Baidu's Ad Revenue Fell 19%: What It Means for Brand Visibility in Chinese AI Answers in 2026

Baidu's online marketing revenue dropped 19% year over year to RMB 13.1 billion in Q2 2026, and management told investors it expects the advertising business to "remain under pressure in the second half." For international brands, the number matters less as a Baidu story than as a signal about where Chinese consumers now get answers. Money that once bought a ranked link is being competed for by a single generated answer, and our own brand data shows that the answer engines do not treat foreign brands equally.

The numbers behind the decline

Baidu reported Q2 2026 total revenue of RMB 31.3 billion. AI-related business accounted for 50% of general business revenue for a second straight quarter, led by AI cloud infrastructure at RMB 7.3 billion (up 50%). The legacy engine, though, is shrinking: online marketing is down 19%, according to the earnings call transcript, and Chinese coverage describes it as the fourth consecutive quarter of double-digit decline.

The user side tells the same story. A Tencent News analysis put Baidu App monthly active users at 644 million, down from 735 million a year earlier, a loss of roughly 91 million users. The same piece estimates Baidu's share of Chinese search fell from about 87% in 2021 to around 51% by mid-2025. Meanwhile QuestMobile's Q1 2026 data shows the native AI apps at 446 million combined MAU, led by Doubao (345 million), Qwen (166 million) and DeepSeek (127 million). Later QuestMobile-based figures cited in September industry guides put Doubao at 382 million MAU.

Two details from the earnings call are worth reading closely. First, Baidu says its priority in AI search is "getting the products and the user experience right" rather than monetizing immediately, and the company has so far delayed ad placement inside AI answers. Second, its AI-native marketing services line, the business that sells brands placement in AI-driven products, brought in RMB 2.6 billion. Baidu is telling advertisers the future is in answers, while not yet showing them a mature way to buy those answers.

Why "search-click-read" is collapsing for brands

Traditional paid search sold position on a page of links. A generated answer compresses ten links into one paragraph that names two or three brands. When the reader never clicks, there is no results page to sponsor, and no bid to win. That is the mechanical reason ad pricing and reach are under pressure across the Chinese market.

For a brand marketer the implication is uncomfortable: the shortlist is now assembled before the customer visits any brand-controlled surface. If your brand is not in the shortlist, no bid can put it there. Baidu's reported monetization delay makes this window unusually clear. Organic presence in AI answers is currently priced at zero, but it also cannot be bought, so it has to be earned through content and entity signals.

What our brand scores show about the Baidu ecosystem

Baidu's answer engine is ERNIE (Wenxin), and it is the model where international brands look weakest in our tracking. Across the 39 brands in the hubGEO database, average scores by model are:

Wenxin posts the lowest score of the six models for 27 of the 39 brands, about 69% of the set. The average gap to DeepSeek is 5.9 points and to Doubao 4.2 points. That is meaningful because Wenxin is the engine embedded in the largest legacy search audience in the country.

The gap is not uniform. It widens sharply by category:

Individual brands show the effect more starkly. Tesla scores 58 on Wenxin against 72 on Doubao, a 14-point gap, and our highest cross-model spread for the brand is 20 points. Porsche scores 50 on Wenxin, 18 points below its best model. Lululemon and Bose both sit at 30 on Wenxin. MAC Cosmetics reaches only 38. These are brands with strong global recognition, which suggests the shortfall is about Baidu-specific signals rather than general awareness.

Why the signals differ: Baidu rewards its own ecosystem

Third-party observation studies of citation patterns, which we treat as directional rather than definitive, point in one consistent direction. A July 2026 study reported by industry guides found roughly 55% of ERNIE's cited sources come from Baidu properties, with Baijiahao around 35% and Baidu Baike around 20%. By comparison, the same study attributed about 60% of Doubao's citations to ByteDance properties (Toutiao 35%, Douyin 25%), while DeepSeek leaned on CSDN (about 20%) and Zhihu (about 15%), and Kimi on Zhihu (about 18%). The same source claims that in ERNIE a definition placed in the first paragraph has better than a 70% chance of direct citation, a claim we could not independently verify and that should be tested per brand.

Our hubGEO citation-source table, built on separate third-party observations from 2025, shows the same structural pattern for Doubao: Douyin (20%) and Toutiao (17%) top its list. The consistent finding across sources is that each Chinese answer engine has a home ecosystem it trusts, and a foreign brand that publishes only on its own website or on Western channels is invisible to all of them.

For Baidu that means the Baike entry and Baijiahao presence are not optional extras. A brand with a thin or outdated Baidu Baike page, no verified Baijiahao account and no structured Mandarin definition of what it is and who it serves will underperform on Wenxin even if its Doubao and Qwen scores are healthy.

Reading the Baidu signal correctly

There are three easy misreadings of the earnings data.

The first is "Baidu is dying, ignore it." Baidu still reported 644 million App users, a larger user base than any native AI app, and its ERNIE assistant's daily active users grew 83% year over year. It is losing share, not relevance. Ignoring the largest search surface in the country is a budget decision that needs evidence.

The second is "AI search will monetize soon, so wait and buy ads." Management's own words describe product experience as the priority. When paid placement inside AI answers does arrive, the brands with existing organic citations will have a stronger baseline and better quality signals to work from. Waiting means starting from zero later.

The third is "one model score represents China." Doubao, Qwen, DeepSeek and Wenxin each draw on different sources and serve different audiences, and our data shows a category-dependent gap of 3.5 to 8.6 points between Wenxin and the strongest model. A brand that only measures on one model will misjudge its exposure.

Takeaway for Brand Marketers

1. Audit Wenxin separately. If your average brand score across models looks healthy but Wenxin trails by more than five points, you have a Baidu ecosystem problem, not a general awareness problem. Beauty, sports and electronics brands should expect the widest gaps.
2. Fix the Baidu entity layer first. Update or create your Baidu Baike entry, verify a Baijiahao account, and publish a first-paragraph Mandarin definition of your brand and flagship products. These are low-cost actions with direct exposure to the engine that still reaches the largest audience.
3. Treat the paid-search budget as a reallocation question. With Baidu ad revenue down 19% and AI-answer ads not yet mature, shifting a measured portion of paid search spend into Mandarin content that engines cite is a defensible test. Track by model, not by a single blended score.
4. Plan for the window. Organic AI-answer presence is unpriced today. Build the citation base now, before placements are sold and incumbents lock in the shortlist.
5. Keep sources honest. Third-party citation studies vary widely in method. Use them to form hypotheses and confirm with your own repeated queries across all six models.

Related: Brand scores by model and category, Wenxin (ERNIE) brand score discount in China AI, Baidu ERNIE 5.1 and brand GEO

Sources: Baidu Q2 2026 earnings call transcript (The Motley Fool, Aug 18, 2026); Yicai and 36Kr Q2 coverage; Tencent News analysis of Baidu user decline (Aug 31, 2026); QuestMobile Q1 2026 AI application insights; Transgenia September 2026 China SEO/AEO guide; hubGEO brand_scores and ai_citation_sources data.