Tesla Couldn't Bring Grok to China — So Doubao and DeepSeek Moved Into the Model Y's Dashboard

Tesla's newest Model Y sold in China doesn't run Grok. It can't — xAI's chatbot is blocked by the country's data localization rules, which require any AI system touching Chinese user data to run on domestic infrastructure. So Tesla did something almost no Western brand has had to consider: it put two of its fiercest AI competitors, ByteDance's Doubao and DeepSeek, directly into its own car. Both models now run side-by-side in the new Model Y's cabin system, hosted on ByteDance's Volcano Engine cloud, the same partnership Tesla struck last August to power in-car voice and interaction. Tesla has also filed "Tesla Intelligence" trademarks in China, suggesting it wants its own AI-branded layer sitting on top of infrastructure it doesn't fully control.

This is a strange position for a foreign automaker to be in, and it's not unique to Tesla. It points to a battleground that most international brands tracking their China GEO scores haven't priced in yet: the AI model that answers "which EV should I buy" in a chat window is often the exact same model now living inside the dashboard of the car being cross-shopped.

The dashboard has become an AI distribution channel

Doubao's rollout inside Chinese vehicles has been aggressive and brand-specific. BYD has deployed Doubao across all five of its sub-brands — Yangwang, Denza, Fangchengbao, Dynasty, and Ocean — inside its DiLink cockpit system. Mercedes-Benz integrated Doubao deeply into its China-market electric CLA and GLC, specifically to handle fuzzy, indirect instructions and dialect-accented speech, the kind of natural Chinese that rule-based voice assistants fail on. SAIC's Roewe M7 DMH was the first production car to ship with Doubao's "deep thinking" model, supporting understanding across 15 use-case categories, multi-turn conversation memory, and control over more than 300 in-car functions.

DeepSeek moved just as fast in the other direction. Since early 2025, more than 20 automakers — Geely, BYD, Chery, Dongfeng, Changan, and FAW-Volkswagen among them — have announced DeepSeek integrations. Dongfeng's Lantu Zhiyin was the industry's first mass-production car built around DeepSeek from the outset. Geely's Xingrui large-model system, now fused with DeepSeek-R1, can parse ambiguous driver intent and route it across roughly 2,000 distinct vehicle-control interfaces.

In other words: the two models that, according to hubGEO's tracking, now account for the large majority of China's AI assistant usage are no longer just where consumers go to research a car. They're increasingly what the car itself sounds like.

Why this matters more than another MAU chart

It would be easy to read this as a supplier story — automakers picking an LLM vendor the way they'd pick a chip supplier. That misses what's actually happening to brand voice. When a consumer sits in a Tesla Model Y and asks the in-car assistant a comparative question — "is this better than the BYD Han?" — the answer is being generated by the same Doubao or DeepSeek model that, in a separate context, is also the one recommending BYD to that same consumer's phone. The automaker doesn't fully own the reasoning happening inside its own product anymore. It has outsourced part of its own showroom pitch to a model built and tuned by a company (ByteDance) or lab (DeepSeek) with no obligation to favor it.

For a domestic brand like BYD, this is close to a non-issue — Doubao and DeepSeek are, if anything, more likely to reflect favorably on categories where Chinese brands already dominate discourse. For a foreign automaker, it's a new and under-discussed exposure: your own dashboard may be quietly reinforcing the same competitive narrative your GEO team is trying to fight in open chat.

The volume gap is widening, not narrowing

The urgency here is compounded by how lopsided China's AI assistant market has become. QuestMobile's Q1 2026 data puts Doubao's monthly active users at roughly 345 million, more than 40x Kimi's 8.3 million — a figure that has now declined for four consecutive quarters. Qwen sits around 170 million and DeepSeek around 130 million MAU. Whatever brand-facing behavior Doubao and DeepSeek exhibit, in a car or in a chat app, now reaches a scale that dwarfs every other Chinese assistant combined, and increasingly dwarfs the reach of a brand's own owned channels.

Meanwhile the EV market these dashboards sit in is reshuffling fast. Leapmotor's 2025 China sales roughly doubled to 597,000 units, overtaking Li Auto to become the top-selling emerging automaker. XPeng and Xiaomi both crossed 400,000 units. Li Auto slipped to 406,000. NIO grew steadily to 326,000. Every one of these brands is now also a distribution point for whichever LLM sits in its cockpit — which means the GEO conversation for auto brands in China can no longer stop at "how do we rank when someone asks DeepSeek to compare EVs." It has to extend to "what does DeepSeek say about us when it's the one running our own passenger's dashboard."

What foreign brands can actually test today

Most international marketing teams have no visibility into how the embedded assistant in their own China-market product talks about them, because nobody on the team has thought to ask. That's a fixable gap, and it doesn't require an engineering team:

1. Identify the model. Ask your China GTM lead, dealer network, or OEM licensing contact which LLM — Doubao, DeepSeek, or another — is licensed into the in-cabin or in-app assistant for your product line, and whether it's the same base model consumers can access independently in Doubao's or DeepSeek's own apps.
2. Prompt it like a customer would. Sit in the vehicle (or use the demo unit at a dealership) and ask comparative questions a real buyer would ask: "What's a good alternative to this car?" "Why should I choose [competitor] instead?" Note whether the assistant surfaces competitors unprompted, and how it frames your brand relative to them.
3. Compare against the open-chat version. Run the same prompts through the standalone Doubao and DeepSeek apps. If the in-car assistant's answers differ meaningfully from the open-chat version, that's a sign of custom fine-tuning or a system prompt layer your OEM partner controls — which means there's room to negotiate how your brand is represented inside your own product.
4. Track it over time. Model updates happen quarterly or faster in this market. A favorable answer today doesn't guarantee one after the next fine-tune. This needs to sit on the same recurring testing cadence as your standalone GEO tracking, not a one-time audit.

The WeChat and dealership layer compounds this

The in-car assistant is also rarely the only touchpoint. Many of the same OEMs distribute a WeChat mini-program or dealership kiosk experience built on the same underlying models, extending the reach of whatever answer pattern the cabin assistant produces into pre-purchase research and after-sales service chat. A buyer who test-drives a car, then goes home and asks DeepSeek the same question on WeChat, is really talking to a close cousin of the assistant that just answered them from the passenger seat. Brands that only monitor one of these surfaces are getting a partial and potentially misleading read on how they're actually being represented across the full AI-mediated buying journey.

Takeaway for Brand Marketers

If your brand sells a physical product with an embedded AI assistant in China — vehicles are the clearest case today, but this extends to appliances, robotics, and eventually retail hardware — your GEO scope now includes the model running inside your own product, not just the model running in a browser tab. Ask your China team, or your OEM partner, which LLM is embedded in-cabin or on-device, and whether that model has been tested for how it discusses your brand versus competitors when prompted naturally. If you're a foreign automaker using Doubao or DeepSeek under license, treat that in-car assistant as a fourth GEO surface alongside Doubao chat, DeepSeek chat, and WeChat search — because it is being trained on, and reasoning from, the same underlying model weights. Track your brand's score across all of them, and don't assume a licensing agreement buys you favorable treatment inside your own product.

Related: see how Doubao and DeepSeek's combined dominance is reshaping GEO budgets industry-wide on hubGEO's brand tracking.