The Language Asymmetry in AI Recommendation

I've been thinking about a problem in overseas GEO lately.

Plenty of non-English brands have solid products and solid content. But when it comes to AI recommendation, their presence is noticeably weaker.

I'm increasingly inclined to think this isn't because they're doing something wrong. The rules themselves aren't symmetrical.

Models like ChatGPT, Gemini, Claude, and Grok are trained primarily on English-language corpora. Their knowledge structures were built around the English internet first. Under that foundation, non-English brands are structurally less likely to enter the layer of information these models reach for most readily.

The reverse is also true. Doubao, Qwen, GLM, and DeepSeek come from different Chinese tech companies, all trained primarily on Chinese-language data. When the query is in Chinese, what these models reach for most readily is an entirely different network of entities.

What's worth paying attention to here: this looks like a technical issue, but over time it settles into something else — a brand visibility issue, a recommendation-rights issue, eventually a question of whose voice gets heard.

AI hasn't dissolved the filter bubble. It has taken the old information asymmetries and amplified them again, through a new mechanism.