Published on November 15, 2026
If you sell in Japan, China, Korea, Thailand or Arabic-speaking markets, translation is rarely what fails. Encoding, normalization, text direction, field lengths and expectations are. A field report from Japanese marketing, from my own WooCommerce shop through international markets to marketplaces and crowdfunding, with a checklist for your PIM.

Before we get to encoding, scripts and data models, a personal part. I worked in marketing in Japan, mostly in international marketing for shops and online retail. Back then, I first built my own shop on WooCommerce and served international markets through it. Later, marketplaces and crowdfunding platforms followed. When someone asks me today what took the most work in all of this, my answer is not “the translation”.
The translation was usually the most manageable part. The hard part came before and after it: when a cleanly translated file showed nothing but question marks after upload, when a product image no longer found its article even though the file name was right, or when a good text still did not sell because it answered the wrong questions.
I started with WooCommerce. For a first shop, that was the obvious choice: quick to set up, flexible, with plenty of extensions. From there I moved into international markets, and that is exactly where I saw how early the problems begin, long before a marketplace comes into play.
Product import runs through CSV, and WooCommerce expects UTF-8 for it. If you maintain your spreadsheet in Excel and save it as a plain CSV, you may import a different encoding, and product names turn into garbled text. With umlauts, you notice immediately. With Japanese text, only people who read Japanese notice.
Then the URLs. WordPress builds the slug from the title and stores non-Latin characters percent-encoded in a column of 200 characters. A Japanese character takes three bytes in UTF-8, so nine characters once percent-encoded. After a little over twenty Japanese characters, the column is full and WordPress truncates the slug. In every log file and every shared link, the address shows up as a long chain of %E3%81 and similar blocks.
And finally, the languages themselves. In WooCommerce, multilingual support comes through extensions, and the language versions of a product then often live side by side as separate entries. When a price, a dimension or a variant changes, every language version has to be updated. With every additional market, it is not just the text that grows, but the number of places where the same detail can drift apart.
What I take from this: The shop system is an output channel, not the source of product data. As soon as more than one market is involved, the truth about a product belongs in one place that feeds the shop, the marketplaces and the crowdfunding pages.
So much for my own experience. From here on, this is about the language and script problems behind it, and they affect every retailer and manufacturer selling into non-Latin markets.
My thesis: if you sell into non-Latin markets, you are not translating texts, you are translating data records. The expensive mistakes do not happen in the language but in data management: in encoding, normalization, text direction, field lengths and buyer expectations.
The Latin script is forgiving. A German or English product title survives almost any chain of ERP, Excel, PIM and marketplace, even if an umlaut gets lost somewhere. With Japanese, Chinese, Korean, Thai and Arabic, every step is a chance to damage characters without anyone noticing who cannot read the language.
That is why I think in five layers. Each has its own class of errors, and each language in the title illustrates one of them particularly well:
In Japan, two worlds still meet. Modern systems run on UTF-8, but many tools, templates and processes still use Shift_JIS or its Windows variant CP932. Even the CSV specification of the government business portal gBizINFO allows Shift-JIS or UTF-8 and explicitly excludes half-width katakana.
The classic case is Excel: a file is exported in UTF-8 and opened in an Excel version that reads Japanese CSV files as Shift_JIS. The result is garbled text, called “mojibake” in Japanese. If the file is saved afterwards, the damage is permanent.
More subtle is the wave dash in specifications like “500ml〜1L”. The same visible character has two Unicode code points: WAVE DASH (U+301C) and FULLWIDTH TILDE (U+FF5E). Microsoft’s mapping table for Shift_JIS and the JIS mapping treat the character differently. Depending on which conversion a file passes through, the wave dash becomes a different character or a question mark. In size, quantity or age information, that is not a cosmetic flaw but a wrong product claim.
China handles the topic through a mandatory national standard: GB 18030 defines which characters software must support, and the GB 18030-2022 edition has applied since August 1, 2023. The Unicode Consortium explicitly warned developers about changes from the previous edition because the mapping of individual characters shifted.
What I take from this: Internally, everything belongs in UTF-8. Conversion happens only at the edge, on export for a specific channel, with a check for characters that cannot be represented in the target encoding. And Excel is not a transport medium for Japanese product data.
Korean syllables can be stored in Unicode in two ways: as a complete syllable (NFC) or decomposed into their components, the jamo (NFD). “한” is either one code point (U+D55C) or three (U+1112, U+1161, U+11AB). On screen, both look the same. To the computer, they are different strings.
macOS stores file names in decomposed form, Windows in composed form. If you name product images on a Mac and pass them on through a Windows system, you get file names that fall apart into single letters there and no longer match the product. The same happens in search, duplicate checks and sorting.
In Japanese, the related problem is full-width and half-width characters. “500ml” and “500ml” mean the same to people, but not to a matching rule. Compatibility normalization (NFKC) aligns them, but it also changes characters that were chosen on purpose, such as circled digits. Normalization therefore needs a rule per field, not a global switch.
What I take from this: Bring everything to NFC on import, file names included. For key fields such as SKU, EAN and sizes, add a defined width rule. Free text stays as it is.
Arabic runs from right to left, while numbers and Latin parts inside it run from left to right. Product data consists almost entirely of such mixtures: model numbers, dimensions, brand names, units. The Unicode bidi algorithm usually arranges this correctly. But the W3C describes exactly the cases where it needs help: part numbers, formatted numbers, brand names and lists. In other words, exactly what specification tables are made of.
