An AI design agent trained on Western portfolios has a predictable failure mode. Crawl a Japanese news portal, a Chinese e-commerce homepage, or a Saudi bank’s site, and the scoring model fires off every alarm at once: too much text, no whitespace, oversized nav, wrong-side alignment. The agent is not wrong about the pixels. It is wrong about the market. This is the first question any design agent must answer — what can I perceive? — and the answer is a stack of DOM signals that only mean something when interpreted per market.
What an agent can actually perceive
Everything an agent needs is already in the rendered page. Text density is computable: count visible text-node characters and divide by viewport area. Direction is computable: read the dir attribute, text-align, and flex/grid axis direction. Font behavior is computable: CJK faces like Microsoft YaHei and PingFang SC have different fallback and weight behavior than Latin faces, and Arabic rendering varies across operating systems. Navigation density is computable: count links in the primary nav. Color is computable: contrast ratios, saturation, and palette statistics never lie.
The trap is not measurement. It is interpretation. A Western-trained agent treats density as an error and left-to-right as a law, so it flags deliberate market conventions as bugs.
East Asia: density is the design
Start with Japan. Japanese sites routinely pack dense content above the fold with smaller fonts — an information-first approach that Western minimalism reads as clutter. Yet the same market produces some of the most restrained award work anywhere: Japanese studios win Awwwards, FWA, and CSS Design Awards at rates that belie the stereotype, working across what one studio guide calls the “density-vs-ma spectrum” (Utsubo, 2026). Alongside that minimalism sits a parallel tradition of vertical rhythm (tategaki columns), offbeat color pairings, and character- and glyph-forward composition (Designmodo). Both poles are native. An agent that only rewards the minimal pole misjudges half the market.
China inverts the Western baseline almost point-for-point. Chinese platforms show high information and feature density with highly saturated colors; Thoughtworks’ analysis attributes this to three forces — the dial-up era when everything had to load in one pass, a logographic script whose square characters are visually dense and whose readers are practiced at processing density, and brutal competition that pushes one-stop “super-app” pages (Thoughtworks). The conclusion there is blunt: “It’s not that they cannot design simple, minimalist interfaces… busy interfaces make more commercial sense.”
The differences are measurable. Where Western sites maximize whitespace and short punchy copy, Chinese sites are busier, longer, and packed with links, banners, and deals — Taobao versus Amazon, or Qunar’s crammed ticketing page versus Skyscanner’s clean one (AZ-Loc). Navigation alone is a computable gap: YouTube offers three top-level categories while Youku offers 21, with deep sub-menus. And the conventions are deliberate: Chinese users accept pop-up advertising, expect QR codes as primary CTAs, and read better in sans-serif faces like Microsoft YaHei at slightly larger sizes than the thin serifs Western brands favor (Nanjing Marketing Group). Typography that “looks elegant” in Latin simply does not render at small sizes in Hanzi — a perception failure an agent can catch by measuring glyph density per font.
The Middle East: direction is the layout
Arabic is the cleanest computable international signal there is: a single dir="rtl" attribute on the document, mirrored grid axes, and a nav that starts top-right. Yet most Arabic sites are still “English templates with the text direction flipped,” and the typography consequences are severe (Voxire). Arabic needs 1.7–1.85 line height against 1.5–1.6 for Latin, because descenders and tashkeel diacritics need vertical room; letter-spacing must be zero because tracking breaks connected-script ligatures; system default fonts render Arabic poorly, so real webfonts like IBM Plex Sans Arabic, Cairo, or Tajawal are required. Mirrored layout is not cosmetic: grid direction, icon direction, sliders, pagination, and breadcrumbs all flip, and navigation begins on the right (ExtraDigital). Arabic also expands — button labels and headings run longer than their English source, so fixed-width layouts break. An agent that flags “nav on the right” or “crowded body text” as a defect is applying LTR assumptions to an RTL market.
Multi-language and Korea: expansion, color, and play
India and other multi-script markets add a third axis: text expansion. Translations can grow 300% for short strings (W3C’s figure), while Japanese and Chinese compress several words into one or two characters — so layout must survive both extremes, and scripts like Arabic and Devanagari need taller line boxes (Weglot). Color and imagery carry market-specific meaning too — yellow reads as divine in China but duplicitous in the US, and even a brand like Victoria’s Secret reskins its homepage per region (LangShop). Korea, meanwhile, trends toward dense-but-clean, pastel, animation-heavy pop-futurism: 3D cartoon typography, metallic liquid effects, and K-pop optimism (Envato Elements). None of this is “bad whitespace.” It is a different target.
Computable rules an agent can implement today
The fix is market-aware thresholds, not a single global heuristic. Two rules to start:
- Info-density bands. Let
d = visible text characters ÷ viewport area in CSS pixels². Calibrate per market on a labeled corpus; plausible starting bands are Western ≤ 0.012, Japanese 0.02–0.05, Chinese 0.03–0.08. The agent detects the market (fromlang, script coverage, anddir) and penalizes density only above that market’s ceiling — never for merely exceeding the Western band. - RTL mirror check. When
dir="rtl", verify: primary nav’s first item sits atinline-start; layout uses logical properties (margin-inline-start,padding-inline-end) rather than hard-codedleft/right; Arabic bodyline-height ≥ 1.7; andletter-spacing: 0on all Arabic text blocks. Any violation is a real defect; a mirrored-but-consistent layout is not.
The answer
So what can an agent perceive? Everything it needs — text density, direction, font metrics, nav structure, color — all of it is computable from the DOM and computed styles. What it cannot perceive, until we encode it, is that those numbers mean different things in Tokyo, Shanghai, Riyadh, and Berlin. An agent that treats density and direction as market conventions rather than errors stops “fixing” Taobao into an Amazon clone and starts judging each page against the criteria its audience actually expects. Perception without market context is not perception — it is prejudice with a gradient descent. Design better, internationally: read the market before you read the layout.