Readability formulas and what a score can tell
Readability formulas estimate features of a text associated with reading difficulty. Their scores can guide inspection or satisfy a specified target, but do not establish whether intended readers understand and can use the information.
Definitions and measured features
Readability concerns how easy content is to read and understand. It includes linguistic and visual aspects and has no single practical definition. Three definitions illustrate its range:
| Definition | What it includes |
|---|---|
| George Klare | Ease of comprehension attributable to writing style. |
| Harry McLaughlin | How compelling and comprehensible a particular group finds the material, recognising shared reading skills, prior knowledge or motivation. |
| Edgar Dale and Jeanne Chall | The interacting features of printed material that affect a group’s understanding, optimal reading speed and interest. |
Educators began developing reading-ease assessments in the 1920s, and formulas became popular in journalism, healthcare and industry in the 1950s. Flesch and Gunning helped reduce newspaper front-page reading levels from grade 16 to grades 9–12, as reported in 2012.
Common formulas use sentence length and word features, such as syllable counts or vocabulary familiarity, as indicators of difficulty. Flesch Reading Ease is conventionally presented on a 0–100 scale, with higher scores indicating easier text, based on average sentence length and syllables per word. Flesch–Kincaid expresses related features as an approximate US school grade. Dale–Chall also considers complex-word frequency against a built-in vocabulary.
Flesch measures are widely used for biomedical articles despite not having been designed to assess them. Traditional metrics may omit semantic relevance, coherence and neurocognitive aspects of reading.
What the scores leave out
Classical formulas do not account for lighting, noise, eyesight, fatigue, background knowledge, information needs, time, motivation or interest. Concept density, familiar words used in unfamiliar contexts and weak connections between ideas can also make low-grade text difficult to understand. These physiological and psychological factors are difficult to quantify, which helps explain the emphasis on linguistic predictors.
Classical formulas apply to running text rather than lists, tables or figures. Broader assessments such as the Suitability Assessment of Materials, PMOSE/IKIRSCH and the User-Friendliness Tool consider features including graphics, layout, typography or cultural appropriateness. A 1935 investigation identified 289 contributing elements and grouped them into content, style, format and organisation.
Shortening sentences alone does not necessarily improve comprehension: longer sentences have more opportunities for grammatical complexity, so length can act as its surrogate rather than its cause. A paragraph and the same words in reverse order can receive the same readability score despite the latter being meaningless. Scores can therefore improve without a corresponding improvement in understanding.
Different formulas can give different results for the same document, and results within a formula are not precise; SMOG has been described as accurate only to about 1½ grade levels. Scores provide an indication of complexity or wordiness rather than an assessment of effectiveness or intention. Readability formulas should not be treated as prescriptions for good writing.
Grade levels and intended readers
A grade-level estimate is not a guarantee that someone with that many years of education will understand the text. Formula development begins by testing passages with readers of known ability, then identifying text features that correlate with their comprehension scores.
A grade-to-age orientation used in a 2018 assent-form account was:
| US grade | Approximate age | Score interpretation |
|---|---|---|
| 5 | 10–11 | Very easy |
| 6 | 11–12 | Easy |
| 7 | 12–13 | Fairly easy |
| 8–9 | 13–15 | Standard |
| 10–12 | 15–18 | Fairly difficult |
| College | 18–22 | Difficult |
| College graduate | Over 22 | Very difficult |
Such labels are mathematical estimates of understanding and do not account for reader motivation, writing style or design and layout. The audience’s age group or reading level should be identified when possible before interpreting a score.
Practical use and conflicting recommendations
Formulas can provide a quick approximation of reading level. By 2010, some medical centres, health departments, agencies and pharmaceutical companies specified maximum grades for health documents, ranging from fourth to tenth and most often sixth grade.
One published critique argues that formulas should not be used to improve comprehension because they omit important influences, can score meaningless text and depend on variable reader samples, passages and test construction. The responding health-literacy practitioners largely agreed about those limitations but noted that formula assessments were commonly accepted evidence of compliance with contractual grade-level requirements. Their practical objection was that refusing to calculate a score leaves the writer needing another form of evidence acceptable to the contract administrator.
Where formatting or punctuation can skew a calculation, the assessment can be prepared on a separate copy; the formatted original should be saved before removing headings, bullets or punctuation. Tools that identify multisyllabic words can help writers decide what to retain, replace or remove, rather than supplying only a composite score.
A worked example illustrates the need for judgement: “See a healthcare provider for treatment without ado” scored at fifth grade despite its unfamiliar wording. Replacing “ado” with “delay” made it clearer but raised the calculated level because the word was longer, while “right away” was clearer still and scored at second grade.
