Hreflang still matters. Just not to the engines you'd expect.

What hreflang actually promises

A hreflang annotation, whether placed in a page's head, in an HTTP header, or in a sitemap, does one narrow job: it tells Google that a given URL is a specific language or regional version of a page, and names the other versions a searcher in a different language or region should see instead. The x-default value is a catch-all for anyone who doesn't match a listed language or region. None of that is a ranking factor. Hreflang cannot make a page rank, cannot create a new page, and cannot force Google to index a variant it has already decided to treat as a near-duplicate of another.

Google has restated that limit directly. Responding to a question from SEO consultant Neil McCarthy on Bluesky, reported by Search Engine Journal on 12 May 2025, Google's John Mueller said hreflang "doesn't guarantee indexing, so it can also just be that not all variations are indexed." He added that when two versions share the same underlying language, a French page aimed at France and a French page aimed at Belgium, Google's systems commonly pick one as canonical and report on that URL even while still swapping in the other for matching searches. Hreflang is read as a hint about intent, not an instruction Google is bound to follow.

Two-thirds of multilingual sites still get it wrong

The most recent large-scale measurement comes from Ahrefs. Technical SEO lead Patrick Stox, with colleague Oleksiy Golovko, analyzed 374,756 domains in a study published 10 August 2023 and last updated 12 August 2026, describing it as roughly ten times larger than any prior hreflang study. The finding: 67% of domains had at least one hreflang issue. A missing x-default tag was the single most common problem, present on 56.3% of the sites with any issue, followed by missing self-referencing tags (18%), tags pointing at redirected or broken pages (16.9%), missing reciprocal tags, where page A points to page B but B doesn't point back (15.3%), and tags pointing at non-canonical URLs (8%).

That is not a new problem. Semrush ran a smaller study of 20,000 multilingual sites back in February 2017, and found 75% had at least one hreflang mistake, with conflicting signals inside the page source, mostly a missing self-referencing tag, the single biggest category at 58%. The two studies are six years apart and differ by roughly nineteen times in sample size, use different tools and different error taxonomies, so the 75% and 67% figures are not a clean before-and-after. What they do agree on, nearly a decade apart, is the general shape: a majority of multilingual sites ship hreflang with at least one real error, and better tooling in the years between the two studies has not solved it.

The new wrinkle: hreflang's reach stops at Google's and Bing's own front door

Everything above describes Google. The newer and less settled question is what happens once the searcher isn't using Google at all.

SEO consultant Glenn Gabe ran the most direct test available, published 1 December 2025. He put the same queries, including one for Google's own Search Console documentation and one for a Cloudflare blog post, through multiple engines with French, Italian and Spanish language or location settings applied, and checked which language version of the page each engine actually surfaced. Google's ten blue links, Google AI Mode and Microsoft Copilot all consistently returned the correct, language-matched URL. Gemini was mixed: it sometimes returned no cited source at all, but gave the correct URL when pushed for one. ChatGPT frequently returned the US English version of the page regardless of the French or Italian setting. Perplexity mostly returned the wrong-language URL. Claude did the same, returning the English version where a French or Spanish one existed and was the correct match.

The pattern lines up with how each system is actually built. Google AI Mode and Copilot sit on top of search indexes, Google's and Bing's respectively, that have already resolved the hreflang signal the way they always have. ChatGPT, Perplexity and Claude run their own independent crawling and retrieval, separate from any search index that reads hreflang, and none of their makers, OpenAI, Perplexity AI or Anthropic, publish documentation describing hreflang support one way or the other. A signal built for one piece of infrastructure does not automatically transfer to a different one built from scratch.

What we could not verify

Gabe's test is a single practitioner's own side-by-side run on a small set of queries, written up as a blog post, not a published dataset with a stated query count, methodology appendix or peer review, and nobody has repeated it at the scale Ahrefs or Semrush bring to the implementation-error question. OpenAI, Anthropic, Perplexity AI, Google and Microsoft have all stayed quiet on whether their independent retrieval respects hreflang, x-default, or any other language-targeting signal; none has published documentation either way. There's also no study connecting a correctly implemented hreflang setup to a measurable change in AI citation rate. The Ahrefs and Semrush numbers measure whether the tag is implemented correctly, not whether implementing it changes what an AI engine shows. Treat the direction Gabe found as credible and worth testing on your own site, and treat any specific percentage from a single informal test as unconfirmed until someone runs this at scale.

What to actually do if you run a multi-language site

Keep hreflang. The evidence against it doing anything for the independent AI engines is not evidence that it stopped mattering for Google and Bing, which still account for most of the traffic most sites get, and a two-thirds-of-sites-get-it-wrong error rate means fixing yours is still a real, mostly unclaimed edge on that side alone. Start with the Ahrefs breakdown as a checklist: confirm every page set has a working x-default, confirm every page in a set references itself, confirm none of the target URLs redirect or 404, confirm the reciprocal links are actually reciprocal, and confirm every hreflang target is the canonical URL for that page, not a parameter variant. That is the technical half, and it is what a crawl-level technical review is built to catch.

For the AI half, treat each language version as its own entity that has to be independently discoverable and citable, not a hidden variant only Google happens to know about. That means real, server-rendered localized content rather than a JavaScript language switcher an independent crawler may never execute, consistent entity signals in each language rather than only the primary one, and genuine in-language citations and mentions, the same mechanism covered in local SEO and AI search for geography rather than language. Then test it the way Gabe did: pick a fixed set of prompts, ask them in the target language through each engine you care about, record which URL variant comes back, and repeat on a schedule using the method in measuring AI visibility. Hreflang tells Google what you intended. Only testing tells you what each AI engine actually did with it.

Sources

The hreflang error-rate figures come from Ahrefs, "We Analyzed 375K Hreflang Tags: Here's What We Found," by Patrick Stox and Oleksiy Golovko, published 10 August 2023 and last updated 12 August 2026. The earlier comparison figures come from Semrush, "The Most Common Hreflang Mistakes [Infographic]," by Anna Lebedeva, published 15 February 2017. Google's hint-not-directive position comes from Search Engine Journal, "Google Reminds That Hreflang Tags Are Hints, Not Directives," 12 May 2025, reporting comments from John Mueller. The AI-engine language test comes from Glenn Gabe, "Does Hreflang Matter For AI Search Engines + Multilingual Queries," published 1 December 2025.

Common questions