95% known, 5% new: what the research says

Read95 gives each reader texts in which they already know about 95% of the words, so that the remaining 5% can be learned in context. The figure comes from reading research, where it is often quoted with more confidence than the studies behind it allow. We look here at what those studies found, how far their thresholds can be trusted, and why we still consider 95% a sensible place to aim.

Learners who reach an advanced level in a language have usually read a great deal in it. The difficulty, for most of them, is finding texts that are worth reading and that they can also follow: authentic writing is often too hard for an intermediate learner, and texts written for learners often contain little that is new. Krashen's input hypothesis held that a language is acquired by understanding input slightly beyond the learner's current level [1]. The hypothesis was influential, but it did not specify how far beyond, which made it difficult to test and of limited help when deciding what a particular person should read next.

Research on lexical coverage asks a measurable version of the same question. Lexical coverage is the percentage of the running words in a text that a given reader knows. A reader with 95% coverage of a page of about 300 words meets, on average, 15 words on it that they do not know [2]. Coverage depends on the reader and on the text together, which makes it a natural quantity to aim at when choosing a text for one person.

How many words do you need to know to understand a text?

Laufer [3] asked learners of English as a foreign language to underline the words they did not know in an academic text, checked those words with a translation test, and compared the resulting coverage with the learners' scores on a comprehension test. Reasonable comprehension was defined as a score of 55%, which was then the pass mark at her university [2]. Readers with at least 95% coverage were more likely to reach that score than readers below it, and 95% has been quoted as a lexical threshold ever since.

Hu and Nation [4] manipulated coverage directly. They replaced words in a short story with invented ones to produce four versions, at 80%, 90%, 95% and 100% coverage, and gave one version to each of four groups drawn from 66 learners in a pre-university English course in New Zealand, so that each group had 16 or 17 readers [5]. Adequate comprehension was defined on a multiple-choice test and a cued written recall. No reader reached it at 80% coverage, three of 16 did at 90%, and at 95% between 35% and 41% of readers did, depending on the test [6]. From a regression of comprehension on coverage, the authors estimated that about 98% coverage would be needed for most learners to read such a text adequately without help [4].

Laufer and Ravenhorst-Kalovski [7] then related the vocabulary size of 745 learners to the coverage it gave them of academic texts and to their scores on a standardised national reading test. They proposed two thresholds: a minimal one of 95% coverage, reached with a vocabulary of 4,000 to 5,000 word families, and an optimal one of 98%, reached with about 8,000. A word family is a word with its inflections and common derivations, and both figures count proper nouns as known [2]. At the minimal threshold, Laufer later wrote, learners can be expected to manage a text with some help from a dictionary [2]. Using corpora, Nation [8] estimated that reading a wide range of written English with 98% coverage requires knowledge of 8,000 to 9,000 word families.

Schmitt, Jiang and Grabe [6] took a different approach with 661 participants from eight countries. Instead of altering the texts, they measured each reader's knowledge of the words in two texts and related it to that reader's comprehension of the same texts. Within the range they could observe, 90% to 100% coverage, comprehension rose roughly linearly, from about 50% at 90% coverage to about 75% at 100%, with no point at which it suddenly improved. In their data, 60% comprehension corresponded to about 95% coverage, and 70% comprehension to 98% or 99%. Readers who knew every word still averaged only about 75%, which led the authors to describe high coverage as "an essential, but insufficient, condition for reading comprehension" [6].

Are 95% and 98% really thresholds?

These studies are often summarised as showing that a reader needs 95% or 98% coverage. Nevertheless, we would be cautious about reading either figure as a threshold in the clinical sense, a value below which comprehension fails and above which it succeeds, for five reasons.

First, the thresholds follow largely from how the outcome was dichotomised. Comprehension was measured as a continuous score, and each study chose its own cut-off for adequate comprehension: 55% in Laufer's study [3], and a stricter criterion, equivalent to about 71% on average across two tests, in Hu and Nation's [2]. Laufer herself attributed the difference between the 95% and 98% figures to these two definitions [2]. Where comprehension rises steadily with coverage, as it did in every study above, any cut-off placed on comprehension will produce a matching threshold on coverage. The threshold then describes the choice of cut-off as much as it describes the reader.

Second, the 98% figure was not observed. Hu and Nation's design had no 98% condition; the figure was inferred from a regression fitted to four groups of 16 or 17 readers, at a coverage level that none of them read [5]. With groups of that size, an estimate of this kind would likely carry considerable uncertainty.

