The Thirty Percent Lost at the Border: on multilingual identity and websitesThe Thirty Percent Lost at the Border: on multilingual identity and websites

On September 23, 1999, NASA's Mars Climate Orbiter burned up in the Martian atmosphere instead of entering orbit around Mars. The investigation board found the cause within a few weeks: the software of the company that built the spacecraft produced thrust data in English units (pound-force seconds), while NASA's navigation team took the same numbers to be metric (newton seconds). There was no arithmetic error; two teams had read the same number in two languages. A $125 million spacecraft was lost to a translation error. The inch-to-centimeter table in an exporter's catalog, the date format on its website, or a single letter in a product name is an error of the same kind; the scale is smaller, the logic the same.

IBM's localization guidelines and the W3C's internationalization documents have given the same table for years: text translated from English into another language grows, and the growth is steepest in short strings. A button label under ten characters can double in translation; a paragraph grows by about thirty percent on average. German, with its compound words, and Finnish and Turkish, with their suffixes, are among the languages that grow the most: the English "Download" becomes "Herunterladen" in German; the Turkish "İndir" looks short, but "Kataloğumuzu indirebilirsiniz" ("You can download our catalog") swallows a whole line.

This is why the grid of a multilingual site is built around the longest language, not the shortest. Menu widths, button sizes, and heading areas are tested with German text; a site designed in Turkish and translated afterward falls apart in the second language. The typeface goes through the same test: a font that does not carry every accent of the Latin alphabet, and Cyrillic and Greek letters where needed, shows square boxes in some market. The Noto family that Google launched in 2014 takes its name from those boxes: "no tofu," no boxes.

The "Can't Read, Won't Buy" study that the language research firm CSA Research conducted in 2014 under Donald DePalma, with three thousand consumers in ten countries, put a number on the assumption that "everyone knows English": 75 percent of participants preferred to buy a product in their own language, and 60 percent rarely or never shopped on English-only sites. The study was repeated in 2020 with 8,709 people in twenty-nine countries and confirmed the trend: 76 percent wanted information in their own language, and 40 percent never bought from a site in another language.

These figures hold for the B2B buyer as well; the purchasing manager reads English, but the technical team, the quality department, and the manager above them do not always. The "EN" button on an export site is a beginning; for a manufacturer selling to Germany, a German technical data sheet brings in more quote requests than its English site does.

In 2003, Lera Boroditsky, Lauren Schmidt, and Webb Phillips ran a simple experiment: they asked German and Spanish speakers to describe objects with adjectives. "Key" is masculine in German and feminine in Spanish; the Germans described a key as "hard, heavy, metal," the Spaniards as "little, lovely, shiny." For "bridge" the genders were reversed, and so were the adjectives. Grammatical gender was changing how people imagined the object. That is why the way a product name or a brand "feels" in another language is learned not through translation but through testing.

Sound itself carries meaning, too. In experiments with invented brand names published in Marketing Letters in 2000, Richard Klink showed that front vowels (i, e) make a product feel smaller, lighter, and faster, while back vowels (o, u) make it feel bigger, heavier, and stronger. This sound symbolism is largely independent of language; it is Köhler's bouba-kiki finding of 1929 applied to brand names. The name of a multilingual brand is accepted only after it has been read aloud in every target market and put to native speakers.

In the software world, Turkish is remembered for a bug: the "Turkish-I problem." In most programming languages, when the computer's locale is set to Turkish, the function that converts text to uppercase turns the English word "file" into "FİLE," because in Turkish the capital of i keeps its dot (the language has a dotless ı and a dotted İ); the file name no longer matches, and the program crashes. The world's largest software companies spent years fixing this bug. In typography the same problem is solved with the OpenType locl feature, on the web with the lang="tr" attribute; when both are missing, "İSTANBUL" becomes "ISTANBUL" in a headline, and in the word "fiyat" (price) the fi ligature swallows the dot.

For the team building a multilingual site this is a warning: language is not a translation file but a setting written into every layer of the page. The font's language tables, the HTML language attribute, the server's character encoding, and the language signal given to the search engine must all say the same thing.

Google announced the hreflang annotation in December 2011: each page links the addresses of its own language and country versions to one another; the x-default value added in 2013 specifies which version to show a visitor who matches none of them. Without the annotation, the English page can be counted as a copy of the Turkish one and drop out of the index; set up wrong, it serves a German user a Turkish result. The rules are plain: each language lives at its own address (/en/, /de/), the links are reciprocal, and the language is not forced by automatic redirection; a British buyer visiting from Istanbul is not sent to Turkish.

Measurement is done by language as well. Search Console's country and page breakdown shows which market arrives in which language; when visits from Germany end on the English page without a form being filled in, that says what is missing is not the design but German. A multilingual site starts working not on the day it goes live but on the day it is updated with this data.

The Mars Climate Orbiter's error is repeated every day in catalogs. "12.500" is twelve thousand five hundred in Turkish and twelve and a half in English; "03/04/2026" is March 4 in America and April 3 in Europe. The international standard ISO 8601 (1988) ties the date to a single form: 2026-04-03. For units of measure the rule is a paired table: millimeters and inches, kilograms and pounds side by side; the buyer does not do the math, a buyer who does the math tires, and a wrong result goes into the order. Even the paper proportions of a catalog bound for America are different: Letter is not a sibling of A4.

Arabic and Hebrew are written from right to left, and the site of a brand entering the Gulf market mirrors not only the text but the layout: the menu moves to the right, the back arrow faces left, the progress bar fills from the right; numbers, however, stay left to right. Unicode's bidirectional text algorithm (UAX #9) resolves this mix in the browser, but unless the design itself is tested in its mirrored state, icons face the wrong way. Color changes meaning at the border, too: white is the color of mourning in some East Asian cultures; red carries luck in China and an association with mourning in South Africa. These differences do not require changing an identity; they require knowing what will be read, where, and how.

Our protocol consists of eight checks: the grid is tested with German text; the font is chosen for its accent coverage, locl support, and Cyrillic coverage where needed; every page carries lang and reciprocal hreflang, and both are verified; there is no automatic language redirection; numbers, dates, and units are written in two forms for every market; brand and product names are read aloud in the target language and put to a native speaker; if there is a right-to-left market, the layout is reviewed in its mirrored state; after launch, measurement is read by country. The spacecraft bound for Mars tripped on a single line of this checklist; the buyer of an export site trips the same way.

NASA Mars Climate Orbiter Mishap Investigation Board (1999). Phase I Report, November 10, 1999.

IBM Globalization Guidelines, "Text expansion"; Ishida, R. (2007). Text size in translation. W3C Internationalization.

Google (2014). Noto fonts, "No more tofu".

DePalma, D. A. et al. (2014). Can't Read, Won't Buy. Common Sense Advisory / CSA Research; (2020) Can't Read, Won't Buy — B2C, 29 countries, 8,709 participants.

Boroditsky, L., Schmidt, L. A. & Phillips, W. (2003). Sex, syntax, and semantics. In Gentner & Goldin-Meadow (Eds.), Language in Mind, MIT Press, 61–79.

Klink, R. R. (2000). Creating brand names with meaning: The use of sound symbolism. Marketing Letters, 11(1), 5–20.

Köhler, W. (1929). Gestalt Psychology.

Unicode Consortium, CLDR and Turkish uppercase/lowercase rules; OpenType locl feature.

Google Search Central (2011). New markup for multilingual content (rel="alternate" hreflang); (2013) x-default.

ISO 8601 (1988). Representation of dates and times.

Unicode Standard Annex #9, Bidirectional Algorithm.