A contrasting recommendation treats formulas and online tools as learning aids for medical writers and new researchers, particularly away from deadline pressure. Experiments can explore four changes:
- Shorter sentences.
- Shorter, nontechnical alternatives to longer words.
- Removal of unnecessary words, such as adverbs.
- A balance between active and passive voice.
In 2018, Hemingway Editor and Readable.IO offered visual analyses that highlighted potentially difficult sentences and words. A 2015 review described ReadabilityFormulas.com calculating seven tests at once and Word reporting sentence and word statistics, passive-sentence percentages and two readability scores. In a 2025 practitioner recommendation, NHS readability tools and Hemingway were suggested as aids to simplification.
A comparison with reader performance
A 2023 medical-device validation compared Readable, Grammarly and Word with testing in three groups of lay participants, numbering 24, 24 and 12. Participants had no prior knowledge of medical terminology, and success required at least 90% of questions to be answered correctly by 80% of participants. The automated tests considered surface features such as sentence length and word simplicity.
All three tools classified the documents or sections as having unacceptable readability, while participants met the success criterion. Participants considered some documents too long but found their detail, information and organisation appropriate. The investigators concluded that participant testing gave more relevant confirmation of adequacy and that the benefit of combining automated and participant testing needed further investigation.
Grade level and cohesion
A 2013 analysis examined 55 online colorectal-cancer and screening texts selected through searches in May–June 2012. Their mean length was 914 words (SD 581; range 113–2,611), and mean Flesch–Kincaid grade was 11 (SD 1; range 7–12).
The analysis used six Coh-Metrix indices: argument overlap and word-stem overlap for adjacent sentences and for sentence pairs within paragraphs, and latent semantic analysis for adjacent sentences and all sentences in the text. Argument overlap concerns shared or referring nouns, pronouns and noun phrases; stem overlap concerns shared word stems. Latent semantic analysis estimates contextual similarity of meaning from large text corpora.
Spearman correlations were tested with a two-sided significance threshold of 0.05.
The results for 55 texts were:
| Cohesion index | Range | Mean (SD) | Grade ρ | Grade P | Length ρ | Length P |
|---|---|---|---|---|---|---|
| Argument overlap, adjacent sentences | 0.39–1.00 | 0.69 (0.12) | 0.25 | 0.07 | −0.13 | 0.33 |
| Argument overlap, all sentence pairs | 0.24–1.00 | 0.55 (0.14) | 0.26 | 0.06 | −0.04 | 0.79 |
| Stem overlap, adjacent sentences | 0.40–1.00 | 0.68 (0.13) | 0.36 | 0.007 | 0.02 | 0.88 |
| Stem overlap, all sentence pairs | 0.28–0.90 | 0.54 (0.14) | 0.31 | 0.020 | 0.03 | 0.85 |
| Latent semantic analysis, adjacent sentences | 0.19–0.54 | 0.34 (0.07) | 0.22 | 0.11 | 0.001 | 0.99 |
| Latent semantic analysis, all sentences | 0.14–0.62 | 0.31 (0.09) | 0.35 | 0.009 | −0.14 | 0.33 |
Three indices showed moderate, statistically significant positive associations with grade level; the other three did not. Text length was not significantly associated with any cohesion index or with grade level (ρ = −0.17, P = 0.21). Lower-grade texts therefore tended to be lower on some cohesion measures.
Revisions aimed at lowering a grade score should preserve connections between ideas rather than merely truncate words and sentences. Coh-Metrix could help check cohesion after revision. In 2013 it generated more than 100 indices, while the simpler Text Easability Assessor offered five, including referential cohesion. The Text Easability Assessor output did not identify where revision was needed. Writers should not revise merely to obtain good scores on these measures.
A published critic questioned the significance of correlating a formula counting word features with a tool counting repeated words, regarding their components as qualitatively different. The investigators replied that formulas remained in practical use and that easy scores might overlook cohesion that could aid comprehension. The study’s own interpretation allowed that its correlations might reflect its selection of a small text sample and six indices, without examining other forms such as conjunction use.
Broader measures in development
By 2025, newer approaches incorporated cognitive mechanisms such as working memory, comprehension and predictive processes, including AI-based approaches, but practical uptake in medical and scientific writing remained limited. Two implementation gaps were identified:
- Effective tools integrating the newer metrics into readability assessment.
- Writer training and methodological frameworks.
User-friendly tools incorporating cognitive factors remained a development opportunity.