Prior art has no language requirement. A Japanese utility model published decades ago, a German Gebrauchsmuster nobody ever translated, a Chinese application that never left China — any of them can be cited against your US claims, and any of them can sink a patent you already paid for. Examiners at the USPTO search foreign collections routinely. If your own search stopped at English results, you searched maybe half the world's technical disclosure and told yourself the field was clear.
The good news: you do not need to read Japanese to evaluate Japanese prior art. Patent families, classification codes, and machine translation tuned on patent text make cross-language searching workable for a founder with a laptop and an afternoon.
Why the language barrier is smaller than it looks
Three structural facts work in your favor.
First, patents are formulaic — field, prior art, object, summary, drawings, claims, in that order, in every jurisdiction. Once you know the shape of the document, a rough translation is enough to navigate it, and our guide on how to read a patent applies almost unchanged.
Second, most commercially serious inventions are filed in more than one country. A Chinese applicant who cares about the US market files here too, and that family member is in English.
Third, classification codes are language-neutral. A code like F16C 33/00 means the same thing in Tokyo, Munich, and Alexandria, Virginia.
Start with the patent family, not the document
A patent family is the set of applications worldwide sharing a priority filing — the same invention, filed in different offices. Espacenet shows the INPADOC family for any document, and it is the most useful button in free patent searching.
The routine: find the foreign document, open its family, look for a US, GB, AU, or EP member, and read that instead. The claims will differ — each office forces amendments — but the disclosure and drawings are close enough to tell you whether the reference is a problem. If the family has only a JP or CN member, that itself is information: the applicant never sought US rights, so the document is prior art but not an infringement risk. That distinction matters when you are running a freedom-to-operate search rather than a patentability search, because an unpatented-in-the-US invention is free for you to practice here.
Search by classification, not by keyword
Keyword searching is exactly where language defeats you. A Japanese application will not contain your English words, and a machine translation of the abstract may render your term three different ways. Classification searching sidesteps the problem entirely.
Find three or four US patents that are close to your invention, note their CPC codes, then search that code with a date range and no keywords at all. You will get every document any office ever put in that bin, in every language. The technique is covered in more depth in our walkthrough of searching patents by CPC and IPC classification, and it is the single skill that most improves a first-timer's search quality.
Drawings help enormously here. Patent figures are language-independent, and skimming the first drawing sheet of 200 classification hits eliminates most of them fast.
Which databases handle foreign text well
| Tool | Coverage strength | Translation |
|---|---|---|
| Espacenet (EPO) | Broadest — 100+ offices, deep historical files | Patent Translate, tuned on patent corpora |
| WIPO Patentscope | PCT applications plus many national collections | WIPO Translate, plus cross-lingual expansion |
| Google Patents | Large, with full-text search of translated text | Automatic, searchable in English |
| USPTO Patent Public Search | US only, but shows examiner-cited foreign art | None — use it to find leads |
Google Patents deserves a special note: it indexes machine-translated full text, so an English keyword search does return Chinese and Japanese documents. That is convenient and also dangerous, because you are searching a translation of unknown quality. Use it to generate leads, then verify in Espacenet. Our step-by-step guides to using the EPO's Espacenet database, searching international applications in Patentscope, and searching Google Patents each cover the interface details.
How far to trust a machine translation
Machine translation of patents is very good at the description and unreliable at the claims. The description is prose written to explain; errors there are usually obvious and recoverable. The claims are a single sentence of nested legal grammar where a mistranslated "comprising" versus "consisting of" flips the scope from open to closed.
A practical rule: use machine translation to decide whether a document is relevant, never to decide whether it blocks you. If a translated reference looks like it reads on your invention, that document graduates to a human translation — typically $0.15 to $0.35 per word, or roughly $400 to $1,500 for a full patent, less if you order only the claims and the relevant passages. Ordering claims-only translation for the three or four documents that actually matter is the cost-effective move.
Country-specific traps
- Japan and China both grant utility models — short-term rights examined only for formalities. They are rarely surfaced by casual searching, they are full prior art, and in China they are enforceable. Search them deliberately.
- Germany's Gebrauchsmuster works the same way and sits in the DE collection alongside full patents.
- Korean applications often pair detailed drawings with thin abstracts, so judging by abstract alone will mislead you.
A search sequence that works
- Define the invention in functional terms, then pick two or three CPC codes that cover it.
- Run the classification search in Espacenet with no language filter and skim drawings.
- For every candidate, open the INPADOC family and read the English member if one exists.
- Machine-translate the rest and triage on the description.
- Repeat the same query in Patentscope and Google Patents — coverage differs at the edges.
- Send the survivors to a human translator, claims first.
Budget a full day for a decent first pass. That is roughly consistent with what we describe in our piece on how long a patent search takes, and it is time spent before you commit five figures to drafting. If the search does turn up something uncomfortably close, the options are laid out in what to do when you find a similar patent — designing around it is more often possible than founders assume.
Projects House runs prior art and freedom-to-operate searches across non-English collections as part of early-stage product work, and folds the findings straight into design decisions rather than handing you a report and walking away. If you want a second set of eyes on a field before you file, reach us through the contact form.