The DeepSeek V4 Paradox: A Stronger Model, a Shrinking Audience — and What It Means for Brand Visibility in 2026

On Monday, July 20, 2026, DeepSeek pushed its V4 series to general availability — V4-Pro (1.6 trillion total parameters, 49 billion active) and V4-Flash (284 billion total, 13 billion active), both built on the company's new sparse-attention design. Days earlier, OpenCompass's mid-July leaderboard had placed DeepSeek-V4-Pro among the strongest domestic large language models, alongside Doubao-Seed-2.0-Pro, Kimi-K2.6, and Qwen3.6-Max-Preview.

And yet, in the same month, DeepSeek's consumer app told a very different story: 129.8 million monthly active users in June 2026, down 20.3% year-over-year, while ByteDance's Doubao climbed to 382.3 million. DeepSeek is now shipping its most capable model to its smallest audience in over a year.

For brand marketers optimizing for AI search in China, this is the DeepSeek V4 paradox — and it forces a question most GEO playbooks have been dodging: when a model gets smarter but reaches fewer people, does it deserve more of your budget, or less?

Capability and reach have decoupled

For most of the AI era, capability and reach moved together. The best model won the most users. That assumption is now broken in the Chinese market.

DeepSeek remains the most globally recognized Chinese AI brand — it reset the industry's cost assumptions in early 2025 and still tops technical benchmarks. But recognition and daily consumer usage are different things. QuestMobile's Q1 2026 data already showed the divergence forming: Doubao at roughly 345 million MAU, Qwen at 166 million, and DeepSeek at 127 million. By June, Doubao had extended its lead to 382 million while DeepSeek slipped further.

The reasons are structural, not qualitative. Doubao is fused into ByteDance's Douyin flywheel, giving it a constant funnel of mass-consumer traffic. Qwen sits inside Alibaba's commercial machine — Taobao, Tmall, 1688, Alibaba Cloud. DeepSeek, by contrast, has no comparable consumer distribution engine; it is a model, not an ecosystem. When its user base skews toward researchers, engineers, and enterprise buyers, raw MAU understates its influence over high-value decisions while a capability jump like V4 does little to widen the top of the consumer funnel.

It is worth being precise about the magnitude here. Doubao is not merely ahead — it is roughly three times DeepSeek's app audience and more than double second-place Qwen's Q1 figure. QuestMobile has described the market entering its "second half," where growth increasingly comes from silver-haired and lower-tier-city users rather than the tech-forward early adopters who first flocked to DeepSeek. That demographic shift matters: the users still joining the category are precisely the mass-consumer segment Doubao captures by default, not the specialist audience that gave DeepSeek its early edge.

The takeaway for brand teams is uncomfortable: you cannot read DeepSeek's importance off a MAU chart. A model can be shrinking in headcount while remaining decisive for exactly the audience your B2B or premium brand cares about. The mistake is to treat all monthly actives as interchangeable units of opportunity. A million DeepSeek users evaluating enterprise software are worth far more to a B2B vendor than a million casual Doubao users asking for dinner recipes — and no aggregate usage chart will ever surface that difference for you.

Why a smarter model changes what gets cited

A generational model upgrade is not cosmetic for brand visibility. It changes retrieval behavior, reasoning depth, and — critically — which sources a model trusts when it composes an answer.

Three shifts tend to follow a capability jump like V4:

Longer, more structured answers. More capable models produce more thorough responses, which can mean citing more sources — but the attention economics stay brutal. Independent analyses of Chinese AI answers in 2026 found that a typical response names only two to three brands, and the first-mentioned brand captures more than 70% of user attention. A smarter model that reasons more deeply does not rescue the fourth-place brand; it just gets more confident about the top two.

Better source discrimination. Sparse-attention architectures like V4's are designed to weigh evidence more selectively. In practice, that rewards brands with clean, citation-ready evidence — structured specification pages, third-party reviews, expert Q&A — and further penalizes brands whose only footprint is a marketing homepage the model can retrieve but won't recommend.

Sharper engine personality. DeepSeek's audience and training diet already skew technical. It cites heavily from Zhihu (China's expert-answer platform), professional documentation, and Simplified-Chinese technical content. A V4 upgrade tends to intensify these preferences rather than blur them. If your category lives on Xiaohongshu lifestyle notes but not on Zhihu expert threads, a stronger DeepSeek may actually widen the gap between where you have content and where this engine looks.

In other words, V4 raises the bar for what counts as a citable brand. The floor for "good enough" evidence just moved up.

The budget question: reach-weighted, not model-weighted

The instinct after a big launch is to chase the shiny new model. The V4 paradox argues for the opposite discipline: allocate GEO effort by audience value, not by benchmark score or launch buzz.

A useful way to frame it is a simple reach-versus-fit matrix for the three leading engines:

Read the table as a spending guide, not a ranking. A consumer beauty brand should not lose sleep over DeepSeek's V4 improvements — its buyers live in Doubao's world. But an enterprise infrastructure vendor whose buyers are engineers should arguably invest more in DeepSeek despite its falling MAU, precisely because those 130 million users are disproportionately the decision-makers it needs, and because V4's stronger reasoning makes DeepSeek's recommendation harder to dislodge once earned.

This reframes the decline entirely. DeepSeek losing 20% of its users is not automatically bad news for you; it depends on which 20% left and whether the ones who stayed are your buyers.

What actually moves the needle inside a V4-class model

Regardless of engine, the evidence signals a capable model rewards are consistent — and they are content problems, not ad-spend problems.

Build a Chinese entity layer so the model can resolve who you are without ambiguity. Publish answer-first category pages that lead with the direct answer a user would ask for, not brand narrative. Create citation-ready evidence — specs, comparisons, and third-party validation the model can lift verbatim. And seed the sources each engine actually reads: Zhihu and technical documentation for DeepSeek, Tmall and product data for Qwen, Douyin and lifestyle content for Doubao. A brand that is retrieved but never recommended almost always has a source-quality or answer-format problem, not a visibility problem.

The through-line is that V4 does not reward louder marketing. It rewards better-structured proof. This is also why paid media and GEO are not substitutes. A brand can dominate Douyin ad placements and still be invisible inside Doubao's answers if it has never produced the structured, comparison-grade content the model can cite. The evidence layer is earned through published proof, and a stronger model simply raises the standard for what qualifies. Brands that mistake ad reach for AI visibility will keep seeing competitors named in answers they thought they had already paid to win.

Takeaway for Brand Marketers

The headline — DeepSeek ships V4, DeepSeek loses users — is a trap if you read only one half of it. The real signal is that capability and reach have split apart in Chinese AI search, and your GEO budget should follow audience value rather than either benchmark scores or MAU alone.

Three moves for the second half of 2026: First, stop treating MAU as a proxy for importance — score each engine by how much of your buyer population it reaches, then weight spend accordingly. Second, assume a V4-class upgrade raises the evidence bar; audit whether your brand shows up as citable proof (structured pages, third-party reviews, expert threads) or merely as a homepage. Third, respect engine personality — DeepSeek's technical, Zhihu-heavy diet means a deep-tech or enterprise brand can win outsized influence there even as the app shrinks, while a consumer brand's energy belongs in Doubao.

The brands that win in China's fragmented AI search are not the ones chasing every model launch. They are the ones who know exactly which engine their buyer trusts — and have built the evidence to be the two-to-three brands that engine names.

Related: see how brands score across China's six major AI models on the hubGEO brand tracker.