Which locales actually show ROI, and the arithmetic nobody checked

The return figure that traces to nothing

Every localization business case starts from the same sentence. Twenty-five dollars back for every dollar spent. It is on vendor blogs, on agency landing pages, inside ROI calculators, on the third slide of the deck that asks for the budget. It is always attributed to the same market research firm, and it never carries a link to a document.

Go looking for that document and you do not find it. The pages carrying the figure cite each other, or cite the firm by name with nothing to open. The firm does publish on localization ROI, and what it publishes is a method for building the case rather than a measured multiple.

Set aside whether the number is plausible. Ask what it would have to contain to be a measurement at all. A return figure needs four things: a cost that was actually incurred, a revenue change, a reason to attribute the second to the first, and a window over which both were counted. The twenty-five to one figure publishes none of them. A ratio with no denominator is not a finding. It is a slogan with a decimal point in it.

This matters more than a pedantic footnote, because the question underneath is a real one and it has a real answer. Someone has to decide which languages to ship next quarter, in what order, and whether the third one is worth it at all. The published evidence can help with that. It just says something narrower, and quieter, than the slide.

The survey watched preferences, not purchases

The largest consumer dataset in this field is a survey of 8,709 people across 29 countries, run in 2020 ¹. It is a serious piece of work with a real sample, and it is where almost every consumer-facing localization statistic in circulation comes from.

Its headline results are the ones you have seen: 76% of online shoppers prefer to buy products with information in their own language, and 40% say they will never buy from websites in other languages ¹.

The country breakdown is more useful than either of those, and it is the part that rarely gets quoted. Germany leads on buying exclusively at local-language sites, at 57%. Asked to choose between two similar products, preference for local-language information runs at 94% in Taiwan, 92% in South Korea and China, 90% in Japan, and 88% in Indonesia. At the other end, 48% of Romanian respondents were satisfied with information in English¹.

That is an ordering which is genuinely decision-shaped. If you can serve two more locales this year, it tells you which two are likely to be sitting behind the most friction.

What it is not is revenue. Every one of those numbers is an answer to a question. Nobody’s card was charged, no cart was abandoned, no renewal lapsed. Stated preference and revealed preference come apart in every field that has checked, and they come apart most where the stated preference costs the respondent nothing to hold. Saying you would never buy from an English-only site is free. Closing the tab when the alternative is thirty dollars more expensive is not.

Every figure arrives without the one beside it

Once you start reading the sources rather than the citations, the same pattern turns up in all of them. The larger number travels. The number in the sentence beside it stays home.

Start with the survey above. On the same page as the 76% are two findings nobody repeats. 69% of respondents said they would choose a major global brand over one that offers information in their own language. And 65% said they prefer content in their language even when the quality is poor ¹.

The first cuts against the whole localization case: brand beats language for most people, most of the time. The second cuts against the argument localization vendors make immediately after quoting the first one, that a bad translation is worse than no translation at all. Both are in the source. Neither is in the deck.

Now the app store figure. A 2012 study of 200 iPhone apps and games found that in the week after a native language was added, downloads rose 128% and revenue rose 26%³. The 128 is one of the most quoted numbers in this industry. The 26 is in the same sentence and is the one that decides whether the work paid for itself. Downloads are what rise when a store’s algorithm starts matching your metadata to a query. Revenue is the part that had to survive somebody actually using the app.

That study has a second problem. It was published by a company that no longer exists, and the report is not retrievable. Every current citation of it is a citation of a summary of a summary. The figure has been in circulation for fourteen years, describing an app store that has been rebuilt twice since, and there is no longer a document behind it to check.

Then the ROI surveys. One widely quoted pair of figures says 96% of marketers report positive ROI from localization and 65% report a return of 3x or greater. The sample is 415 respondents, director level or above, responsible for translation at companies with more than 100 employees, in France, Germany, Japan and the United States, surveyed for a translation vendor ².

Read that description again. Those are people who already bought, grading their own purchase, in a survey commissioned by the seller. The result is not evidence about whether localization pays. It is evidence about how buyers of localization feel about localization, which is a real thing to know and a different thing entirely.

None of this is a conspiracy. It is selection. A page assembling statistics is optimizing for the statistic, and the qualifier is always the less quotable half of the sentence. The practical consequence is simple: when a figure reaches you without the number that was next to it, the missing one is usually the one that would have priced the decision.