Third, the 98% estimate has replicated poorly. Kremmel and colleagues [5] repeated Hu and Nation's design with 104 adult learners in Sri Lanka, added a 98% version and an informational text, and used both multiple-choice and short-answer questions. They confirmed a mostly linear relationship between coverage and comprehension for the original story and question format, but could not replicate 98% as sufficient for adequate comprehension, and their results differed by genre and by question format. An earlier approximate replication in Thailand, which they summarise, found that coverage explained less than a third of the variance in comprehension that it had explained in the original study [5]. In a small study using the original text and task with Arabic-speaking secondary school learners, Laufer [9] found no significant difference in comprehension between 90%, 95% and 98% coverage [5]. Most recently, Webb, Pellicer-Sánchez and Wang [10] gave a short story to 94 advanced learners at 90%, 95%, 98% or 100% coverage. Comprehension tended to increase with coverage, but the differences between 90%, 95% and 98% were not statistically significant, and coverage explained 11% of the variance in comprehension scores.

Fourth, coverage is itself measured with error. Laufer's readers reported their own unknown words [3]; other studies inferred each reader's coverage from a vocabulary test and the frequency profile of the text [6,7]. Readers are known to take unfamiliar words for familiar ones [2], and a vocabulary test estimates what a reader knows from a sample of words. If this misclassification is unrelated to comprehension, it would tend to weaken the observed association, and the true relationship could be somewhat steeper than reported.

Fifth, the populations are narrow. Almost all of these studies tested learners of English, most of them in universities, reading one or two short texts under test conditions. Kremmel and colleagues suggest that the learning context and the learners, non-academic adults in place of the university students of the original study, might account for their partial failure to replicate [5], and we suspect that a similar caution applies across languages. The word family, the counting unit in most of this work, was defined for English, and coverage in languages where a single word takes many more forms, such as Finnish, Russian or Arabic, has been studied far less.

Against these limitations, one finding held in every study we reviewed: comprehension rises with coverage. The relationship appears with manipulated texts [4,5,10] and with naturally varying vocabulary knowledge [3,6,7], in samples from several countries, and it shows a gradient. Consistency and a dose-response gradient are two of the considerations Bradford Hill proposed for judging whether an association is causal [11], and here they come with experimental evidence from the studies that manipulated coverage. Text difficulty also depends on topic familiarity, background knowledge, syntax and cohesion [5], so coverage explains only part of the variation in comprehension. What the evidence supports, then, is a direction: knowing more of the words helps, and the benefit continues past 95%. It does not support a sharp boundary at 95% or at 98%.

Why not read at 100%?

If comprehension were the only goal, the evidence above would argue for coverage as close to 100% as possible. A reader who wants to learn words from a text, however, has to meet words they do not know. Carver [12] found that school and graduate students reading material matched to their ability met about 1% unknown basic words, and almost none in easy material, and he argued on this basis that easy free reading offers few opportunities to learn vocabulary. The question for a learner is how many unknown words a text can carry before they stop being learnable.

Words are learned from reading, but slowly. In a meta-analysis of 20 experiments on incidental word learning during normal reading, Swanborn and de Glopper [13] estimated that students learn around 15% of the unknown words they meet. A more recent meta-analysis of second language studies found that learners learned 9% to 18% of the target words immediately after meaning-focused input, and 6% to 17% at delayed tests, with similar gains from reading and from listening [14].

Repetition matters. Across 26 studies with 1,918 participants, the number of times a word was met correlated moderately with learning it (r = 0.34) [15]. Waring and Takaki [16] gave 15 Japanese learners a graded reader in which 25 target words appeared between once and 18 times. On average, the meaning of only one target word was remembered three months later, and none of the words met fewer than eight times. Pigada and Schmitt [17], who followed one learner of French through a month of extensive reading and tested 133 words, found more: knowledge of 65% of the words improved in some respect, spelling most and meaning less.

The context around a word also affects learning. Learners who met target words in more informative contexts did better on tests of their meaning than learners who met them in poorer ones [18], and Liu and Nation [19] found that unknown words were easier to guess at a density of one in 25 running words than at one in 10. Laufer notes that the clues needed to infer a word often lie in other words the reader does not know [2]. Since the number of known words around each unknown one falls as coverage falls, these findings suggest, although they do not prove, that the chance of learning any one unknown word also falls as coverage falls.

