Doubao starts charging 12% commission on AI-booked hotels
China's AI assistant race just found its first real business model: ByteDance is pricing AI-sourced hotel bookings at a premium over its own Douyin traffic, while Zhipu's coding agent crosses the million-user mark — two signs the country's AI boom is moving from user acquisition to monetization.
ByteDance's Doubao assistant now takes a roughly 12 percent commission on hotel orders that originate in its chat and close on Douyin's local-services platform — the first time an AI assistant's referral traffic has been priced as its own channel. The rate, effective August 10, splits into an 11.4 percent software service fee plus a 0.6 percent payment processing fee, up from the 8 percent Doubao-sourced bookings previously paid when they were lumped in with Douyin's natural traffic. Doubao PR chief Liu Xing responded on Weibo that the charge is not an advertising fee — merchants cannot pay to influence recommendations or ranking, and they pay only when a booking actually closes, a "pay-per-conversion" model he contrasts with traditional travel platforms' commission-plus-ads structure. The move matters because Doubao is China's biggest AI app at roughly 382 million monthly active users (QuestMobile), and it makes the assistant the first AI entry point in the country with an explicit transaction toll — a dry run for turning AI "intent" into a tollbooth. The risk is the trust question that hangs over every AI recommendation: once an assistant's answers are a revenue channel, how neutral are they? At 12 percent, the rate sits between Douyin's 8 percent and traditional online travel agencies' 13–15 percent, which suggests ByteDance is pricing its AI channel as premium traffic while it still can. Doubao has spent the summer testing paid tiers, and this commission is the second leg of that monetization push — a signal that the era of free AI concierge service is ending.
Zhipu's coding agent ZCode crossed 1 million users and shipped four new features aimed at autonomous, multi-step work. The upgrade adds Goal, Subagents, Remote Control and idle-time task execution, pushing the tool from conversational code help toward handing off complex jobs that span files and stages. Zhipu's internal Z.ai Code Bench shows GLM-5.2 paired with ZCode passing tasks 2.39 percent more often than when paired with Claude Code, with the edge concentrated in cross-file work requiring final acceptance. The company also says cache optimization raises effective quota for its GLM Coding Plan subscribers by about 30 percent — roughly doubling usable capacity through an August 31 promo — and reset weekly limits for all users to mark the milestone. It's a modest but telling data point in the coding-agent arms race: Chinese labs are now competing on the harness around the model, not just the model itself.
What to watch: whether other Chinese AI assistants — Alibaba's Qwen assistant reportedly began charging for some tiers the same day — copy Doubao's channel-fee playbook, and whether hotel merchants actually see net-new bookings from AI referrals.
Would you trust a hotel pick from an assistant that takes a cut of the booking? Tell us in the comments.
Sources: 36kr · Securities Times (证券时报) · STAR Market Daily via CLS (科创板日报) · NetEase (网易) · AI TOP100 · Sina (新浪) · Beijing News (新京报) · Sina Tech (新浪科技) · Guandian (观点网)
