Doubao, DeepSeek and ChatGPT Don't Read the Same Internet

Before doing AI search optimisation (Generative Engine Optimization), one premise has to be clear: China's Doubao, DeepSeek and ERNIE Bot, and the overseas ChatGPT, Gemini and Perplexity, do not read the same internet. Their retrieval pools differ and their source preferences differ — run one playbook against both and one side is bound to miss.

This isn't a ranking of Chinese versus overseas. More precisely, they are two parallel tracks serving two groups of customers that only partly overlap. An export factory meets both groups, which is why both tracks need building.

Chinese AI reads a Chinese pool; overseas AI reads a mainly English web

This is the most fundamental difference, and the most easily overlooked.

China's major AI engines — Doubao, DeepSeek, ERNIE Bot and Qwen — retrieve mainly from a Chinese-language content pool. Your website, your WeChat Official Account, Zhihu, Baijiahao, trade media: if it's written in Chinese, it sits in that pool.

The major overseas engines — ChatGPT, Gemini and Perplexity — read a predominantly English open web. When an overseas buyer asks in English, "who makes small-batch hardware parts with CE certification", the retrieval happens inside the English pool.

What does that mean? A factory with only Chinese material effectively doesn't exist in English-language retrieval. Not because the product is poor, but because it sits outside the pool.

One easily missed detail: Google's official product page states that AI Overviews now covers more than 120 countries and regions and supports 11 languages. So this isn't a feature only US users can access — buyers in Germany, Brazil and Japan can now ask AI directly in their own language.

One technical switch many companies have closed on themselves

This deserves its own section: checking it takes a minute, and it affects everything downstream.

Google's official crawler documentation is explicit: Google-Extended is a control token in robots.txt, not a crawler. It has no separate HTTP identity; crawling is still done by Google's existing crawlers. The token controls exactly one thing — whether content already crawled may be used for Gemini training and grounding. Google also states clearly that it does not affect a page's inclusion in Google Search, and is not used as a ranking signal.

The key word is grounding. Grounding means feeding content from the search index to the model to improve the truthfulness and relevance of answers — which is precisely the step where your content gets cited into an AI answer.

Here's where it goes wrong: through 2023 and 2024, many sites bulk-disallowed Google-Extended in robots.txt to "stop content being used to train AI". Many assumed they were only refusing training — in fact they gave up eligibility for real-time grounding as well. And it happens silently: no notification, you simply fade out of AI answers over time.

So whether or not you do any optimisation, the first step is to check what that line in your robots.txt currently says. It takes a minute.

Two source logics — one playbook won't fit both

Gemini is crawled by Googlebot and its retrieval layer is the Google index. For it, English content on your own site plus assets inside the Google ecosystem (YouTube and the like) carry more weight.

On the ChatGPT side, third-party content and community discussion weigh more heavily. Applying the same "flood the communities" playbook to Gemini is likely to miss — the two source structures are not the same thing.

One study is worth consulting: research from the University of Toronto (arXiv:2509.08919, not peer-reviewed) found that AI engines prefer to cite third-party content, with a preference between 53% and 95%, and that a brand's own content loses out in that comparison.

Scope note: this is a preprint, not peer-reviewed; "53% to 95%" is a range, not a precise figure. But the direction it points in is clear — your own website isn't enough; someone else has to be mentioning you.

Why this article gives no exact source-share numbers per platform

You may have seen figures elsewhere such as "72% of Gemini's cited sources are company websites" or "Reddit accounts for 46.7% on Perplexity". I've checked this type of data, and I'd advise against quoting it directly.

The reason is simple: almost all of it comes from third-party analytics firms or tool vendors, with inconsistent methodologies. For the same "share of brand websites", I've seen both a 14% version and a 72% version — the gap comes from the counting method (counting cited domain types versus a distribution of source types) and from what is being measured (AI Overviews and the Gemini app are, after all, two different products).

So this section gives direction, not absolute values: Chinese AI leans on the Chinese content pool, overseas AI leans on English content and the Google ecosystem; both value your own website, but both trust more what third parties say about you.

Three questions about Chinese versus overseas AI

Can we run the China track only?

It depends where your customers are. For a purely domestic business, the China track is largely enough. For exporters and companies going global, overseas customers are the main body — they ask in English, and retrieval happens in the other pool. The two tracks are a division of labour, not a substitute for each other.

If we translate our website into English, is that enough?

That's only a start. Translation solves "can be read", but citation also depends on credibility — whether third parties mention you in English, and whether your sources agree with each other. Both have to be done.

How soon will the overseas track show results?

No timeline and no results are promised. Both vendors' official positions are the same: neither guarantees content will be crawled, indexed or shown. What can be said clearly is the pattern: English content has to be indexed first, and after that it needs to accumulate at the source level — the nature of it is saving up, not buying ads.

In closing

To sum up: Chinese and overseas AI are two retrieval environments, not two languages of the same ranking. For an exporting company, if your content exists on only one side, customers on the other side will never see you.

And the fix isn't complicated: English content, third-party sources, site-wide consistency, plus a check of that switch in robots.txt. The hard part isn't the technique — it's whether you're willing to start.

Have you searched for your own product in AI, in English? Did you appear in the answer? Try it once and see which companies the answer lists.

Yihang Network, based in Shenzhen, previously worked in cross-border B2B and Alibaba.com-related services, and expanded into GEO from 2024. Having worked on going-global business, one thing is clear to us: the retrieval environments at home and abroad were never the same thing.

Data and scope note: The Google rules and product coverage information cited here come from Google's official product pages and crawler documentation; they are public information, not a promise of results. The University of Toronto study is a preprint (arXiv:2509.08919), not peer-reviewed, and its range does not represent a precise proportion. The statement about how Chinese content performs in English-language queries is an inference about mechanism, not an official conclusion. The methods described raise the probability of being retrieved and cited; they do not guarantee a company will appear in AI answers, and no timeline or business result is promised.

References

1. Google official product page, AI Overviews (more than 120 countries and regions, 11 languages)
2. Google official Help Centre, list of countries and languages supported by AI Mode
3. Google official crawler documentation, "List of Google's common crawlers" (definition and scope of Google-Extended)
4. Google Search Central, "AI features and your website"
5. University of Toronto, arXiv:2509.08919 (preprint, not peer-reviewed)