China's Parallel Answer Engines: Why DeepSeek, Doubao, and Baidu ERNIE Are a Separate Visibility Game

A market larger than most countries, arriving fast

Start with scale, because it is the one part of this story that is cleanly measured. China's official internet regulator, CNNIC, reported 602 million generative-AI users by the end of December 2025, a jump of 141.7 percent from a year earlier, against a total internet population of 1.125 billion at 80.1 percent penetration. National adoption of generative AI moved from roughly 17.6 percent to 42.8 percent of the population in a single year. Layered underneath that top-line figure, QuestMobile's tracking of standalone AI-native apps, meaning dedicated AI apps rather than AI features bolted onto an existing app, put monthly active users at 499 million by May 2026, up 85.4 percent year over year. Engagement inside that number is notably heavy: an average of about 93 uses and 183 minutes per user per month.

Put plainly, whatever you think you know about how people discover and evaluate brands using AI in your own market, there is now a population roughly the size of the United States and Indonesia combined running that same behavior on a completely different set of apps, under a completely different set of rules.

Who actually holds the audience, and how unsettled that ranking still is

Inside China's AI-native app market, ByteDance's Doubao led with 382 million monthly active users as of May 2026, Alibaba's Qwen followed at 167 million, and DeepSeek, the model that briefly dominated Western headlines, ranked third domestically at 130 million, per QuestMobile. Separately, the Wall Street Journal reported that Baidu's ERNIE Assistant crossed 200 million monthly active users in January 2026, a figure that traces to sourcing described as people familiar with the matter rather than a confirmed Baidu filing, and is treated here as emerging rather than established.

This ranking should not be read as settled. Different trackers measure different things: monthly active users in one report, weekly active users in another. The resulting figures for the same platforms have been found to vary by a factor of two or more depending on the source. Even the analysts closest to this market do not agree on the order.

Doubao leads at 382 million monthly users, Qwen sits second at 167 million, DeepSeek third at 130 million. Ask a different tracker and the order can change.

A wall built by regulation, not by accident

The separation between China's AI-answer stack and the Western one is not a side effect of language or preference. ChatGPT has never been officially accessible inside mainland China since its November 2022 launch, and Google Search has been blocked there since 2010. Any generative-AI service operating inside China has to register with, store user data inside, and comply with the content rules of the Cyberspace Administration of China under its 2023 Interim Measures for Generative AI Services.

That is a designed border, and it runs in both directions: it keeps Western engines out of China's answer layer just as effectively as it keeps China's dominant answer layer invisible to most Western brand-visibility tooling, which was built to crawl and monitor ChatGPT, Google's AI Overviews, and Perplexity, not Doubao or ERNIE.

Google has been blocked in China since 2010. ChatGPT has never officially launched there. This is a regulatory border, not a language gap.

What the moderation regime actually changes, and where the evidence stops

The clearest measured effect of China's content rules comes from a peer-reviewed study by Stanford and Princeton researchers, published in PNAS Nexus. Testing Chinese-origin models against non-Chinese models on politically sensitive prompts, the researchers found refusal rates as high as 60.23 percent for one Chinese model against 0 to 2.8 percent for non-Chinese models, along with a higher rate of complete factual inaccuracy on those same sensitive prompts, roughly 8 to 22 percent for China-origin models (DeepSeek came in near the top of that range) versus 6 to 10 percent for non-China models. The researchers frame this as consistent with regulatory compliance rather than a general capability gap, and the numbers are strong enough to support that reading.

What this study does not tell you is what happens on an ordinary commercial question: whether your product gets mentioned accurately, whether your brand name is spelled correctly, whether a competitor is favored. The measured censorship effect documented here sits specifically on politically sensitive topics. Extending it to mean that everyday commercial answers are also being reshaped is a plausible guess, not something this research established.

A different citation substrate, reported but not yet proven

Practitioner sources working in China's answer-engine-optimization space describe a citation hierarchy that looks nothing like the Western one. A guidance piece from the China-market agency KAWO names Zhihu, a long-form question-and-answer platform, Baidu Baike, China's dominant online encyclopedia, and WeChat Public Account articles as the sources Chinese generative engines lean on most heavily, rather than the broad open web that ChatGPT and Google's AI Overviews draw from.

