Hreflang for AI search: audit language alternatives
Check hreflang clusters, return links and canonical conflicts so localized pages have clear language signals and usable destinations for search.
By Crawlito · 5 min read
In short
- Hreflang describes language alternatives; it does not translate a page or override access restrictions.
- Check the complete alternate cluster, including return links and the actual content at each destination.
- A valid language annotation cannot repair a missing translation or a redirect to the wrong region.
On this page
Hreflang helps Google understand which URLs are localized versions of the same content. For AI search readiness, use it as part of a sound international search setup, not as an AI citation switch. The useful audit question is whether each language alternative leads to a readable, appropriate page with consistent signals.
A translated pricing page can fail in several independent ways: its alternate link can be malformed, its canonical can point to the English original, or the server can send every visitor back to the default language. Checking only whether the word hreflang appears in the source misses those failures.
Build clusters around equivalent pages
Start with a content identity rather than a URL prefix. A single product guide may have English, Arabic and French editions, each with a different translated slug. Store those relationships in the CMS so your template does not guess them by replacing /en/ with /fr/.
Google's localized-version documentation describes reciprocal annotations, self-references and fully qualified alternate URLs. It supports HTML, HTTP headers and sitemap annotations. Pick a maintainable method and verify its output instead of maintaining three separate lists that can drift apart.
For an illustrative guide, the HTML might contain:
<link rel="alternate" hreflang="en"
href="https://example.com/en/guides/delivery/">
<link rel="alternate" hreflang="ar"
href="https://example.com/ar/guides/delivery/">
<link rel="alternate" hreflang="fr"
href="https://example.com/fr/guides/livraison/">These addresses are examples, not a recommendation to change your existing URL structure. Review the same cluster on all three pages. A translation unpublished last week should not remain advertised as a working alternative today.
Keep a spreadsheet with content ID, language, URL, publication status and owner. That turns an intermittent SEO defect into a visible publishing dependency. It also gives translators a precise answer when they ask which English article their version corresponds to.
Separate language targeting from canonical selection
Hreflang and canonical tags answer different questions. An alternate annotation identifies a localized counterpart; a canonical expresses a preferred representative among duplicate or similar pages. They should describe a coherent publishing model.
A frequent template mistake is copying the original page's canonical into every translation. Diagnose this at the template level before editing hundreds of records. The canonical tags guide explains the broader duplicate-URL problem; here the specific task is checking whether your localized cluster contradicts itself.
Use this review table for each content family:
| Check | Healthy example | Investigate when |
|---|---|---|
| Main content | French body on the French URL | Only navigation is translated |
| Destination | Intended locale responds directly | It redirects to another language |
| Canonical | Consistent with the chosen locale policy | Every locale names the English URL |
| Return relationship | Counterparts list one another | One template omits all alternatives |
| Availability | Published alternatives only | Draft or deleted translations remain listed |
Do not automatically change every canonical to self-referencing without understanding regional duplicates. Two English editions with almost identical content may need a different approach from a complete French translation. Record the reason for the chosen policy and ask the content owner which edition should serve each audience.
Test the destination without a saved preference
Your normal browser is a poor first test because it may remember a language choice. Open a fresh session and request each alternate directly. Then fetch the response without cookies and compare the final address, status and body language.
curl -sS -L --max-redirs 5 \
-D /tmp/locale-headers.txt \
-o /tmp/locale-page.html \
-w 'final=%{url_effective} status=%{http_code}\n' \
'https://example.com/fr/guides/livraison/'Replace the example URL with a page you own. This command saves the response for inspection; it does not establish how any named search crawler behaves. Read the saved body and look for a distinctive translated sentence. A successful response containing a region selector is different from the actual article.
Check the source before relying on browser screenshots. If the alternate list or translated body arrives only after scripts execute, document that dependency. The comparison in raw versus rendered HTML helps isolate whether the problem sits in delivery or in the content record.
Test a few representative templates: an article, a product, a category and a page without a translation. The last case matters. A missing French product should not silently become a French category page while still claiming to be an equivalent product alternative.
Handle incomplete translations deliberately
International sites rarely launch every language on the same day. A reasonable publishing rule is to advertise only reviewed, live counterparts. Keep an untranslated draft out of the alternate cluster until the destination is ready.
Avoid sending every missing translation to the locale homepage. That may let visitors continue browsing, but it does not make the homepage equivalent to the original article. Provide a clear language switcher that communicates unavailable translations, or let readers choose the original language with an honest label.
Regional differences need editorial attention as well. If the US page offers a feature that is unavailable in Germany, the German text should explain its own offer. Automatically copied pricing tables can create more confusion than a missing annotation. Technical consistency is useful only when the information itself matches the audience.
Give deletion and replacement workflows the same care as publishing. When retiring a translated guide, update its partners' alternate lists, internal links and sitemap entries in the same change. A translation relationship is a small graph: deleting one node leaves stale references elsewhere unless the workflow updates them together.
Turn the audit into repeatable release checks
Create a small fixture set with one complete cluster, one partially translated item, one regional variation and one retired URL. After a deployment, fetch all fixtures and compare them with the expected map. Save evidence of failures rather than only a pass count.
The release check should identify the exact edge that failed: English points to French, but French no longer points back; the Arabic destination redirects; or the German canonical names an unrelated record. These findings are much easier to fix than a general message saying international SEO is broken.
Prioritize defects that affect readers first. Wrong-language redirects and missing article bodies deserve attention before harmless ordering differences in a tag list. Then repair template defects that affect many pages, followed by isolated CMS mistakes.
Use sitemap discovery checks to confirm published language pages also appear in your intended discovery system. Hreflang should not be the only way to locate a translation. Readers need ordinary links from a useful language selector or relevant locale hub.
Keep the audit's conclusion narrow. Passing these checks establishes consistent localized delivery and annotations. It does not establish that every AI service reads hreflang, selects your preferred edition or cites the page. Those claims require separate evidence from the service concerned.
Frequently asked questions
Does hreflang make AI assistants use the correct language?
Do not assume universal support. Hreflang has a documented role in Google's international search setup, but it is not a general instruction that binds every assistant or retrieval system.
Should untranslated pages be included in a language cluster?
Only advertise a counterpart when it genuinely serves that language or region. A draft, empty template or unrelated locale homepage is not an equivalent translation.
Can one broken alternate affect the whole site?
The impact depends on where the defect originates. A shared template can spread an error widely, while an incorrect CMS relationship may affect only one cluster; sample multiple page types to distinguish them.
What should I check after fixing the tags?
Request every destination again without stored cookies, confirm the body language, and verify the complete cluster's relationships. Also check canonicals and normal internal links so the surrounding signals agree.
Start with a representative translated page in the free readiness check, then inspect its counterparts using the same evidence-based process.
Filed under Structure and schema. Topics: hreflang, canonical, international-seo