Ranking by economic weight, and what happens when you divide

The best idea in this literature is that speakers are not buyers. Ranking target languages by how many people speak them puts Hindi and Bengali near the top of a list that has nothing to do with who will pay you. The alternative that has caught on is to rank by the online economic activity reachable in each language, which vendors call eGDP.

The most cited application of it is a study of African languages, and it is worth looking at closely because it is doing something right. Of over 2,000 languages spoken on the continent, an examination of 2,800 major brand sites found 22 receiving any support at all⁴. The published figures put the ordering by economic weight distinctly out of step with the ordering by population.

The headline the study draws from its own table is the Afrikaans case, and it is the line that gets repeated everywhere: Afrikaans offers 50% more accessible eGDP than Portuguese, while Portuguese has 40% more speakers ⁴.

The speaker half checks out. Eleven million against eight million is about 40% more.

The money half does not. Seventy-one billion against fifty-three billion is 34% more, not 50%. There is no rounding of those two published numbers that produces a half. The claim and the table contradicting it are on the same page, a few hundred pixels apart.

The finding survives it. Afrikaans really does carry more reachable economic activity than Portuguese on fewer speakers, and that really is the interesting thing about the dataset. But this is the most repeated anomaly in the field, and everyone who repeated it had the arithmetic in front of them and did not do it.

There is a larger caveat that matters more than the division. eGDP measures the aggregate online economic activity reachable in a language. It is not your addressable market, it is the ceiling of everybody’s. A number that is identical for a payroll product and a mobile game is not telling you about either of them.

The one causal estimate, and what it actually measures

There is one causal estimate in this area that meets the standard the rest of the field claims to meet. In 2014 eBay replaced Bing Translator with its own machine translation system for search queries and item titles, rolling it out by region on dates the researchers did not choose. That is a natural experiment, and the resulting difference-in-differences study ran on actual transactions across a platform that mediated over fourteen billion dollars of trade⁶.

The result is the strongest evidence anyone has that language barriers cost real money. It is also smaller than every vendor figure in this post, and it moved.

Both numbers are real and neither is an error. That is what peer review looks like when it works. It is worth sitting with, though, because it puts a floor under how much precision anyone should claim here: the single most careful measurement in the field moved by a third between draft and print, and the figures being quoted around it are given to one decimal place with no interval at all.

The more important detail is what the study measured. The change was an upgrade of translation quality on search queries and listing titles, judged by linguists as correct 91.4% of the time against 84.4% for the system it replaced ⁶. The eBay interface itself was already presented in local languages, and that part was not what changed.

So the best causal evidence that language barriers suppress revenue is about improving machine translation of user-generated listing text on a marketplace that was already localized. It is not a measurement of whether translating your onboarding flow into German pays for itself. It is adjacent to your question and it is honest about being adjacent, which is more than can be said for the numbers that answer your question directly ⁵.

The evidence you already own

Here is the part the industry figures are standing in for. You already hold better evidence about your own product than any of them, and it is sitting in systems you pay for.

Then run the experiment the literature did not run for you. Take the two locales sitting highest on the friction ordering, ship one of them, hold the other, and read the difference over a quarter. It is not a randomized trial and it does not need to be. It needs to be better than a ratio you cannot source, and that is a low bar.

We have written about what localization actually costs now, and the cost side is the half that has moved most. When translating a locale took a quarter and a purchase order, quoting a return figure instead of measuring one was a reasonable economy. It is not one any more. The reason to distrust twenty-five to one is not that it sounds too good. It is that you can produce a real number for your own product, in one quarter, for less than the meeting spent arguing about whose figure to believe.

References

1. 1.CSA Research, 2020 Survey of 8,709 Consumers in 29 Countries Finds That 76% Prefer Purchasing Products with Information in Their Own Language csa-research.com
2. 2.DeepL and Regina Corso Consulting, 2023 The state of translation and localization in 2023-2024 DeepL white paper, 415 respondents
3. 3.Hill, 2015, citing Distimo, 2012 What is localization and why should I care? Game Developer, March 2015
4. 4.Lionbridge, 2023, with data from CSA Research Embrace the Online Opportunity of African Languages lionbridge.com, updated June 2023
5. 5.Brynjolfsson, Hui and Liu, 2019 Does Machine Translation Affect International Trade? Evidence from a Large Digital Platform Management Science 65(12), 5449-5460
6. 6.Brynjolfsson, Hui and Liu, 2018 Does Machine Translation Affect International Trade? Evidence from a Large Digital Platform NBER Working Paper 24917