Accept-Language negotiation and its limits

What it is

RFC 9110 defines Accept-Language as a list of language ranges with optional quality values: da, en-gb;q=0.8, en;q=0.7 means "I prefer Danish, but will accept British English and other types of English". Order among equal weights cannot be relied upon. Matching is delegated to RFC 4647: basic filtering treats a range as a prefix (de matches de-CH), lookup finds the single best tag by truncating the range from the right (de-CH-1996, then de-CH, then de). An absent header expresses no preference and leaves the choice to the server. A user agent that offers no user control over the preference MUST NOT send the header, and the RFC flags the privacy cost of sending full preferences on every request. The response names its language in Content-Language and, because it varies by request header, needs Vary: Accept-Language for caches.

Why it matters

The header describes a device or browser configuration, not necessarily the reader: shared computers, corporate images and travellers all send the wrong thing. Google's documentation states that its crawler sends requests without Accept-Language, warns that it might not crawl, index or rank all content for different locales, and recommends separate URLs per locale with hreflang annotations; content chosen only by header may therefore go unseen in some languages. Scripts and agents usually send nothing. A cache that ignores Vary serves the first language it stored to everyone.

How to apply

• Give every language its own URL (/de/, /en/) with hreflang alternates; use the header only to choose where a visitor to the language-neutral root goes, with a temporary redirect and Vary: Accept-Language, never a cached permanent one.
• Implement lookup: exact tag, then primary language, then the site default; treat * and an absent header as the default.
• Store an explicit choice (cookie or profile) and let it override the header on every later visit.
• For APIs, accept a documented lang parameter that overrides the header, and always return Content-Language.
• Do not infer language from IP address; countries are multilingual.

Pitfalls

A request for en-GB against a site that only has en fails under basic filtering, because the tag is shorter than the range; lookup handles it, and RFC 9110 notes that user agents ought to suggest adding en to such a list. Equal quality values with implied order differ between clients. Language negotiation says nothing about region-specific formats; treat locale for dates and numbers separately.