Coaley Peak GEO Research Centre Engine guide
Baidu and ERNIE 百度 · 文心 How China's leading search engine became an answer engine, and how to be its source.
Baidu (百度) is China's leading search engine, and it has rebuilt its results around AI: by October 2025 about 70% of its mobile results pages carried AI-generated content. Its answers come from Baidu's own index and its ERNIE (文心) models. This guide covers its timeline, the four layers behind its answers, Baiduspider and how to opt out.
Independent guide: Coaley Peak is not associated with Baidu, Inc. Trademarks
- Owner
- Baidu, Inc.
- Launched
- Baidu 2000; ERNIE Bot March 2023; AI answers in Baidu Search 2025
- Home market
- China: 46.65% of searches in September 2026 (StatCounter)
- Languages
- Chinese
- Open weights
- ERNIE 4.5 under Apache 2.0; ERNIE 5.x closed
- Crawlers
- Baiduspider and its variants; no AI-specific token
Overview
What Baidu and ERNIE is, and who uses it
Baidu was founded in 2000 and has led Chinese search ever since. StatCounter put it at 46.65% of searches in China in September 2026, ahead of Bing, Yandex, Haosou and Sogou, with Google at 1.46%; StatCounter samples sites that carry its tracker, so treat the figures as a guide. The Baidu App had 644 million monthly users in June 2026, according to Baidu's own results.
Baidu unveiled ERNIE Bot (文心一言) in March 2023 and opened it to the public that August, after regulatory approval. It has since brought AI into search itself: an 'AI搜' entry on the Baidu homepage in January 2025, DeepSeek R1 and ERNIE deep search in Baidu Search in February 2025, and in July 2025 what it called the biggest redesign of Baidu Search in ten years. Baidu says link-based results are being progressively replaced by AI answers.
The models are Baidu's own. ERNIE 4.5 was open-sourced under Apache 2.0 in June 2025; ERNIE 5.0, a 2.4-trillion-parameter model that works across text, images, audio and video, was released in January 2026, and ERNIE 5.1 followed in May 2026. The assistant app has changed names twice, from 文心一言 to 文小言 in 2024 and to 文心 in November 2025, and had 202 million monthly users in December 2025.
For brands entering China, Baidu is the first search engine to get right, and its controls are blunt: Baiduspider is one crawler for everything, there is no AI-specific opt-out, and hosting a Chinese-language site outside mainland China can hurt how it appears. In September 2026 Baidu announced an AI traffic analysis tool for its webmaster platform, which will report how often a site is cited in AI answers.
Timeline
Baidu and ERNIE, from launch to today
Baidu, Inc. is founded.
Baidu renames its webmaster platform the Baidu Search Resource Platform (搜索资源平台).
Robin Li unveils ERNIE Bot (文心一言) at a launch event in Beijing.
ERNIE Bot opens to the public after Chinese regulatory approval.
The 文心一言 app is renamed 文小言 and repositioned as a 'new search' assistant.
An 'AI搜' entry, built on ERNIE, is added to the Baidu Search homepage.
DeepSeek R1 and ERNIE deep search go live in Baidu Search.
Baidu open-sources ten ERNIE 4.5 models under Apache 2.0. By the end of June, more than half its mobile results pages contain AI-generated content.
Baidu Search gets its biggest redesign in ten years: a larger search box, structured multimedia results and an AI assistant with deep search.
ERNIE 5.0 is unveiled at Baidu World as a natively omni-modal model, and Baidu says about 70% of mobile results pages carried AI content in October.
The 文小言 app is renamed 文心.
ERNIE 5.0 is officially released, with 2.4 trillion parameters.
ERNIE 5.1 launches, more compact and with stronger reasoning.
Baidu's Search Resource Platform announces an AI traffic analysis tool, reporting AI citations, exposure and clicks for a site's content.
Things to know
Things to know about Baidu and ERNIE
Baidu says about 70% of its mobile results pages carried AI-generated content by October 2025, up from 35% in April 2025, and that link-based results are being replaced by AI answers.
Blocking Baiduspider blocks every Baidu product and removes the site from Baidu Search. Baidu publishes no separate token for AI training or AI answers.
Two of Baidu's documented crawlers, Baiduspider-cpro and Baiduspider-ads, do not obey robots.txt.
Baidu publishes no IP list. Verify Baiduspider with a reverse DNS lookup to a .baidu.com or .baidu.jp hostname, then a forward lookup.
robots.txt changes take about 48 hours to take effect, but pages already indexed can take months to drop out.
Baidu has warned that Chinese-language sites hosted outside China can display poorly in its results, and advises against overseas servers for them. Hosting in mainland China needs an ICP filing or licence.
Do not block Baiduspider-render, CSS or JavaScript, and do not load valuable content only through JavaScript, which Baidu has said risks indexing losses.
Baidu's AI traffic analysis tool, announced on 15 September 2026, will report AI citations, exposure and clicks for your content. It is opening in phases.
ERNIE 3.0's training corpus drew heavily on Baidu's own platforms: Baidu Baike, Baijiahao, Zhidao and Tieba.
