world · 4 min read
37% vs. 6%: same kind of launch, two languages, opposite outcomes
Traditional Chinese and Malay went live a week apart with the same goal. Two months later, one made up 37% of its region's orders. The other barely cracked 6%.
by Arthur
published September 26, 2026
Two and a half months after becoming the default language for Taiwan, Hong Kong and Macau, traditional Chinese already accounts for 37.2% of orders from those three places. Malay, which shipped a week later and only for Malaysia, barely reaches 6.2% over the same kind of window. Same type of launch, two countries starting from nearly the same spot, and one result six times bigger than the other.
Before and after each language
Before either language existed, American English already dominated both groups almost equally: 98.2% of orders from Taiwan, Hong Kong and Macau, and 98.8% of orders from Malaysia. The starting point could not have been more similar. What happened after each language went live is where the story splits.
| Locale shipped | Countries | Went live | en-us before (orders) | en-us after (orders) | Native language after (orders) | Native language after (visits) |
|---|---|---|---|---|---|---|
| zh-hant (traditional Chinese) | Taiwan, Hong Kong, Macau | Jul 11, 2026 | 98.2% | 59.4% | 37.2% | 34.4% |
| ms (Malay) | Malaysia | Jul 19, 2026 | 98.8% | 91.2% | 6.2% | 9.7% |
orders.destino->>'pais' grouped by locale, counted before and after each dictionary's go-live date (commits 97938e4 and eba0569); visits from sum(weight) of the locale prefix in the path (1:20 sampling, never a raw count). There is a 22-day gap with zero orders anywhere in the orders table (July 20 to August 12, 2026, logged as Supabase Fair Use throttling, not an actual traffic drop) that cuts into the start of both windows, and cuts deeper into Malay's, which lost almost its entire first week; there is no way to tell whether that pulls the ms percentage up or down, so the number is a snapshot, not a fine-grained measurement.
Reading the table: in the Chinese-speaking group, English dropped almost 39 percentage points, from 98.2% to 59.4%, and zh-hant reached close to 4 in every 10 orders. In Malaysia, English gave up only 7.6 points, from 98.8% to 91.2%, and Malay landed just above 1 in every 16 orders. Inside the Chinese group itself, the pattern holds country by country: 39.6% in Taiwan (132 of 333 orders) and 33.8% in Hong Kong (75 of 222). Macau had just 1 order in the whole period, too small a sample to mean anything on its own.
What this number does not prove
This measures observed adoption, not motive. We know which language someone actually browsed or bought in, not why they preferred one over the other. Malaysians being more comfortable with everyday English is a reasonable guess, but it is not something we measured here, it is an educated hunch wearing an explanation's clothes.
The two launches also did not get the same amount of runway. Zh-hant ran for about 9 days before hitting a 22-day gap with zero orders logged anywhere in the database, a Supabase Fair Use throttling effect, not a real traffic collapse. Malay had about a day and a half of runway before the same gap hit, losing almost its whole first week. The gap cuts into the start of both windows, and cuts deeper into Malay's: there is no way to tell whether that pushes the 6.2% figure up or down, so it counts as a snapshot, not a fine-grained measurement.
There is also something the table cannot see: people who never noticed the option to switch languages and people who saw it and chose to stay in English both land in the same en-us row. The middleware detects language by the visitor's country, but anyone can switch manually, and that choice is not logged separately. And there is the backdrop to keep in mind: total site-wide weekly orders fell from 9,737 in one June week to a range of 569 to 1,597 per week in September. Some of the drop showing up in both countries after launch is just the whole site cooling off, not a language-specific effect.
The same kind of launch produced two different audiences: one switched languages almost entirely, the other took a look and went back to English.
We assumed every new language behaves roughly the same way, with a similar adoption curve. It does not. That surprise is worth more than confirming the obvious hypothesis: it is the reason to keep measuring each language on its own instead of treating 'ship one more locale' as a checklist item that always plays out the same way.