hreflang fails as a set, not a page, and your per-page audit will never see it

The arithmetic is why this defect class survives review. Five language versions are not five checks. Each of the five pages has to name itself as well as its four alternates ^[2][5], which is twenty-five entries across the cluster ^[1], and every entry has to be answered by a matching one on the page it points at ^[3]. Add a sixth language and the count is thirty-six ^[2]. Miss a single reciprocal and Google has grounds to distrust the whole group ^[3], while each page in it continues to look correct when you view its source ^[4].

There is no shortage of ways for one entry to go bad without looking bad. `en-UK` is invalid because the region code comes from ISO 3166-1, which lists GB and has no UK; the author of the post calls it easily the most common hreflang error on the web ^[6]. Targets have to be final URLs, so an entry pointing at http when the server redirects to https, or omitting a trailing slash the server adds, resolves to a redirect rather than a page and does not count ^[7]. That one, according to the post, bites hardest after a migration, when the site still works fine for humans ^[7]. And the canonical conflict is the quiet one: hreflang says five pages are alternates, the canonical tags say four of them are copies of the fifth, and the two signals cancel ^[8]. Declaring the same alternates in HTML and in a sitemap is another reliable way to produce conflict, which is why the post says to pick one method ^[9].

The disclosure in the post is the more useful part, because the author's company builds a translation widget that generates hreflang tags for other people's sites ^[15]. That widget injects the tags from JavaScript after page load ^[10]. Google renders JavaScript, so those tags can be picked up, but rendering is a second pass on no guaranteed schedule, server-rendered tags are stronger, and other search engines are less reliable about running JavaScript at all ^[11]. The alternates are also built as a query parameter, `?lang=de`, which avoids duplicating anything in the CMS and avoids re-translating when a paragraph changes, at the cost of a clean path per language; a query parameter can be indexed, but `/de/` is a stronger signal ^[12]. On that basis the author points readers whose main goal is multilingual organic search towards GTranslate, which serves translated sites with translated URLs, or WPML and Polylang, which create a post per language ^[13]. The widget's case is the other one: visitors reading in their own language, no second copy of every page, hreflang as a useful signal rather than the whole strategy ^[16].

The first diagnostic the post recommends is worth reading as a statement about tooling rather than a tip: view source on two different language versions and compare them ^[14]. The unit of inspection is the pair, not the page. Anything that grades one URL at a time is measuring the wrong object, which is exactly how a cluster can be discarded in full without a single failing check anywhere in the report ^[1].