Since 1 September 2025, AI-generated content published in China must be labelled under the Cyberspace Administration of China's rules.
The four layers
How Baidu and ERNIE builds an answer
Every AI answer is made in four layers: what the model learned in pre-training, how it was tuned, what it retrieves live, and how it puts the answer together. Here is each one for Baidu and ERNIE, and what it means for a brand that wants to be cited.
Layer 1
Pre-training: the base (foundation) model
ERNIE (文心) is Baidu's own model family. ERNIE 4.5, open-sourced under Apache 2.0 in June 2025, includes a 300-billion-parameter mixture-of-experts model with 47 billion parameters active. ERNIE 5.0, released in January 2026, has 2.4 trillion parameters and works natively across text, images, audio and video; ERNIE 5.1 followed in May 2026. Training data and cut-offs for these are not published, but ERNIE 3.0's paper describes a corpus built largely from Baidu's own platforms, including Baidu Baike, Baijiahao, Zhidao, Tieba and Baidu Search data, plus web text and a knowledge graph.
What it means for you: Baidu's own platforms are part of what ERNIE learns from. A brand with an accurate Baidu Baike entry, articles on Baijiahao and good answers on Zhidao is more likely to be known to ERNIE than one present only on its own website.
Layer 2
Post-training: tuning and alignment
Baidu describes supervised fine-tuning and preference optimisation for ERNIE 4.5, and for ERNIE 5.0 multi-turn reinforcement learning on long tasks, guided by a programme of 835 outside experts who calibrate the model, including on values. The service is filed with the Cyberspace Administration of China; Baidu reviews inputs and outputs, and AI output carries visible and hidden labels, as Chinese rules require.
What it means for you: Answers follow Chinese content rules and Baidu's own review. For most commercial topics that shapes tone rather than which brand is named, but claims that would breach Chinese advertising law are unlikely to be repeated.
Layer 3
Retrieval: live search
Baidu answers from its own index, built by Baiduspider, and from its own content properties: its AI search launched connected to Baidu Search, Baidu Health, Baidu Lawyer, Baidu Wenku and Baidu Education content. Since February 2025 users can switch to DeepSeek R1 or ERNIE deep search, and the July 2025 redesign added deep search to its AI assistant and opened search to more than 18,000 MCP services. The assistant also uses DeepSeek's open-source models.
What it means for you: Ranking in Baidu is the way into its answers, and so is a presence on Baidu's own properties, which it draws on directly. Classic Baidu SEO, compliant hosting in China and content on Baidu's platforms work together.
Layer 4
The answer: what people read
Baidu says link-based results are being progressively replaced by structured, multimodal AI answers, and its July 2025 redesign introduced '百看' results that mix text, images, audio and video. Baidu has not published a citation format, but its new webmaster tool will report 'AI citations' (AI引用) of a site's content, which confirms that its AI answers cite and link to sites. AI output is labelled.
What it means for you: Clicks on classic links are falling on mobile, so citations in AI answers are where visibility is moving. Baidu's AI traffic analysis will be the first official way to measure them; set it up as soon as your site has access.
Crawlers
Baidu and ERNIE's crawlers and user agents
The user-agent tokens to look for in your server logs and to name in robots.txt.
| User-agent token | What it does | robots.txt |
|---|---|---|
| Baiduspider Documentation | Baidu's web and mobile search crawler, which builds the index behind Baidu Search and its AI answers. PC user agent: Mozilla/5.0 (compatible; Baiduspider/2.0; +http://www.baidu.com/search/spider.html). | Follows robots.txt, with * and $ wildcards and case-sensitive paths. Changes take about 48 hours. |
| Baiduspider-render Documentation | Fetches CSS, JavaScript and images so Baidu can render and understand pages. User agent: Mozilla/5.0 (compatible; Baiduspider-render/2.0; +http://www.baidu.com/search/spider.html). | Not separately documented. Baidu asks sites not to block it. |
| Baiduspider-image Documentation | Crawls for Baidu image search. | Follows robots.txt, and can be given its own rules. |
| Baiduspider-news Documentation | Crawls for Baidu news search. | Follows robots.txt, and can be given its own rules. |
| Baiduspider-video Documentation | Crawls for Baidu video search. | Unclear: Baidu's own pages disagree on whether it follows robots.txt. |
| Baiduspider-cpro Documentation | Crawls for the Baidu Union advertising network. Pages it fetches are not added to the search index. | Does not obey robots.txt, Baidu says. |
| Baiduspider-ads Documentation | Crawls for Baidu's search advertising. Pages it fetches are not added to the search index. | Does not obey robots.txt, Baidu says. |
Only the main Baiduspider and Baiduspider-render user agents are published in full; the image, news, video and advertising crawlers are documented by token only. Baidu documents no crawler specifically for ERNIE training or AI answers.
Crawl sources
Where Baidu and ERNIE crawls from, and how to verify it
Baidu publishes no IP ranges or networks for its crawlers. Its documented check is a reverse DNS lookup followed by a forward lookup: genuine hostnames end in .baidu.com or .baidu.jp, such as baiduspider-123-125-66-120.crawl.baidu.com, and the forward lookup must return the original IP address. Baidu says any other hostname is an impostor.
