Your Site Is Translated. Does Each Page Say So?
Google Ads now infers language rather than asking for it. A five-step afternoon check of how your multilingual pages declare what language they are in.
The site is translated, so the language of every page is obvious. That is the objection, and it is a reasonable one. A German buyer who lands on the German product page can see perfectly well that it is in German. The trouble is that the buyer is no longer the only reader who needs to know, and most of the others do not look at a page the way a person does.
The prompt for writing this is a change on the advertising side. On 13 August 2026 Google announced that, starting in September, campaign-level language targeting is being removed from Search campaigns. In the words of the Google Ads help page, Search ads "will automatically match based on the language of your ads", and Google says it prioritises ads that match the language of the user's search term. An advertiser used to tick a box. Now the system infers.
This piece is not about ads. It is about the fact that more and more systems are being asked to work out what language your pages are in, and that none of them has the context your buyer has. On a well-kept multilingual site that inference is trivial. On a site that grew by accretion, with a CMS migration or two and a translation agency that handled the body copy but not the template, it is less so. What follows is a procedure you can run in an afternoon with a browser and the view-source shortcut.
1. Read the lang attribute on every language version
Open one product page in each language, view the source, and look at the first element: <html lang="...">. Write down what it says for each version.
The W3C's guidance is unambiguous: always put a lang attribute on the html element, use a BCP 47 code such as de or fr-CH, and never use a meta http-equiv="Content-Language" element to do the job instead. The fault you are looking for is dull and common: every version of the site declaring lang="en", because the attribute lives in a shared template that nobody localised.
A point of precision here, because it is often got wrong. Google's own documentation on localised versions states that Google does not use hreflang or the HTML lang attribute to detect the language of a page; it uses algorithms, working on the visible content. So a wrong lang will not, on Google's word, mislabel the page in Google Search. Fix it anyway. It is the one place in the markup where a page says what language it is in, and any software that would rather read a declaration than guess will read that one. A declaration that contradicts the content is worse than none.
2. Count the hreflang lines
Still in the source, find the <link rel="alternate" hreflang="..."> lines in the head. They may instead be in HTTP headers or in the XML sitemap; Google accepts all three methods, so check wherever yours are.
The arithmetic is simple and makes a good test. If the page exists in five languages, each of the five versions should carry the same set: five hreflang lines, one of them pointing at itself, plus an x-default if you use one. Six lines, identical on all five pages. Google's documentation is explicit that each language version must list itself as well as all the others, and that if two pages do not both point to each other, the annotations are ignored. One German page that forgot to link back to the French one quietly cancels the pair.
Three further things to check while you are there. The URLs must be fully qualified, with https:// in front. The codes must be real: de on its own is valid, en-GB is valid, a region code on its own is not. And x-default is optional. Google describes it as designed for language selector pages, a fallback for visitors whose language you do not serve. On an export site with a sensible default version, it is usually worth having.
3. Read the page as someone who speaks only that language
This is the step that no validator does for you, and it is where multilingual industrial sites most often come apart.
Go through the German page and mark everything that is not in German. The usual suspects:
- Navigation and footer, left in English because they live in the template.
- Specification tables, pasted from the English master and never translated, so the German page describes its product in German and specifies it in English.
- Downloads: the datasheet, the installation manual, the certificate, available only as an English PDF behind a German link.
- Buttons, form labels and the cookie banner, which the translation agency never saw.
Google's advice on multilingual sites is to keep a single language for content and navigation on each page and to avoid side-by-side translations. The broader point does not depend on Google. Text in the wrong language on a page is text in the wrong language for any reader, human or machine. Nothing documented says that AI assistants read hreflang, and it would be foolish to assume they do. What an assistant retrieving your German page certainly gets is its text, and if the part that decides a technical purchase is in English, that is what it gets.
4. Look at the URL
Put the five versions of one page side by side in the address bar. The language should be legible in each: a country domain such as example.de, a subdomain such as de.example.com, or a subdirectory such as example.com/de/. Google lists all three as workable and advises against URL parameters such as ?lang=de.
The version to worry about is the one where there is no separate URL at all: a single address that shows German or English depending on a cookie, a browser setting or the visitor's IP. Google recommends letting people choose their language through links rather than redirecting them automatically, and notes that its crawlers mostly come from the United States. A page that only exists in German for visitors who arrive with the right settings, for most automated readers, does not exist in German.
5. Check the title and meta description
Last, and quickest. In the source of each language version, read the <title> and the meta name="description". These are frequently the last things to be localised, because the CMS fills them from a default, or because they sit in a separate SEO field the translators never opened. A German page with an English title is making two contradictory statements about itself in the first few lines of its code.
Count as you go. Five languages, a dozen key pages, five checks per page: it is a spreadsheet with a few hundred cells, and a surprising proportion of them will be wrong for reasons nobody decided.
What this does not tell you
A site that passes all five steps is correctly labelled. That is all. It says nothing about whether any assistant names you when a buyer asks, in German, who supplies what you supply. That is a separate measurement, and it is the one PSentry runs: the same questions across ChatGPT, Claude, Gemini and Perplexity, in each language and market you sell in, twice a month. It does not audit your markup, and it will not tell you which of your pages carries the wrong lang. The afternoon above is yours to spend.