I'd prefer static analysis that extracts strings from calls like t("Check out our Jobs page {{url}}!") and auto-populates the appropriate locale files/databases, etc. Referencing keys all over the place seems like a poor substitute to just writing the content in a natural way and having the tools do the heavy lifting for you. For instance, I know that Twitter and Facebook both handle I18N content in such a way, and you can see the text-strings that they use in the translation dashboards.

There's also a good technical overview of Facebook's I18N framework here: https://www.quora.com/Facebook-Internationalization/What-was... .