We asked 200 chats what our app was for. They disagreed with our marketing.

10 November 2026 · ThaiPo

There were two theories about who uses ThaiPo, and both of them were ours. The marketing site held the first: a foreign man and his Thai girlfriend, a love story with a translator in the middle. The developer held the second: that story was out of date, real usage had drifted to business, to suppliers and condo committees, and the site was advertising to a customer who had quietly become the minority.

We were about to rewrite the whole site on the strength of theory two. Then it occurred to us that we run the product, and could simply go and count.

Both theories were wrong. Usefully, they were wrong in different directions.

A developer holding a clipboard in front of a wall of chat bubbles sorted into labelled piles, looking surprised at how the piles turned out

Where the hunch came from

The business theory was not invented. A few months earlier we had looked at a small sample of active chats and come away with "about 60% business operations". That number stuck.

It had two problems we noted at the time and then politely forgot. The sample was 18 chats. And it only looked at longer messages. Guess who writes long messages: people explaining a repair, a schedule or a delivery. Guess who writes short ones: people in love. We had built a filter that removed couples and then been struck by how few couples there were.

The census

So we did it properly. The method fits in a list, and the list is most of the value.

  • Start with metadata for everything. Every chat with any traffic in the last six weeks: 605 chats, about 107,000 messages. Message counts and language pair per chat, no content yet. This is the denominator, and you need it before you look at a single message.
  • Sample by volume, plus a random tail. The 120 busiest chats, then 80 drawn at random from all the rest. The top slice tells you what the product spends its day doing. The random tail tells you what a typical chat looks like. Those are different questions, and the whole story turned out to live in the gap between them. Together the 200 chats covered 96% of message volume.
  • Twenty messages per chat, read against a fixed rubric. The readers were language-model instances with one rubric and one list of categories, and each label came back with a confidence: 100 high, 83 medium, 17 low. Twenty messages is plenty. You can usually tell a factory from a marriage in about four.
  • Report categories and counts, nothing else.The output of the census is a table of labels. No message text, no names, no chat identifiers leave the job. The question was "what kinds of chat are there", and a tally answers it completely.
  • Run it nowhere near production traffic. Our web server also fronts the chat webhook, and the messaging platform gives a webhook about one second to answer. A CPU spike on that server once delayed translations by minutes, so a heavy archive read does not get to run there, ever. The census ran on a scratch machine booted from the same image, which existed for the job and nothing else.

What the chats said

By message volume, couples are 76%.Not a minority. Not even close. The fourteen busiest chats on the entire service are all couples, and the top ten alone carry 44% of every message we translate. The developer's hunch, measured by volume, was simply wrong.

By chat count, it is an even split. Just under half of the 200 chats were couples. Just under half were work. About one in twenty was family or friends. The marketing site, measured by count, was ignoring half its users.

Both are true at once because the two groups behave differently. A couple talks all day, every day, about everything. A work chat says "the technician comes Thursday" and goes quiet until Thursday. Work chats are numerous and calm; couples are fewer and relentless.

The work half was more varied than we had guessed:

  • property and condo committees
  • factories and warehouses
  • suppliers and buyers
  • hotels and guesthouses
  • other workplaces
  • customer service

Nearly half the volume is not English at all. This was the finding nobody had a theory about. Thai and English is 53% of volume. Chinese is 34%, then Japanese at 7%, Korean at 4% and Burmese at 2%. The site had been written for a Thai and English couple.

And the obvious explanation for the Chinese was wrong too. We assumed suppliers. Supplier chats are indeed about 90% Chinese, but they are roughly 2% of volume. Most of the Chinese traffic is couples. Theory one, the love story, was right about love and wrong about who was in it.

99% of volume is in groups. Even a couple is technically a group of three: two people and a translator. Almost nobody uses the product as a one-to-one chat with a bot, which is how most translation apps imagine themselves.

What we changed

The rewrite we had planned, "pivot the site to business", would have thrown away three quarters of our message volume to chase a hunch. The site we had would have kept ignoring half our chats and half our languages. The data suggested something duller and better: stop picking.

  • The headline says what the product does."Your LINE chats, translated automatically." It no longer says who you are.
  • Languages come first.The line above the headline names English, Chinese and Japanese before it says "40 languages", because a third of our volume needed to see itself in the first second.
  • A "who it's for" section built from the tally.One card per kind of chat we actually found: couples first, then property, factories, suppliers, hotels, family, friends and customer service. Its title is "One translator for every chat you have".

Couples still lead that section. They earned it, at 76%.

What we did not measure

One honest gap: we never joined the labels to billing, so we cannot say which kind of chat pays. The census answers "who uses it", not "who funds it". That is a separate query, and we would rather say so than let a usage chart pass itself off as a revenue chart.

The takeaway

Volume and count are two different stories about the same product, and most arguments about "who our users are" are one person quoting each. Get both numbers before the meeting. Sample the heavy head and a random tail, because either one alone will lie to you with a straight face. Keep the content out of the report. And when your marketing and your gut disagree, do not assume one of them must be right. Ours had managed to split the error between them.

ThaiPo is designed, built and run by one developer, Steven Acreman.

Try it in your own chats

ThaiPo lives inside LINE and translates every message in both directions, free forever. Get started free or add @thaipo.ai as a friend on LINE and it sends your signup link right in the chat.