What We Learned About Readability
Does the language a page is written in change how hard it is to use? We tested seven languages across three script families to find out.
The Short Answer
Yes — and the effect is measurable. A first-timer reading your Spanish docs page hits 11% more readability friction than the same content in Korean. An ADHD user faces 13% more cognitive load on German pages than Chinese ones. These aren't translation quality problems. They're built into the writing system itself.
Three Script Families, Three Cognitive Profiles
We found three distinct clusters. They group by writing system, not by language:
Latin-alphabetic (English, German, Spanish) — Words are split by spaces and vary in length. Long compound words (German Wahrscheinlichkeitsmaße) force the reader to hold more in working memory per eye fixation.
CJK logographic (Japanese, Chinese, Korean) — Each character packs more meaning into less visual space. Scanning is easier, but each token carries denser concepts.
Abjad (Arabic) — Words use short root stems with attached prefixes and suffixes. The right-to-left reading direction adds extra cognitive load, especially for attention-impaired users.
The radar chart shows why generic tokenization fails. CJK has a near-zero longWordRatio because words are only 1-4 characters. It also has high technicalDensity because each character is the concept. This creates a profile that looks nothing like alphabetic scripts.
The ADHD Effect
The ADHD persona's readability cost follows a clear pattern: Germanic and Romance languages are hardest. CJK is easiest. Arabic falls in between.
Arabic's position stands out. It scores higher than all CJK languages, even though its technicalDensity is lower. This lines up with research showing that RTL (right-to-left) reading direction adds cognitive load for attention-impaired readers. The effect is separate from script complexity.
Where the Effort Comes From
CBrowser's 6-layer Cognitive Transport Chain (CTC) breaks total effort into parts. The language effect shows up in just two layers:
Saliency (visual prominence), decision, motor, and frustration are nearly identical across languages. They measure DOM structure, not text content. All the difference comes from cognitive load and readability. That's exactly what you'd expect if the engine correctly separates language effects from layout effects.
Five New Metrics
These findings were possible because we added five script-aware linguistic metrics to the readability layer:
| Metric | What It Measures | Strongest Signal |
|---|---|---|
longWordRatio | Complex words per total | CJK vs everything else (30x difference) |
technicalDensity | Domain-specific terms | CJK 2.5x higher than alphabetic |
avgWordLength | Normalized per script | Distinguishes CJK from alphabetic |
lexicalDiversity | Vocabulary richness | Relatively stable across scripts |
avgSentenceLength | Clause complexity | Stable (content-driven, not script-driven) |
Each metric uses calibrated tokenization tuned to its script family. For CJK, the engine estimates 1.8 characters per word with a 6-character Latin equivalence. For Arabic, it normalizes 3-character root stems at 1.6x to reflect how much information each character carries.
What This Means for Your Site
If your site serves content in multiple languages:
German pages are measurably harder to read than English. Compound words demand more working memory.
CJK pages produce lower readability friction. Each logographic character carries more meaning per visual unit.
Arabic pages add RTL cognitive load. This hits ADHD users hardest.
The effect spans 11-13% across seven languages. That's real enough to shift abandonment predictions.
CBrowser's site_cognitive_assessment captures all of this in a single call:
site_cognitive_assessment url="https://your-site.com" userLanguage="de-DE" personas="first-timer,cognitive-adhd"Honest Caveats
We tested one prose page per language. All pages had the same content, translated via Google Translate.
Three languages per family for alphabetic and CJK. Only one abjad language (Arabic).
No abugida scripts tested (Hindi, Thai, Amharic).
Engine determinism confirmed: byte-identical scores across repeat runs.
Try It
Free tier: run cognitive_effort on any page with any persona. Pro: full site_cognitive_assessment pipeline with multi-language support, session reuse, and the five new linguistic metrics.
Upgrade to Pro — 2,000 credits/month, 110 MCP tools, cognitive journeys with value-modulated abandonment prediction across any language. Plus 500 bonus credits that never expire.
CBrowser v18.56 — Script-aware readability across Latin, CJK, and abjad. 110 MCP tools. Seven languages validated. The only framework that predicts when users give up — in any language.