EN
Back to the archive

The archive · AI & Models · Strategic decision · 2025–2026

Doubao wins China's AI race by spending less: 227M MAU as ads fall 60%

Doubao cut quarterly ad spend from ¥161M to ¥65M while MAU climbed 99.8M to 227M in 2025, and still held #1 in China's iOS AI downloads in April 2026.

Doubao (ByteDance)

The betThat a free assistant's data-model flywheel makes paid acquisition optional: cut ads, keep retention high, let model quality pull users faster than rivals can buy them.Scaling

What the business is

Doubao is ByteDance's consumer AI assistant for mainland China, the domestic counterpart to the overseas Dola and built by the same Flow organization. It answers questions, edits images and generates content for free, while the underlying Doubao model is also sold to enterprises through ByteDance's Volcano Engine cloud.

How it started

Doubao went out as ByteDance's free answer in China's consumer AI race and was already winning by early 2025: MAU hit 99.8 million in Q1 while that quarter's user-acquisition spend was ¥161M. The premise, per TMTPost, was a flywheel: more users produce more data, data improves the model, and a better model retains and attracts more users.

What happened

Across 2025 the flywheel turned. MAU climbed each quarter to 227 million by Q4, up more than 127% for the year, while quarterly ad spend fell to ¥65M by Q3 - growth moving opposite to marketing. TMTPost cites average 30-day retention of 44.5% from January to November 2025 against 32.1% for second-place Kimi. When the Spring Festival 2026 marketing war began, rivals outspent ByteDance massively: Qwen reportedly poured in ¥6B and held the App Store free-chart top spot for eight days, while Doubao spent ¥1.5-2B, joined the fight only on February 13 and still retook first place on February 17, holding it through February 26 and leading iOS downloads at 12.43 million for December 20-February 20 (Jiemian/Diandian data). DataEye's April 2026 report then showed the whole mainland AI-app market contracting 16.95% month over month while Doubao kept first place on iOS at 5.89 million downloads (22% share) and its share of ad materials fell to 2.0% - which DataEye read as proof that a brand moat had replaced paid traffic.

How it ended up

Still running as China's #1 AI app as of DataEye's April 2026 report; TMTPost describes the domestic flywheel as the playbook ByteDance then exported overseas through Dola.

Background

Doubao is ByteDance's consumer AI assistant for mainland China, built by the same Flow organization that later shipped the overseas app Dola. Its bet, per TMTPost, was a flywheel: a free assistant that keeps users generates more data, data improves the model, and a better model retains and attracts more users - so growth should not depend on buying traffic.

The numbers moved opposite to marketing convention. MAU rose from 99.8 million in Q1 2025 to 227 million in Q4, up more than 127% for the year, while quarterly user-acquisition spend fell from ¥161M to ¥65M. Average 30-day retention ran 44.5% from January to November 2025, against 32.1% for second-place Kimi, and TMTPost called Doubao the first Chinese AI-native app to pass 100M daily active users.

The Spring Festival 2026 marketing war tested the thesis. Qwen reportedly spent ¥6B and held the App Store free-chart top spot for eight days; Doubao spent an estimated ¥1.5-2B, joined late on February 13 and still retook first place on February 17, holding it through February 26 with 12.43 million iOS downloads over December 20-February 20 (Jiemian/Diandian data).

By April 2026 the mainland AI-app market was contracting 16.95% month over month, yet Doubao stayed first on iOS with 5.89 million downloads and a 22% share while its share of ad materials fell to 2.0%. DataEye read the combination as proof that a brand moat built on retention had replaced paid acquisition.

What has to be true

  • MAU rose from 99.8M in Q1 2025 to 227M in Q4 while quarterly acquisition spend fell from ¥161M to ¥65M - growth inversely tracking marketing.
  • Average 30-day retention of 44.5% (Jan-Nov 2025) versus 32.1% for runner-up Kimi meant every acquired user was worth more, so buying less did not cost share.
  • Outspent by Qwen's ¥6B Spring Festival blitz, Doubao spent ¥1.5-2B, joined late, and still retook the free-chart top spot and led downloads at 12.43M.
  • DataEye's April 2026 reading, #1 iOS with 5.89M downloads and 22% share while ad materials shrank to 2.0%, showed the flywheel holding as rivals' paid spikes faded.

What can be applied

When retention compounds, ads are optional: Doubao outgrew China's AI rivals while cutting acquisition spend, because retained users fed a better model - the flywheel did the buying.

Aftermath

Doubao remained #1 on mainland iOS as of DataEye's April 2026 report, and the retention model survived its hardest test: Qwen spent roughly ¥6B, three to four times Doubao's outlay, yet Doubao retook the App Store free-chart top on February 17, 2026 and led iOS downloads at 12.43M (Jiemian). The market shrank 16.95% month over month in April, so growth is not guaranteed; Doubao's ad-material share fell to 2.0%, an edge only while retention holds. TMTPost calls this flywheel the template ByteDance exported as Dola, which passed 200M downloads in Q1 2026; monetization is still open.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases