Journal · 28 Mar 2026

Activation curves in mixed Thai–English teams

If your screenshot deck is English and your largest cohort reads Thai, a single activation curve is a diplomatic fiction.

Thailand product orgs often ship bilingual surfaces. Design critiques happen in English. Support macros happen in Thai. Event names stay in English because the warehouse was born that way. None of this is a moral failure. It becomes an analytics failure when first-value completion is reported as one line.

What we saw on a wallet checklist

A Pulse Seat student brought an onboarding checklist. English screenshots looked clean. Production first-value — a verified wallet ready to transact — lagged. Splitting the curve by UI language showed Thai users stalling on a step whose label still said “Verify wallet” in English inside an otherwise Thai screen. QA had never seen that mixed string because QA accounts defaulted to English.

After the label was rewritten and the tap was re-instrumented, the Thai curve moved toward the English one over about six weeks. Revenue did not magically jump. What jumped was honesty: the team stopped telling the GM that “activation is 41%” when that number was an average of two different products wearing one skin.

Language is a cohort property

In Feature Adoption Analytics as we teach it, language belongs on the cohort contract. Who enters: users whose UI language at first exposure was Thai, English, or other. If you only store device locale and not UI language, you will mis-split people who switched languages after install. Write that limitation on the uptake map. Do not hide it in a footnote nobody reads.

Mixed-language teams also disagree about first value itself. An English-speaking PM may call “opened settings” a success. A Thai-speaking support lead may know that users open settings to find the phone number, not to finish the job. Sit them in the same critique. The sentence that survives is the one you instrument.

Practical splits without a localisation department

You do not need a giant localisation org to start. Add UI language as an event property. Look at first-value completion for the two largest languages every week. If they diverge by more than you can explain with sample noise, read the actual strings. Often the bug is one untranslated verb.

Cloudeventgrid clinics in Lat Krabang keep printed screenshots in both languages on the table. It is harder to wave away a bad Thai label when it is sitting next to the English one under a lamp. Remote seats do the same with a shared folder. The method is unsophisticated. It works.

Continue with the marketplace case study or programmes.