From "Baidu It" to "Doubao It" — A Change in More Than Method

11/16 2026 472

Are search engines going backwards?

Image source: Internet (please contact us for removal if infringing). Partly AI-generated.

A friend of mine who worked in SEO switched careers last year. Before leaving, he said one thing that stuck with me: clients used to ask "how do I get Baidu to index me?" Now they ask "how do I get Doubao to recommend me?" The question changed, and the answer changed with it.

He no longer needs to study Baidu's algorithm updates. Instead, he needs to figure out how to feed content into the retrieval chain of large language models.

That single shift says more about the problem than any set of monthly active user data ever could. Baidu is still here. The search box is still here. But the entire business logic, content ecosystem and user habits built around "search" are being replaced at the root.

Baidu is not the victim of this shift. It was the first to turn "search" into a toll business — and the first to be bitten back by the very tolls it once collected.

In the PC era, Baidu was never really about "giving answers"

Many people are nostalgic for the "Baidu It" era, describing it as a clean, efficient information-retrieval experience. That nostalgia filters out a key fact: what Baidu truly monopolized in the PC era was not the quality of answers, but the must-pass-through node users had to traverse before reaching information.

In an era when web pages were scattered and every website went its own way, the search engine's power came from "locating" — it did not produce content; it decided who was seen, who got ranked on page three, and who did not exist at all.

Advertisers were not buying exposure. They were buying the right to intercept users along their decision path. The premise of this logic was simple: users had no other way to go.

Baidu made this road very wide — and then built the toll booth in the middle of it.

But this power structure had a built-in fragility: it depended on the existence of the "middle." Once content supply began concentrating on a few super-platforms, users could bypass the middle node and complete retrieval directly where the content lived.

WeChat Search has over 800 million MAU, with nearly half of WeChat users using it every day. Douyin averages 5.3 billion searches per day — more than ten times Baidu App's 300 million. Xiaohongshu reaches 800 million daily searches, and 72% of users go there to search before making consumption decisions.

These platforms do not require users to "go search specifically." Search happens where they already are.

Baidu's predicament in the mobile era is often summarized as "having its traffic taken by ByteDance and Tencent."

But a more accurate statement is this: Baidu never owned "search." What it owned was an unavoidable node in a particular historical stage. Once that node can be bypassed, the power ends.

In the AI era, Baidu keeps on taxing

Facing the loosening of the search entrance, Baidu did not stand still. In 2026, Baidu upgraded the search box into a "smart box," embedded AI-generated content into 70% of mobile search results pages, and packaged the Wenxin task engine as an Agent entrance capable of "delivering tasks."

From a product narrative standpoint, this is a generational upgrade.

But what users actually experience is another matter entirely.

In January 2026, Kimi officially called out Baidu: when searching for "Kimi official website," the first four results were all ads. The light-colored small-print labels were almost invisible, the real official website was pushed to the end of the results page, and some users mistakenly paid 199 yuan for a third-party membership.

What is more noteworthy than this single case is how Baidu handles competitors' brand words: when searching for DeepSeek, Yuanbao, or Qwen, the brand words are likewise sold to bidding advertisers — yet when searching for "Wenxin Yiyan," the results are entirely undisturbed.

Rules can be selectively enforced — as long as the target is one's own people.

After Wenxin 5.1 launched, two phrases repeatedly appeared in user feedback: "got dumber" and "ads mixed in." This is not a sudden degradation of model capability, but the projection onto the product layer of a structural contradiction: when the answers of AI search are priced by bidding rank, the "intelligence" of those answers can never exceed the needs of commercialization.

Baidu stuffing AI into the search box is essentially not about redefining search with AI. It is more like giving the old toll booth a prettier sign that says "AI."

The "forced sale" design — where clicking a search suggestion jumps to an AI conversation — was complained about by users as "I clearly wanted to search the web, but you insist on showing me AI." That sentence could almost serve as the epitaph of Baidu's AI search transformation: it did AI technically, but logically it is still doing traffic interception from the search era.

