Whatever Language You Ask AI In, That Is the Language It Cites — We Ran the Test

A Question We Didn't Want to Answer by Feel

You wrote a strong article in Chinese. Solid content, thorough data. Now someone who speaks English asks AI about the very same topic. Will your site appear in AI's answer?

We didn't want to answer that question by feel, so we ran an experiment.

How We Ran the Experiment

We built our own citation probe. The method is direct: take the same set of questions, ask AI in different languages, and log which sites appear in the citation list at the end of each answer. "Cited" means your URL appears on that list; if it doesn't appear, you weren't cited. No gray area.

We used two platforms: Perplexity sonar and Gemini grounding. We picked these two because both list their sources plainly, so every entry can be checked line by line. This probe run took place on 2026-08-07.

The Result: Each Language Keeps to Its Own

The result was almost uncomfortably clean.

Japanese questions drew citations from Japanese sites: clabel.co.jp, note.com, prtimes.jp. English questions drew citations from English media: LinkedIn, Reuters, JapanToday, tradingeconomics, cnbc, cambridge.org.

And our own site was cited 0 times. Not ranked low — never in the room at all.

Why This Happens

Think of AI as a waiter taking orders. When the customer speaks English, the waiter only opens the English menu; when the customer speaks Japanese, only the Japanese menu. Your restaurant printed its menu in Chinese only — however good the food is, English-speaking customers will never know what you serve.

The waiter doesn't dislike you. It simply follows the customer's language. That is procedure, not a grudge.

Worth noting: this doesn't look like a comprehension problem. AI can clearly read Chinese, and it can translate. But the behavior pattern observable from outside is stable: whatever language the answer is in, the cited sources are in that language. How the platforms work internally, we cannot see, and we won't guess. For creators, the cause hardly matters — what matters is that the pattern shows up consistently: if you don't have a version in that language, you are not on the list.

AI picks sources in the language of the question. Chinese content is nearly invisible in English answers.

Being Crawled Is Not Being Cited

Some will say: AI crawlers visit my site every day — surely I'm not short on exposure?

Being crawled and being cited are two different things. In our earlier piece "AI Crawler Volume Is Buzz, Not Revenue" we tallied five months of accounts: AI crawlers hauled our content away roughly 30 million times, while humans who came back via an AI referral numbered only 476 (a visible lower bound). Source: https://km.idaeo.ai/ai/ai-crawler-shop-ledger

Crawling is moving inventory; citing is serving the dish. Inventory leaving the warehouse doesn't mean your dish reaches the table — especially when the customer's language doesn't match your menu.

Another reference point: in an earlier piece we measured the language structure of our own traffic — English 74.7%, Japanese 22.0%, Chinese 3.3%. Source: https://km.idaeo.ai/ai/visibility-lesson-1

Most of the world's questions to us do not happen in Chinese. Content in Chinese only means most of the doors are closed from the start.

Our Approach: One Text, Four Languages

So the conclusion is not "write better Chinese." It is "the same article needs to exist in other languages too."

This site's approach is one text in four languages: zh-Hant on the bare path as canonical; zh-Hans, en, and ja as ?lang= variants; hreflang cross-references in the sitemap.

In menu terms: one menu, printed in every language your customers can read, and the back cover of every copy notes "this restaurant also carries these language editions." Canonical states which copy is the master; hreflang lets the versions recognize each other as one family. Without that cross-referencing, the search side may treat them as unrelated pages, and the effort of keeping one text is wasted.

The point is "one text": not separate articles written per language, but the same content in different language versions. That way, whichever language AI searches in, it finds the same thing — and knows they are the same thing.

A multilingual version is not extra credit; it is the ticket. Without a version in that language, you don't even qualify to be picked.

Limitations, Stated Up Front

The sample in this experiment is small. Please read it at that scale:

• Only two platforms: Perplexity sonar and Gemini grounding. Other AI platforms were not tested.
• Few questions. With a different question set, we cannot rule out different results.
• The site is n=1: we only tested our own site. Your site and your topics may behave differently.
• Our own observation instrument only came online at 2026-08-10 07:52 UTC. The sample is still young.

The only conclusion we are willing to state is directional: the language of the question determines the language of the cited sources. Treat the numbers as one observation, not a law.

What You Can Do

• Pick your most important content and check whether it has a version in your target readers' language. No need to do the whole site at once — start with the topics you are asked about most.
• If you already have multilingual versions, check that canonical and hreflang cross-reference each other, so the language versions aren't treated as unrelated pages.
• Ask AI about your topic in the target language and look at the citation list with your own eyes. If you are not there, the ticket isn't bought yet.
• After buying the ticket, ask again after a while. Citation is not a one-shot verdict; it is worth watching continuously.