That claim is worth taking seriously and worth tiering carefully: it comes from a single agency source, not from an independently audited crawl of what these engines actually cite, and it is carried here as contested rather than established. If it holds, it means the entire mechanism for becoming citable in China runs through a different set of platforms than the ones most Western content and PR programs are built around, and a strong Western web presence would not automatically transfer.

"Chinese AI" is not one market, and that complicates the picture further

It is tempting to treat China's AI apps as one contained domestic phenomenon. The data does not support that. Of roughly 22 Chinese-developed apps that appeared on a16z's ranking of the top 50 global consumer AI web products, only three, Quark, Doubao, and Kimi, primarily serve the domestic China market. The rest were built by Chinese teams but are used mostly outside China. That breakdown comes from a single industry tracker, not an official census, so it is worth reading as a directional picture rather than a settled count.

Even within that smaller group, the split is uneven. Kimi draws more than 75 percent of its traffic from inside China despite its global ranking, while DeepSeek, the model most associated with China in Western coverage, had its web usage fall more than 40 percent and its mobile usage fall 22 percent from its February 2025 peak by the time a16z's report was published in August 2025. And DeepSeek itself is proof that the domestic-versus-global line can blur fast: it reached number one on both the US Apple App Store and Google Play Store in late January 2025, the same period it was being framed almost entirely as a Chinese domestic story.

DeepSeek hit number one on the US App Store the same month it was being described almost entirely as a domestic Chinese product. The category is messier than the label suggests.

Where AI-assisted shopping actually stands in the data

On the buying side, the strongest available signal is what is missing rather than what is measured. Bain's China Shopper Report 2026, now in its fifteenth consecutive year and covering 27 FMCG categories with Worldpanel by Numerator, is arguably China's most established annual shopper-behavior tracker. It reports e-commerce reaching 38 percent of urban FMCG spending in 2025, accelerating at 6 percent growth. It does not contain a discrete measurement of AI-chatbot-assisted shopping at all.

A separate, single-source marketing claim reports that Xiaohongshu (RED) has become the primary discovery surface for Chinese women aged 18 to 35 in beauty, fashion, and food, with category search volume reportedly overtaking Baidu. That claim is not corroborated by an independent measurement source and is carried at contested tier here. Taken together, the read is that China's own best shopper-behavior research has not yet found AI-chatbot shopping significant enough to track as its own line item, even as usage of the underlying apps climbs into the hundreds of millions.

The central gap this study will not paper over

The seed idea behind this research was sharper than what the evidence supports: that a brand well-cited on ChatGPT or Google's AI answers could be functionally invisible in Chinese consideration sets. Every structural fact needed for that claim to be true is independently verified here: a walled-off audience of hundreds of millions, a separate regulator with real teeth, and, by practitioner report, a different citation substrate. What does not yet exist, in the sources reviewed for this study, is a transparent, independent census of brand-citation rates across DeepSeek, Doubao, ERNIE, Qwen, and Kimi, the kind of study that already exists for Google AI Overviews and ChatGPT citations in Western markets.

That means the core claim of this study is a well-grounded inference, not a measured fact, and it is precisely the kind of gap that gets closed only by someone actually going and measuring it.

The wall is a fact. The audience size is a fact. Whether your brand actually disappears from a Chinese consumer's chat window is something nobody has counted yet.

What this means if China is anywhere in your growth plan

If you sell into China, source from China, compete with China-based brands globally, or simply have customers who increasingly ask AI apps for recommendations before they ask a search engine, the practical takeaway is not to assume your existing AI-visibility work reaches this market, and not to assume it definitely does not. Both are guesses right now. What is provable is that this is a large, fast-growing, differently governed answer layer, reportedly built on different sources too, running in parallel to the one most Western visibility work targets.

The next step is not a bigger claim. It is mapping, specifically for your category and your market, where attention actually sits before deciding where to spend.