China's Workplace AI Agents Route Around Doubao — And That Rewrites B2B Brand Visibility in 2026

Tencent's WorkBuddy pulled 20.97 million visits in June 2026, more than the second- and third-placed products combined. It is now the largest AI-native desktop office agent in China. It is also, for brand marketers, the most consequential surface almost nobody is measuring — because WorkBuddy does not run on a single model. It routes queries to DeepSeek, GLM, Kimi, and MiniMax.

Doubao is not on that list.

That single fact breaks the assumption most international B2B brands have been operating on since the start of 2026: that winning Doubao — 345 million monthly active users, the largest consumer AI app in China — is the same thing as winning China's AI search. In the consumer layer, it mostly is. In the workplace layer, where vendor shortlists and procurement comparisons actually get made, the model mix is close to the inverse.

The 14-Day Consolidation That Created a New Layer

The workplace agent surface did not exist as a coherent category a year ago. It was assembled in about two weeks.

On July 30, 2026, ByteDance folded the Feishu product team into Doubao, with Feishu head Xie Xin reporting to Doubao's lead; Feishu's sales organization moved to Volcano Engine. Four days later, on August 3, Alibaba opened QwenWork to public beta, consolidating three previously separate products — QoderWork (coding workflows), MuleRun (task automation), and Wukong (knowledge management) — into one workplace agent platform. Tencent, already in the lead, kept pushing WorkBuddy across PC, WeChat mini-program, and mobile.

Three of China's largest platform companies restructured their office software around agents inside a fortnight. That is not product iteration; that is a land grab for a new query surface.

The volume backs it. Across 17 mainstream desktop AI-native office agent platforms, combined monthly visits passed 60 million in June 2026. WorkBuddy took roughly a third of that on its own, growing from 8.85 million visits at its March launch to 20.97 million in June — a 137% increase in three months. ByteDance's domestic TRAE IDE placed second at 12.79 million; Alibaba's QoderWork third at 7.88 million.

Sixty million monthly visits is small next to Doubao's 345 million MAU. Judged on raw reach, a brand team would deprioritize it immediately. That would be a mistake, and the reason is what those queries are.

Consumer Queries Ask "What Should I Buy." Workplace Queries Ask "Who Should We Buy From."

A consumer asking Doubao for a sunscreen recommendation produces one low-value impression. An operations manager asking a workplace agent to "compare three ERP vendors for a 200-person manufacturer and summarize implementation timelines" produces a document that circulates internally and shapes a six-figure purchase.

The workplace agent layer concentrates exactly the queries that matter most to B2B brands and to the B2B side of consumer companies: vendor comparison, supplier verification, pricing benchmarks, compliance checks, RFP drafting. IDC projects 5 million enterprise AI agents deployed across China in 2026, and the broader Chinese AI agent market hit RMB 80.4 billion in 2025, up 123.2% year over year, with forecasts reaching RMB 696.8 billion by 2030.

Put differently: the consumer layer generates most of the impressions. The workplace layer generates most of the contract value per impression.

The Routing Layer Is the Part Brands Keep Missing

Here is the structural point that changes GEO strategy rather than just adding a checkbox to it.

WorkBuddy is not a model. It is an orchestration layer that dispatches queries to DeepSeek, GLM, Kimi, and MiniMax and earns revenue on each call. Your brand's visibility inside WorkBuddy is therefore derived — it is a weighted function of how those four models describe you, mediated by routing logic you cannot see and that changes without announcement.

Three consequences follow, and each one contradicts a common 2026 playbook.

One: a Doubao-first China GEO budget is misallocated for B2B. Doubao's dominance is a consumer-surface fact built on the Douyin content flywheel. If your buyer is a procurement lead using WorkBuddy, your Doubao score is not in the chain of causation at all. The models that decide your visibility there are DeepSeek and GLM — models where citation behavior leans toward technical documentation, structured specifications, and verifiable third-party sources rather than social content.

Two: your weakest model becomes your ceiling, not your average. In a single-model surface, a low score on one platform costs you that platform's traffic. In a routed surface, a low score on one of four backends means a randomly assigned share of workplace queries returns an incomplete or absent version of your brand — and neither you nor the user can tell which backend answered. Averaging your scores across models hides this. A brand at 82 on Kimi and 31 on GLM does not behave like a 56; it behaves like a coin flip.

Three: routing weights are an unpriced volatility risk. A router can rebalance toward a cheaper or faster backend overnight for reasons that have nothing to do with your category. Brands that concentrated their optimization on one strong model will see workplace visibility move without any change in their own content, their competitors' content, or the underlying model.

Why GLM and MiniMax Deserve Attention They Are Not Getting

Most China GEO tracking — including the six-model panel that powers our own brand scores — is built around consumer-facing assistants: Doubao, Kimi, DeepSeek, Qwen, Wenxin, Hunyuan. That panel is correct for consumer discovery and incomplete for workplace discovery, because GLM and MiniMax appear in the routing pool while carrying almost no independent consumer mindshare.

This is the same measurement gap that showed up when China's agentic checkout surfaces began selecting products without ever rendering a recommendation list. The pattern repeats: a new selection layer forms, it consumes model outputs rather than producing them, and the industry keeps measuring the models instead of the layer.

There is also a jurisdictional wrinkle worth flagging for international brands. QwenWork's enterprise deployment has drawn scrutiny over Chinese state-law exposure inside corporate workflows. Whatever position your legal team takes, the practical marketing consequence is that multinational buyers in China may end up split across different agent platforms by policy rather than preference — which means your visibility has to hold across more than one routing pool, not just the market leader.

Takeaway for Brand Marketers

If you sell to businesses in China, or sell to businesses that sell to Chinese consumers, four moves are worth making this quarter.

Separate your consumer and workplace GEO scorecards. Do not report a single "China AI visibility" number. Report a consumer score weighted toward Doubao and Qwen, and a workplace score weighted toward the routing pool — DeepSeek, GLM, Kimi, MiniMax. They will diverge, and the divergence is the actionable part.

Test your floor, not your mean. Run your top 20 category prompts against each routed backend individually and record the worst result. In a routed surface that worst result is a live outcome for a real share of queries, not a statistical tail.

Write for the citation behavior of technical models. DeepSeek and GLM reward structured specification data, published documentation, verifiable pricing and compliance detail, and named third-party sources. They do not reward the social-proof density that lifts you inside Doubao. If your China content operation is built entirely on Xiaohongshu and Douyin, you have optimized for the surface your B2B buyer is not using.

Re-audit quarterly, and treat routing shifts as a monitored risk. Assume the routing mix will change without notice. The defense is not predicting it; it is holding an acceptable score across all four backends so that any rebalancing is survivable.

The consumer AI search race in China looks largely settled — Doubao won reach, and the remaining fight is over monetization. The workplace layer is at the stage the consumer layer was in early 2025: fast-consolidating, under-measured, and cheap to enter. Sixty million monthly visits is not where the volume is. It is where the contracts are.

Related: see how the same measurement problem plays out in China's agentic checkout layer, or check your current cross-model position on the hubGEO brand index.

Sources: Sina Finance and NetEase Technology reporting on the July 30, 2026 Feishu–Doubao merger; C114 and Sina Tech coverage of the June 2026 desktop AI-native office agent rankings (17 platforms, WorkBuddy 20.97M visits); ChinaTechInsights on WorkBuddy's routing architecture and growth from 8.85M March visits; IDC China enterprise AI agent deployment forecast for 2026; Alibaba QwenWork public beta announcement, August 3, 2026.