Typical symptoms are a model number like “AB-500/2” whose parts suddenly appear swapped, a punctuation mark on the wrong side, or a unit that jumps away from its number. The fix lies in rendering: parts with a different direction are isolated, in HTML with the dir attribute or the bdi element. That only works if model number, value and unit are separate fields in the record and not buried in running text.
On top of that come decisions someone has to make: Western or Arabic-Indic digits, Modern Standard Arabic or a regional tone, mirrored layouts in images and graphics.
What I take from this: Structured attributes instead of concatenated strings, and a target market decision that is recorded in the data model rather than in the head of a single translator.
Thai does not separate words with spaces. Spaces separate phrases. Software and browsers usually determine where a word ends with dictionaries, and according to the W3C even then there are different, equally defensible results.
For product data, this means a title truncated to a fixed number of characters can end in the middle of a word. On top of that, vowel and tone marks are separate code points placed above or below a consonant. If you count and cut by code points, in the worst case you separate a tone mark from its letter.
A second length problem affects all scripts in this article, and the WordPress slug above is just one example: in UTF-8, a Latin letter takes one byte, while a Japanese, Chinese, Korean or Thai character takes three. A field limited in bytes therefore holds only about a third as many characters, and truncating by bytes can split a character in half.
What I take from this: Define length rules per channel in visible characters and in bytes, truncate only at word boundaries, and write short titles deliberately rather than cutting them automatically.
No. Simplified Chinese for mainland China and Traditional Chinese for Taiwan or Hong Kong use different characters, sometimes with different vocabulary. A stainless steel insulated bottle becomes 不锈钢保温瓶 or 不鏽鋼保溫瓶. A language field “zh” is not enough. The data model needs language, script and region, such as zh-Hans and zh-Hant.
Even subtler is an effect many people only notice when they sell in Japan: Chinese, Japanese and Korean characters of the same origin share one code point in Unicode. The font decides what the character looks like. The Unicode Consortium therefore recommends showing Japanese text in a font intended for Japanese. A Japanese product text without a language tag can appear in a Chinese font and then looks foreign to Japanese buyers, even though not a single character is wrong.
What I take from this: Locale codes with script and region in the PIM, a language tag that travels all the way into shop, data sheet and image, and fonts that fit the target market.
For me, this is the biggest difference between translating and localizing, and I learned it during my time in Japanese marketing.
Japanese buyers read product pages thoroughly. They expect exact dimensions, weight, material, country of origin, contents of the package, care instructions and warranty, and they notice when a detail is missing or only approximate. A German description that relies on mood does not get better through a clean translation if the specification underneath stays thin. Tone is a decision of its own: polite and precise rather than loud.
On Japanese crowdfunding platforms, the project page is often a long story told in images. For the campaign, that works. For the product data afterwards, it is a trap: text locked inside graphics cannot be searched, translated or handed over to a marketplace. If you want to move into shop and marketplace after the campaign, you end up typing the same information a second time, this time from image files.
And there are decisions that have to be made once per market and then kept: how the brand name is written in katakana or hangul, which size systems and units apply, which plugs and voltages are stated. If three people transcribe the brand name three different ways, the brand has three names in that market.
What I take from this: Complete the specification first, then write the text. Information belongs in the PIM as data, not in images. And transcriptions, units and terms live in a glossary that people and AI agents follow alike.
In ainavio, language versions are part of the data model: attributes are maintained in multiple languages on one product, not as separate copies per language. The AI agents work on this model.
Whatever AI writes lands as a draft and is approved in the PIM. For non-Latin markets, this is the most important point: the work shifts from typing to reviewing where language sense and market knowledge matter. Clarify in advance which channels in your target markets are connected and which encoding and field lengths they expect.
A single “〜” in a quantity looks insignificant. In fact, a whole chain of decisions hangs on it: which encoding stores it, which table converts it, which field follows which rule. If you want to sell in Asia and the Arabic-speaking world, you do not need better translations first. You need product data that carries every script intact all the way to the buyer.
Usually the encoding is mixed up: a UTF-8 file is read as Shift_JIS or the other way around, often when it is opened in Excel. In addition, individual characters such as the wave dash are mapped differently depending on the conversion table.
WordPress stores non-Latin characters in the slug percent-encoded in a column of 200 characters. A Japanese character takes nine characters there, so the limit is reached after a little over twenty Japanese characters. Short, deliberately set slugs per language avoid this.
Internally, yes. But some channels, templates and legacy systems expect other encodings. In that case, convert on export and check that every character can be represented in the target encoding.
Yes. Taiwan uses traditional characters and partly different vocabulary than mainland China. In the PIM, this belongs in its own locale such as zh-Hant.
Because left-to-right parts inside right-to-left text are reordered by the bidi algorithm. The fix is separate fields for model number, value and unit, plus isolated rendering with dir or bdi.
No. Thai has no spaces between words, and vowel and tone marks are separate code points. Truncate at word boundaries and check length in visible characters and in bytes.
Talk to us about your range, your channels and your legacy systems. We will show you where PIM with AI fits into your setup.
Conclusion: In non-Latin markets, product data rarely fails at translation. It fails at five layers before and after it: encoding, normalization, text direction, length and word boundaries, and language variant and buyer expectation. If you work with UTF-8 and NFC throughout, keep locales with script and region, store values as structured attributes, define length rules per channel and keep information as data rather than in images, your products reach Japan, China, Korea, Thailand and Arabic-speaking markets intact. A PIM with AI takes over the busywork, while approval stays with people who know the language.