Every Language We Launched Grew 70–120% a Week. The Demand Was Visible Before We Built Anything.

In 2023, CoinDesk was an English-first publication covering one of the most borderless markets there is. Someone trading Bitcoin in Seoul needs the same breaking news as someone trading it in San Francisco, and we were only really serving one of them.

So I built a machine translation system to change that, and every language we launched grew traffic 70–120% week over week. The number gets people's attention, but how we chose where to launch is the more useful part of the story, because the demand was visible before we built anything. We didn't have to ask anyone whether they wanted it.

How do we validate a new market before we build for it?

Look for behavior the audience is already exhibiting – what they buy, hold, use, read and choose – rather than sizing the market or asking people what they'd want. Then run the cheapest launch that can confirm or refute it, and make sure you can reverse it if the answer comes back no.

We didn't rank languages by market size

Most companies decide which languages to translate into by looking at GDP, internet penetration or general market size. Those are reasonable-sounding numbers, and they answer the wrong question. A country can have an enormous internet population and a small crypto community, and a smaller country can have a disproportionately active one.

So I took a different approach built on two data sources. The first was the countries with the top cryptocurrency holdings, meaning the populations holding the most crypto. The second was the countries with the most cryptocurrency transactions, because a place can have a lot of trading activity without a lot of holding, and the reverse. I brought those together with Google and SimilarWeb data on how people engaged with CoinDesk and with other news outlets, and that produced a prioritized list that looked very different from a "biggest internet markets" ranking.

I won't share the order. It was proprietary competitive intelligence for CoinDesk, and it would only matter to you if you were also translating crypto news. What matters is that every language we launched had already been validated by real activity before a single article went out in it.

The demand was already in what people were doing

None of those data sources involved asking anyone a question. Holdings, transactions and reading habits are all things people had already done, with their own money and their own attention.

There was one more signal, and it's the one most teams would read the wrong way. We were late. Other major crypto news outlets were already translating, and that was part of what told us we needed to start. A competitor already serving an audience can look like a reason to stay out, but it's also evidence that the audience exists and that someone has found it worth the effort to reach. When the traffic came in, it told us something else: those readers preferred CoinDesk and had been waiting for us to show up in their language.

The broader data says the same thing about language. In the European Commission's Flash Eurobarometer survey of 13,752 internet users, 9 in 10 said that when they're given a choice of languages, they always visit a website in their own – even though 55% use at least one other language to read or watch content online. CSA Research's 2020 survey of 8,709 consumers in 29 countries found that 76% prefer to buy products with information in their own language and 40% will never buy from websites in other languages. Those are buying numbers rather than reading numbers, but the pattern is the same one we saw: people who can use English still choose their own language when you offer it, and a company publishing only in English never sees that demand at all.

Why behavior beats asking

The reason to look at what people already do, instead of asking them, is that what people say they'll do is a discounted signal.

Webb and Sheeran's 2006 meta-analysis of 47 experiments found that a medium-to-large change in people's intentions produced only a small-to-medium change in their behavior. The study comes from psychology rather than product research, so I wouldn't stretch it into a claim about customer interviews specifically, but the direction is clear. Intent is real and it leaks on the way to action. Behavior that has already happened doesn't leak, because it's already happened.

This is why I'm wary of validation plans that start with a survey. A survey asks people to predict their own future behavior, and if the behavior you need is already happening somewhere, you can measure it directly instead. If it isn't happening anywhere, that's worth knowing too – it's the third question most expansion plans never ask.

Make the launch the test

The second half of the method is that the launch itself should be the experiment, and it should be cheap enough and reversible enough that being wrong doesn't hurt much.

Our launches were cheap. API translation fees were pennies per article, and because translation engines kept leapfrogging each other, we built middleware that let us switch the engine for any language in about five minutes. If a language had flopped, the cost of finding out would have been small and the cost of backing out smaller still.

The expensive part was the glossaries, which were the most labor-intensive part of the whole project and also the most important. General-purpose translation engines don't know crypto. Left alone, they'd translate "crypto" into a word that meant "to encrypt something," and Ethereum's "gas" fees would come out as the chemical substance. We hired a translation agency to build crypto glossaries for every target language and kept sending them new terms as the vocabulary changed, and those glossaries grew to hundreds of entries per language. That was real work, but it was work on quality, after the demand question had already been answered.

Cheap, reversible tests beat confident forecasts because even experts are bad at predicting which ideas will work. Ron Kohavi and Stefan Thomke reported in Harvard Business Review that at Google and Bing only about 10% to 20% of experiments produced positive results, and at Microsoft roughly a third were effective, a third neutral and a third negative. Jeff Bezos made the organizational version of the point in Amazon's 2015 shareholder letter: most decisions are "two-way doors," and treating them with heavyweight process leads to "failure to experiment sufficiently."

This is how you keep a cheap yes from turning into an expensive no. A forecast gives you a confident yes for the price of a spreadsheet, and you find out it was wrong after the build. A small, reversible launch gives you a real yes or a real no for pennies, before the big money moves.

The 70–120% weekly growth was one case, without a control group, and it had help: search engines indexing the translated pages, crypto communities passing articles around Telegram groups and Discord servers, and an automated pipeline that translated the full daily output instead of a few highlights. I wouldn't promise that number to anyone. What I'd stand behind is the method that told us where to launch before we spent anything meaningful.

What this looks like for a software team's next audience

You probably aren't translating news. But if you're weighing a new audience for your product, the same questions apply, and most of them can be answered with data you or your market already has.

Where are people already doing the thing your product would help with, without you? Look for existing spend, workarounds, usage of adjacent tools, and traffic or sign-ups arriving from places you never targeted. Which competitors are already serving this audience, and what does their persistence tell you about whether it pays? And what's the smallest version of the launch you could run, with a clear definition in advance of what result counts as a yes and what counts as a no?

If the behavior is there, the launch confirms it and you've earned the right to invest properly. If it isn't, you've found out cheaply, and validation that ends in a no is still a result worth having.

Those readers were already out there, reading other outlets in their own language. Our job was to notice.