A shampoo brand launches in five cities with five different weather, five dialects, and five price sensitivities. A human team would shoot five films. A generative system writes five hundred in an afternoon and asks which one landed. The interesting question is no longer "can AI make an ad" but "how many versions of the truth do you need."
From single hero film to infinite variation
Traditional brand creative optimized for one perfect asset: the hero TVC, the flagship poster, the single voiceover. Chinese marketing clouds have quietly inverted that logic. The asset is now the template, and the system produces the variations — by region, by platform, by audience segment, by time of day.
Two platforms show the pattern most clearly: Baidu's (百度) AI-native marketing system and Alibaba's (阿里) Alimama (阿里妈妈) creative suite.
Baidu: the AI-native marketing stack
Baidu began injecting large-model (大模型) capability into its commercial products in 2023, narrowed its focus through 2024, and by 2025 described the work as scaled and validated. The system rests on three pillars it markets as "new search, new infrastructure, new system": a redesigned search experience, merchant agents (商家智能体), and an AIGC production tool called Qingduo (擎舵).
The loop is meant to run without a human touching every step. A small-business owner sets a budget and a goal in the simplified delivery console; the system calls Qingduo to generate batches of AIGC video; the material runs inside the new-search surface; and a merchant agent answers inquiries 24/7, filters leads, and books callbacks.
The financial signal is concrete. In its third-quarter 2025 earnings, Baidu reported AI-native marketing service revenue of 2.8 billion yuan (≈ US$389 million / HK$3.04 billion), up 262% year on year, and said the AI marketing system had reached more than 30 first-tier industries. Those are company-reported figures from the earnings release, not third-party estimates.
A documented case makes the mechanics visible. China Oriental Education, a vocational-training group, used Baidu's search-answer surface, Qingduo AIGC, and merchant agents across seven education brands. Baidu said the setup lifted daily impressions by 26% and click-through rate by 59%, while Qingduo produced hundreds of thousands of materials to cover long-tail search queries the company could never staff by hand.
Alimama: creative factories inside the Taobao machine
Alibaba's Alimama built the counterpart inside e-commerce. Its AIGC creative platform, anchored by a standalone tool called Wanxiang Yingzao (万相营造, "create with AI"), is designed for the merchant who has one product photo and needs a hundred placements.
The tool turns a single product image into video, swaps in models, generates copy, and adapts scenes — a garment flat-lay becomes a model wearing it in a lifestyle setting, without a photographer. Alimama says the platform leans on its self-developed Taobao Star (淘宝星辰) model and Taobao's consumption data to keep the output on-brand and commercially legible.
The scale figures are large and company-stated. At the 2025 China International Fair for Trade in Services, Alimama said its AIGC creative system had helped millions of merchants generate more than 100 million marketing assets and saved over 10 billion yuan (≈ US$1.39 billion / HK$10.85 billion) in production cost. During the 2025 Tmall Double 11 festival, Alimama reported that Wanxiang Yingzao alone produced 190 million-plus materials, saved merchants more than 3.5 billion yuan (≈ US$486 million / HK$3.8 billion), and raised creative-production efficiency by up to 150%. Its assistant, AI Xiao Wan (AI小万), was invoked over 10 million times and served more than 2 million merchants.
Why localized variation is the killer feature
The reason brands adopt this is not "cheaper video." It is local fit at volume. A national campaign in China spans dozens of regional markets, short-video platforms, and shopping festivals, each with its own tone and format. Hand-producing a tailored cut for every combination is economically impossible; generating it is nearly free at the margin.
So the work shifts from crafting one message to governing a system that crafts many. The human sets the guardrails — brand voice, banned claims, approved spokespeople — and the model fills the matrix.
The risks nobody puts in the launch deck
Bulk generation has a failure mode: a wrong claim multiplied ten thousand times. Both platforms pair generation with compliance filters, and Alimama's figures emphasize "pass-rate" for scaled commercial use, signaling that review, not output, is the bottleneck. There is also the question of creative sameness — when every merchant uses the same templates, differentiation collapses into a sea of competent, forgettable assets.
For now, the platforms compete on throughput and integration. The brand that wins is the one that treats the system as a creative partner with rules, not a vending machine.
Honest limitations
Baidu's 2.8 billion yuan revenue and 262% growth, and the 30-plus industries figure, are drawn from Baidu's Q3 2025 earnings disclosure and associated press coverage; they confirm revenue existence, not independent quality judgment of the ads produced. The China Oriental Education +26% impressions / +59% CTR result is a single vendor-documented case, not a controlled study, and outcomes vary by category. Alimama's 100 million-asset, 10 billion yuan saved, and 190 million-material figures are company-stated at CIFTIS and Double 11 retrospectives, presented without independent audit; "up to 150%" efficiency is a best-case ceiling, not an average. No authoritative source here measures brand-lift or long-term creative-quality impact, so those remain open questions.
What readers can do now
- If you run paid acquisition in China, pilot one product line through Baidu Qingduo or Alimama Wanxiang Yingzao and measure cost-per-qualified-lead against your current studio pipeline, not just cost-per-asset.
- If you brief creative agencies, write the compliance and brand-voice rules into the prompt library up front — govern the matrix, because a bad claim scales as fast as a good one.
- If you evaluate vendors, ask for pass-rate and review-queue data alongside output volume; the bottleneck is review and fit, not raw generation.