Baidu documents the commands for Linux, Windows and macOS, and asks site owners to report abnormal crawling through its complaint platform, with access logs. Complaints about the Baidu Union advertising crawler go to union1@baidu.com.
Opting out
How to opt out of Baidu and ERNIE, or limit what it uses
Block all of Baidu
User-agent: Baiduspider with Disallow: / blocks every Baidu product and removes the site from Baidu Search and the search services Baidu supplies. Product tokens such as Baiduspider-image allow finer rules.
Official pageTest your rules first
Baidu's Search Resource Platform has a tool for checking and updating robots.txt, which Baidu recommends using after any accidental block.
Official pageHide cached snapshots
A noarchive meta robots tag lets Baidu index a page without showing its cached snapshot. Baidu says the change can take two to four weeks to show.
Official pageUrgent removals and crawl problems
Send urgent de-indexing requests and reports of abnormal crawling to Baidu's complaint platform or the webmaster feedback centre, with the domain and crawl logs.
Official pagePersonal data in results
Baidu says it cannot edit third-party pages: ask the source site to change or remove the content, or use the report button at the bottom of the search result to ask Baidu to de-link it.
Official pageAdvertising crawlers
Baiduspider-cpro and Baiduspider-ads ignore robots.txt. Contact union1@baidu.com about the first, and your Baidu account representative about the second.
Official page
There is no way to stay in Baidu Search while opting out of ERNIE training or AI answers: Baidu publishes no AI-specific token, so the only control is blocking Baiduspider, which removes you from Baidu altogether. For a brand that wants Chinese customers, the realistic goal is to be cited accurately, not to opt out.
How we help
How we get brands cited in Baidu and ERNIE
The Coaley Peak GEO Research Centre researches and implements GEO for SEO agencies and in-house teams, with foreign markets as our speciality.
Baidu AI answer audit
We search the Chinese queries your customers use, on mobile, and record which AI answers appear, which sources they cite, and whether you are among them.
Baidu SEO foundations
Compliant hosting in China, Baidu Search Resource Platform set-up, crawlability for Baiduspider and Baiduspider-render, and server-rendered content, so Baidu can index you properly.
Presence on Baidu's own platforms
Baidu draws on Baike, Baijiahao, Zhidao and Wenku. We help you build an accurate, compliant presence there, which both its search and its models learn from.
Native Chinese content
With native Chinese writers, we create pages that answer your customers' questions in simplified Chinese, in the form Baidu's AI answers quote.
AI traffic measurement
We set up Baidu's AI traffic analysis as it opens, and track your AI citations in Baidu alongside DeepSeek, Doubao, Qwen and Yuanbao.
More on our research, services and team at the Coaley Peak GEO Research Centre.
FAQ
Baidu and ERNIE GEO: questions
How much of Baidu is AI answers now?
Baidu says about 70% of its mobile results pages carried AI-generated content by October 2025, up from 35% in April 2025, and that link-based results are being progressively replaced by AI answers.
Can I stop Baidu using my site for AI but stay in its search?
No. Baidu publishes no AI-specific crawler token or opt-out. Blocking Baiduspider removes your site from Baidu Search and every other Baidu product.
How do I verify Baiduspider?
Run a reverse DNS lookup on the IP: a genuine Baiduspider hostname ends in .baidu.com or .baidu.jp. Then look that hostname up and check it returns the same IP. Baidu publishes no IP list.
Does hosting outside China affect Baidu?
Baidu has warned that Chinese-language sites on overseas servers can display poorly in its results, and advises against overseas hosting for them. Hosting in mainland China requires an ICP filing or licence.
What is ERNIE?
ERNIE (文心) is Baidu's family of AI models. ERNIE 4.5 was open-sourced under Apache 2.0 in June 2025, ERNIE 5.0, with 2.4 trillion parameters, was released in January 2026, and ERNIE 5.1 in May 2026. They power Baidu's AI search and its assistant, now called 文心.
Can I measure how often Baidu's AI cites my site?
Soon. On 15 September 2026 Baidu announced an AI traffic analysis tool for its Search Resource Platform, reporting AI citations, exposure and clicks for a site's content. It is opening in phases.
Want a free chat about Baidu and ERNIE?
Tell us which market and brand you are working on. If we can help, we will set up a free, no-obligation chat with Stephen or Ethan.
More engine guides
Related engines
Independence and trademarks. Baidu, 百度, ERNIE, 文心 and 文心一言 are trademarks of Baidu, Inc. or its affiliates. DeepSeek is a trademark of its respective owner. Coaley Peak is independent: we are not associated with, endorsed by or a partner of Baidu, Inc. This guide is our own research from public documentation and reputable reporting, checked in October 2026.
Reviewed by Stephen Grindley, research centre lead. Last checked .
Document reference: ISO_webpage_geo_engine_baidu_v1
Last modified: 7 October 2026
GEO Research Centre·Engine guides·Baidu and ERNIE