Direct evidence on coverage and learning is scarce. In an eye-tracking study, Pellicer-Sánchez, Webb and Wang [20] gave advanced learners a text at 90%, 95%, 98% or 100% coverage. Reading was less effortful at 98% on one of three global measures, but coverage did not significantly change the attention readers gave to the unknown words, and longer processing of those words predicted larger vocabulary gains.

Taken together, the two bodies of research describe a trade-off. Lower coverage puts more new words on each page and leaves less understood context around each of them. Table 1 sets out the arithmetic for a page of about 300 words, the page length Laufer uses to illustrate coverage [2].

Counts are averages over a text. In practice, unknown words are spread unevenly.

At 98%, a reader meets about six unknown words on a page and understands most of the text. At 90%, they meet 30, and answer about half the comprehension questions. At 95%, most readers can follow a text with some help, and they meet two to three times as many new words per page as at 98%. We are not aware of a study that has measured where, between these points, the number of words learned per hour of reading is greatest, and the choice of 95% therefore rests on an inference from the two bodies of research.

What helps the new words stay

Two further findings bear on what a reader should be able to do at 95%. The first is that help with unknown words improves learning. Hulstijn, Hollander and Greidanus [21] gave advanced Dutch learners of French a short story with marginal translations of unknown words, with a bilingual dictionary, or with neither. Recurrence of a word helped retention more when its meaning had been supplied in a gloss or looked up than when readers had no help; without help, readers often ignored unknown words or inferred their meanings wrongly.

The second concerns when a word is met again. In a synthesis of 317 experiments, Cepeda and colleagues [22] found that the interval between study sessions and the time until the final test operate jointly, and that the best interval grows with the time over which the material has to be remembered. A meta-analysis of 48 second language experiments with 3,411 participants found a medium-to-large benefit of spaced over massed practice, with longer spacing better at delayed tests. Expanding intervals were no more effective than equal ones [23].

At the scale of whole programmes, meta-analyses of extensive reading, the practice of reading large amounts of text at a comfortable level, report a medium effect on reading proficiency compared with control groups (d = 0.46 across 34 studies) [24], and small to medium effects across 49 studies [25]. These studies evaluate reading programmes as a whole, not a particular coverage level.

How Read95 puts this into practice

Read95 gives each reader texts from real books, essays and articles in which they know about 95% of the words. At that level, the studies above suggest, most readers can follow most of a text and still meet a useful number of new words on each page. Because coverage is only as accurate as the estimate of what a reader knows, Read95 first estimates the reader's vocabulary from a short calibration, in which they select the words they know, and then refines the estimate with every text they read and every word they look up, in both directions.

Any word can be pressed and held (or clicked, on a Mac) to show what it means in that sentence. This is the help that Laufer's minimal threshold assumes [2], and the condition under which Hulstijn and colleagues found that meeting a word again improved its retention [21]. Words that are looked up come back for review at widening intervals, in line with the evidence on spacing [22,23], although that evidence favours spacing itself more clearly than any particular schedule.

Coverage explains only part of how hard a text is [5,10]. For this reason, a reader can mark a text as too hard or too easy and the following texts adjust, and texts are chosen according to the reader's interests, which may help with the topic familiarity that comprehension also depends on.

What we still do not know

The case for 95% combines two bodies of research that were not designed to answer our question: the comprehension studies did not measure learning, and, with one exception [20], the learning studies did not vary coverage. Most of the evidence also concerns learners of English, and the studies discussed here are the ones most often cited, selected by hand; a systematic search might have found others. Nor have we evaluated Read95 itself in a controlled study, so nothing here should be read as evidence that it works better than other ways of reading.

What the research does support is a consistent, roughly linear relationship between the share of words a reader knows and how well they understand a text, and a slow but real rate of learning new words from reading, which improves with repetition, with informative context and with help in finding the meaning. It does not support a fixed threshold at 95% or at 98%. We chose about 95% as a working compromise between understanding and learning, and we expect to revise it as direct evidence on coverage and vocabulary learning accumulates.

The author works on Read95, the app discussed in this article.

