Distrify

Multilingual Content for AI Search: One Page per Market

September 8, 2026·14 min read
A woman in her early thirties in a forest green sweater compares two open notebooks on a large wooden table in a shared workspace, in front of a bookshelf, in warm afternoon light

If you sell in two countries, one translated page will not make you findable in both. Multilingual content for AI search works when every market has its own page, indexed, written with the words people actually type there, and linked to its equivalents by annotations that Google requires to be reciprocal.

The question usually arrives in a very concrete form: we sell in several countries, how do we know where our product shows up in assistant answers, market by market. The answer starts earlier than measurement. It starts the moment you decide how many pages you publish and where they live. What follows separates what Google states in writing from what is only a field observation, because on this topic the two blur fast.

TL;DR

  • One language, one page, one URL. Google offers four URL structures for targeting several markets and explicitly advises against one of them, URL parameters.
  • hreflang only works both ways. The documentation is unambiguous: "If two pages don't both point to each other, the tags will be ignored".
  • Google uses neither hreflang nor the HTML lang attribute to detect the language of a page. It works it out with its own algorithms.
  • To appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and snippet eligible. No special markup exists for it.
  • Machine translation is not banned as such. The spam policies target mass produced pages with no value, translation appearing only as one example of an automated transformation in that context.

Table of contents

Multilingual content for AI search: the short answer

An assistant answering a question in German mostly pulls from pages written in German. An assistant answering in Portuguese does the same with Portuguese pages. Your English page, however good, does not compete in both tournaments at once.

What makes you eligible in a market comes down to three conditions that stack. The page exists at its own address. It is indexed by the engine feeding the assistant. It is written in the language of the query, with the vocabulary people really use in that country.

None of this is specific to AI. It is the same mechanism described in our complete guide to generative engine optimization, applied to a site that speaks to several countries. The new part is not the method, it is the number of doors you have to keep open at the same time.

What Google documents, and what is only an observation

Plenty of claims circulate about international SEO without a source. Here is the sorting, with the status of each one.

Claim Status
You need a distinct URL per language or region Documented by Google (ccTLD, subdomain, subdirectory structures)
URL parameters such as ?loc=de should be avoided Documented, marked "Not recommended"
hreflang must be reciprocal or it is ignored Documented, explicit wording
Google infers page language from hreflang False, Google states the opposite
Auto redirecting by browser language is good practice Advised against by Google
Adapting content based on visitor IP address Advised against by Google
Special markup makes a page eligible for AI answers False, Google states no dedicated schema is needed
Assistants favour sources written in the language of the question Observable, not documented as a rule
A machine translated page is penalised by default Not documented in that form

The first four lines come from Google's documentation on multi-regional sites and on localized versions. The two lines marked observable should not be presented to your team or your client as rules: they can be noticed, they cannot be cited.

One page per market, not one page translated on the fly

The most tempting shortcut when you are on your own is the translation widget dropped onto the site. The visitor clicks a flag, the text changes, and the page address stays exactly the same.

The problem is mechanical. If there is one URL, there is one page to index, and the engine keeps one version. Google puts it its own way when it covers content adapted dynamically through cookies or browser settings: in that case, "Google might not find and crawl all your variations". Your other languages exist for your visitors, not for the index.

One page per market means one address per market. That is the entry condition, not a refinement. Until it is met, everything else, hreflang included, does nothing.

The reasoning matches what governs product content for AI answers: information a machine cannot reach at a stable address does not exist for that machine, however good it is for a human.

Choosing the URL structure for your markets

Google documents four options and gives the pros and cons of each. The table below reflects them as written.

Structure Example Stated pros Stated cons
Country specific domain example.de Clear geotargeting, server location irrelevant, easy separation of sites Expensive, requires more infrastructure, can only target a single country
Subdomain de.example.com Easy to set up, allows different server locations, easy separation Users might not recognise geotargeting from the URL alone
Subdirectory example.com/de/ Easy to set up, low maintenance on the same host Users might not recognise geotargeting, single server location
URL parameter example.com?loc=de None stated Not recommended, segmentation is difficult, geotargeting unclear

For a site run by one person, the subdirectory nearly always wins. It adds no domain to renew, no certificate, no hosting, and every new language benefits from the history of the main domain. It is the structure of this blog: English pages sit at the root, French pages under /fr.

A country specific domain earns its cost when the brand genuinely differs from one market to the next, or when a legal or logistics constraint really separates the entities. Otherwise it is a fixed monthly cost for a gain you will not measure for a long time.

Want one campaign to go out in several languages without you holding the calendar? Join the waitlist

hreflang: reciprocity is not optional

hreflang is the annotation that tells Google: this page has an equivalent over there, in that language. It has three documented implementations, your choice of a link tag in the HTML head, an HTTP header, or a dedicated element in the XML sitemap.

The rule that breaks most setups fits in one sentence of the documentation: "If two pages don't both point to each other, the tags will be ignored". If your French page points to the English one but the English one does not point back, both annotations fall. You do not get half a result, you get nothing.

Two useful details. Each version must reference itself in addition to referencing the others. And the code is a language in ISO 639-1 format, optionally followed by a region in ISO 3166-1 alpha 2 format: a country code on its own is not valid.

Finally, one belief worth dropping: Google writes that it uses neither hreflang nor the HTML lang attribute to detect the language of a page, and that it determines it algorithmically. hreflang points the right user to the right version, it does not declare a language and it does not win positions.

Do not redirect your visitors automatically

Google's documentation is blunt: "Avoid automatically redirecting users from one language version of a site to a different language version". On geotargeting it adds: "Don't use IP analysis to adapt your content".

