknown in that sentence. Ten sampled unknowns suggest a poor fit. Two invite a closer look at their role.
The denominator also needs care. A profiler may split “ice cream” into two tokens, count a possessive ending separately, or group “decide,” “decision,” and “decisive” in ways that exceed the learner's knowledge. Run the same settings across candidate texts, and treat the result as an estimate. For one class, the teacher's marked copy is better evidence than a decimal printed by software.
When the estimate sits near the chosen boundary, read the first two paragraphs aloud with the learner and note every hesitation caused by an unfamiliar form, because that six-minute sample reveals false matches that a frequency list cannot see.
Adjust the text without emptying it
When coverage is low, deletion is only one option. Replace a rare word that contributes little. Add a short marginal gloss for a term the argument needs. Choose another text on the same subject so topic knowledge and recurring vocabulary can accumulate. Choc Education's article on reading several texts about one topic explains why that sequence gives words another chance to appear.
Keep some friction. A learner who knows every running word can still meet unfamiliar meanings, references, or sentence structures, and a reading course should allow that work. The counterexample is an assessment of unaided comprehension: glossing the decisive verb would change what the test measures.
A teacher can run the check tomorrow with one page. Count 50 running words, circle what the intended reader would miss, and read the whole sentence around each circle. If three circles land on details and one lands on the claim, the page needs a different repair for each of its four words.