References

1. Krashen SD. Principles and Practice in Second Language Acquisition. Oxford: Pergamon; 1982.
2. Laufer B. Lexical thresholds for reading comprehension: What they are and how they can be used for teaching purposes. TESOL Quarterly. 2013;47(4):867–872. doi: 10.1002/tesq.140.
3. Laufer B. What percentage of text-lexis is essential for comprehension? In: Laurén C, Nordman M, editors. Special Language: From Humans Thinking to Thinking Machines. Clevedon: Multilingual Matters; 1989. p. 316–323.
4. Hu M, Nation P. Unknown vocabulary density and reading comprehension. Reading in a Foreign Language. 2000;13(1):403–430. doi: 10.64152/10125/66973.
5. Kremmel B, Indrarathne B, Kormos J, Suzuki S. Unknown vocabulary density and reading comprehension: Replicating Hu and Nation (2000). Language Learning. 2023;73(4):1127–1163. doi: 10.1111/lang.12622.
6. Schmitt N, Jiang X, Grabe W. The percentage of words known in a text and reading comprehension. The Modern Language Journal. 2011;95(1):26–43. doi: 10.1111/j.1540-4781.2011.01146.x.
7. Laufer B, Ravenhorst-Kalovski GC. Lexical threshold revisited: Lexical text coverage, learners' vocabulary size and reading comprehension. Reading in a Foreign Language. 2010;22(1):15–30. doi: 10.64152/10125/66648.
8. Nation ISP. How large a vocabulary is needed for reading and listening? The Canadian Modern Language Review. 2006;63(1):59–82. doi: 10.3138/cmlr.63.1.59.
9. Laufer B. Lexical coverages, inferencing unknown words and reading comprehension: How are they related? TESOL Quarterly. 2020;54(4):1076–1085. doi: 10.1002/tesq.3004.
10. Webb S, Pellicer-Sánchez A, Wang A. Lexical coverage and reading comprehension revisited. Reading in a Foreign Language. 2025;37(1):1–21. doi: 10.64152/10125/67517.
11. Hill AB. The environment and disease: Association or causation? Proceedings of the Royal Society of Medicine. 1965;58(5):295–300. doi: 10.1177/003591576505800503.
12. Carver RP. Percentage of unknown vocabulary words in text as a function of the relative difficulty of the text: Implications for instruction. Journal of Reading Behavior. 1994;26(4):413–437. doi: 10.1080/10862969409547861.
13. Swanborn MSL, de Glopper K. Incidental word learning while reading: A meta-analysis. Review of Educational Research. 1999;69(3):261–285. doi: 10.3102/00346543069003261.
14. Webb S, Uchihara T, Yanagisawa A. How effective is second language incidental vocabulary learning? A meta-analysis. Language Teaching. 2023;56(2):161–180. doi: 10.1017/S0261444822000507.
15. Uchihara T, Webb S, Yanagisawa A. The effects of repetition on incidental vocabulary learning: A meta-analysis of correlational studies. Language Learning. 2019;69(3):559–599. doi: 10.1111/lang.12343.
16. Waring R, Takaki M. At what rate do learners learn and retain new vocabulary from reading a graded reader? Reading in a Foreign Language. 2003;15(2):130–163. doi: 10.64152/10125/66776.
17. Pigada M, Schmitt N. Vocabulary acquisition from extensive reading: A case study. Reading in a Foreign Language. 2006;18(1):1–28. doi: 10.64152/10125/66611.
18. Webb S. The effects of context on incidental vocabulary learning. Reading in a Foreign Language. 2008;20(2):232–245. doi: 10.64152/10125/66826.
19. Liu N, Nation ISP. Factors affecting guessing vocabulary in context. RELC Journal. 1985;16(1):33–42. doi: 10.1177/003368828501600103.
20. Pellicer-Sánchez A, Webb S, Wang A. How does lexical coverage affect the processing of L2 texts? Applied Linguistics. 2024;45(6):953–972. doi: 10.1093/applin/amae062.
21. Hulstijn JH, Hollander M, Greidanus T. Incidental vocabulary learning by advanced foreign language students: The influence of marginal glosses, dictionary use, and reoccurrence of unknown words. The Modern Language Journal. 1996;80(3):327–339. doi: 10.1111/j.1540-4781.1996.tb01614.x.
22. Cepeda NJ, Pashler H, Vul E, Wixted JT, Rohrer D. Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin. 2006;132(3):354–380. doi: 10.1037/0033-2909.132.3.354.
23. Kim SK, Webb S. The effects of spaced practice on second language learning: A meta-analysis. Language Learning. 2022;72(1):269–319. doi: 10.1111/lang.12479.
24. Nakanishi T. A meta-analysis of extensive reading research. TESOL Quarterly. 2015;49(1):6–37. doi: 10.1002/tesq.157.
25. Jeon EY, Day RR. The effectiveness of extensive reading on reading proficiency: A meta-analysis. Reading in a Foreign Language. 2016;28(2):246–265. doi: 10.64152/10125/66901.