Data is more honest than narrative. Baidu's search market share fell from 86.8% in 2021 to 50.92% in mid-2025, and further declined to the 44.6%–50.8% range in 2026.

Baidu App's MAU dropped by 69 million in one year, from 724 million to 655 million. It is not that Baidu has no AI revenue — in Q1 2026, AI business revenue accounted for more than half for the first time — but profits are shrinking, and Wenxin has been pushed out of the first tier of large models by Doubao and DeepSeek.

Li Yanhong himself admitted in an internal speech that Baidu "has tried and also made mistakes," and that its top-level AI strategy "has always been wavering."

Baidu's real problem is not "doing AI late." It did it very early — Wenxin Yiyan was hastily launched in March 2023, earlier than most domestic players.

But every time it tried to fit new technology into the shell of the old business model, that shell was leaking.

Doubao makes "search" no longer important

If measured by search capability alone, Doubao is far from perfect.

It prefers citing Toutiao and Douyin Baike from the same corporate group, and most of its citations of individually published content also come from the Toutiao ecosystem.

Some tests found that Doubao produces "misattribution"-style hallucinations in fact-checking, stitching information from different sources into a seemingly reasonable but inaccurate answer.

The source-quality problems and GEO gray-market pollution facing AI search overall do not spare Doubao. Some agencies charge 299 yuan to help merchants "buy" recommendation slots in AI search by mass-producing list-style content to cater to large models' retrieval preferences.

But these problems have not stopped users from migrating. The reason is not on Doubao's side — it is on Baidu's.

What Doubao has achieved is essentially not "searching better," but turning search from an action that must be specifically performed into a step in a conversation that does not need to be named.

Users do not need to open something called "search." They do not need to filter and compare among blue links. They just say the question, get an answer, and ask follow-ups if unsatisfied.

ByteDance merging the TRAE, Coze and Feishu teams into the Doubao system and launching "Doubao Work" further shows that its goal is to make Doubao a unified entrance for all tasks, with search merely one of the capabilities hidden behind conversation.

The driving force of this migration is less "AI is smarter than search" than the fact that users' patience with the product form of "search" has been exhausted.

Baidu's search experience over the past decade has taught users one thing: you enter a word and get a bunch of results you must judge yourself, filter yourself, and bear the consequences of yourself — and above and in the middle of those results, commercial content you do not want is always waiting for you.

When a new interaction method promises "you no longer have to flip through it yourself," user migration is almost instinctive — even if the new answers are also biased, hallucinated, and at risk of contamination.

Users are not choosing between two search engines. They are choosing between "continuing to endure a known bad experience" and "betting on a possibly better interaction method."

Baidu still has a share in the stock migration of AI search. Its index depth on PC and in professional verification scenarios has no short-term substitute, and traditional search has receded into a tool for source verification.

The change search is undergoing is not replacing one product with another, but turning from a business of "intercepting users" into a capability of "embedding into tasks."

Baidu is betting that it can split search capability into APIs sold to developers and Agent calls, surviving by becoming infrastructure.

This path may not be wrong, but its commercial value is inherently lower than the taxing right of the entrance era. Tools can survive, but tools do not hold power.

What is truly worth asking is not "whether Baidu can turn things around," but who will bear the public nature of information access once the interface form of search is dissolved.

Baidu spent twenty years proving one thing: handing the information entrance to a company centered on advertising revenue costs users having to use their own judgment to hedge against commercial erosion.

Doubao and ByteDance behind it are repeating the same structure in another way. Search hidden behind conversation does not mean the commercial impulse has disappeared; it has merely moved to a place less easy to identify.

The GEO gray market is already reminding everyone: the "objective answers" of AI search can likewise be priced.

Search has changed from "Baidu It" to "Doubao It." Users changed one action, but did not change the fundamental question: when the middleman of "finding information" changes, who decides for users what should be seen — and who should supervise that decision-making power?

Baidu has not answered this question well. Doubao has not yet either.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.