The reason is about what a crawler sees. A crawler arriving from a United States address and force redirected to the English version will never see your German pages. You published ten pages per market and made nine of them invisible with three lines of configuration.

The good practice is unglamorous and it holds: let every URL serve its own content, and offer a visible language selector. The visitor chooses, the crawler sees everything.

Machine translation and the line Google draws

Many founders believe translating automatically exposes them to a penalty. That is not what the spam policies say. What they define is scaled content abuse: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users".

Translation appears there only as an example of an automated transformation, next to synonymising, when it is used to mass produce pages with no value. The criterion is not the tool, it is the intent and the result for the reader.

In practice the line is easy to hold. A translation that has been reviewed, corrected, with examples and amounts adapted to the market, is localized content. An automated export of three hundred pages that no human has read is exactly the case the policy describes. The same reasoning applies to generated articles, which we covered in what Google really says about AI written content.

A keyword is not translated, it is researched

This is the costliest mistake and the least visible one. You take your best English page, translate the title word for word, and end up with a phrase nobody types in the target language.

A common pattern in online retail: the English query and the local query on the same subject do not share length, structure or level of precision. The French queries reaching this site are often full sentences, phrased the way you would ask a person. The equivalent English queries are shorter and more technical.

The consequence lands straight on your method. Every market deserves its own keyword research, with its own phrasings, as described in keyword research with buying intent. Slug, title and description get rewritten, not translated. Two twin pages can perfectly well target two phrases with no word in common.

What assistants see of your site, language by language

On the Google side, the eligibility rule is written down: "To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet". The documentation adds that no specific file or schema is required: "There's also no special schema.org structured data that you need to add".

Google also publishes the list of countries, territories and languages where AI Overviews are available. That list exists precisely because availability is not uniform: your German market and your Brazilian market do not necessarily live the same search experience at the same time.

On the OpenAI side, the crawler documentation separates three agents. OAI-SearchBot is "used to surface websites in search results in ChatGPT's search features", and opting out has a written consequence: those sites "will not be shown in ChatGPT search answers". GPTBot concerns training of the foundation models. ChatGPT-User covers visits triggered by a user action, and the documentation notes that robots.txt rules may not apply to it.

What is documented nowhere, and therefore has to be presented as an observation, is the preference assistants show for sources written in the language of the question. We notice it, we do not guarantee it. To measure what actually happens on your own site rather than assume it, the method is in tracking your AI visibility, and the comparative case in measuring brand share of voice in AI answers.

What no tag can buy

A correctly implemented hreflang does not lift you. It stops a German reader landing on your Spanish page. That is useful, it is not a ranking lever.

Google says as much about its AI features: "Just because a page meets all requirements, best practices, and complies with the policies, doesn't mean that Google will crawl, index, or serve its content". Technical compliance is a ticket in, not a seat.

What stays decisive is slower and less configurable: the depth of what you write, other sites talking about it, and time. On a young domain the first weeks read in impressions, not in clicks, and that is normal. We described what can legitimately be watched before the first clicks in signs SEO is working.

Where to start with one writer

Order matters more than completeness. Here is a sequence you can actually hold when nobody is dedicated to this.

  1. Pick a URL structure and stop changing it. The subdirectory is the sensible default.
  2. Open one second market only, the one visitors or orders already come from. Two markets held well beat five started.
  3. For every page, redo the keyword research in the target language before writing a line.
  4. Write the pair at the same time, and place both hreflang annotations right away. Placing them later means never placing them.
  5. Check reciprocity page by page. An annotation that goes out without a return does not count.
  6. Publish, then handle the follow up like any other release, which what to do after publishing a blog post covers step by step.
  7. Wait before concluding. A new page on a young domain takes weeks to exist in the index.

If you sell online, the logical next step is the store level, covered in AI search optimization for ecommerce, then the product page level.

FAQ

Should I translate the whole site at once? No. One complete, reviewed pair of pages beats fifty machine translated ones. The spam policy targets exactly that kind of mass production without added value.

Is a subdirectory worse than a country specific domain? Google presents both as valid, with different trade offs. The country domain gives clearer geotargeting but costs more and targets a single country. The subdirectory is simpler and benefits from the domain history.

Does hreflang improve my ranking? No. It helps serve the right version to the right user. Google even states that it does not use it to detect the language of a page.

What happens if only one of the two pages carries the annotation? Nothing. The documentation says that if two pages do not both point to each other, the tags are ignored.

How do I know if my pages appear in AI answers per market? There is no dedicated official report. You proceed by regular sampling, asking the same questions in each language and noting what gets cited. The full method is in our article on visibility tracking.

My local competitors are older, is it lost in advance? No, but time works against you in the short term. On a generic query held by established sites, targeting a more precise phrasing that matches the real question gives better odds than a head on fight.

Conclusion

Multilingual content for AI search does not need a magic tool. It needs one address per market, reciprocal annotations, keyword research redone in each language, and the acceptance that none of it shows up in week one.

Distrify exists so that this machinery runs without you holding it by hand: one campaign, written, published and adapted across your channels, aligned on the same keywords. The product is not open yet, and the engine behind it already publishes every day for real sites.

Join the waitlist

For the wider picture, see content distribution strategy across every channel, and for the citation side getting mentioned in ChatGPT as well as how to appear in Google AI Overviews.

Sources cited: Google documentation on multi-regional and multilingual sites, Google documentation on localized versions (hreflang), Google documentation on AI features in Search, Google spam policies, Google Search Help page on AI Overviews availability, OpenAI documentation on its crawlers.