Coaley Peak GEO Research Centre Engine guide
Doubao 豆包 How China's most-used AI assistant finds its sources, and how to become one.
Doubao (豆包) is ByteDance's AI assistant and, by third-party estimates, the most-used AI app in China. It answers in Chinese from ByteDance's own Seed models and from third-party web pages it searches, showing its sources above each answer. Here is its history, the four layers behind its answers, Bytespider and how to opt out.
Independent guide: Coaley Peak is not associated with ByteDance. Trademarks
- Owner
- ByteDance
- Launched
- August 2023 (iOS app 24 August 2023)
- Home market
- Mainland China; abroad, ByteDance offers Dola instead
- Languages
- Chinese
- Open weights
- No: Doubao's Seed models are proprietary
- Crawlers
- Bytespider (search and training); no Doubao-specific bot documented
Overview
What Doubao is, and who uses it
ByteDance, the company behind Douyin, Toutiao and TikTok, released Doubao in China in August 2023, just as China's rules for generative AI services took effect. It opened the models behind it, formerly called Skylark, to businesses through its Volcano Engine cloud in May 2024. Doubao grew from about 26 million monthly users then to an estimated 382 million in May 2026, according to QuestMobile: the largest of any AI-native app in China.
Doubao is a mainland-China, Chinese-language service. Its operator is a Beijing company, it stores users' data in China, and its official app was not in the UK, US, Singapore or Indian App Stores when we checked in October 2026. Outside China, ByteDance offers a separate assistant, Dola (formerly Cici), run from Singapore. Apps called 'Doubao AI' in the UK App Store are not ByteDance's.
Doubao is woven into ByteDance's platforms. It runs inside Toutiao, a bound Douyin account brings personalised Douyin search results into its answers, and its language model is filed for use across Toutiao, Douyin, Jianying, Fanqie Novel, Xigua Video and Feishu. The basic tier is free, with paid tiers at ¥68, ¥200 and ¥500 a month.
For brands selling into China, Doubao is the first AI assistant to plan for. When it detects search intent it searches public third-party pages and shows them as sources above its answer, which makes being one of those pages the visible prize.
Timeline
Doubao, from launch to today
ByteDance, founded in 2012 by a team led by Zhang Yiming and Liang Rubo, launches its first flagship app, Toutiao.
ByteDance's overseas assistant Cici, now Dola, first appears on the App Store, in markets including Mexico.
China's Interim Measures for Generative AI Services take effect, requiring security assessments and algorithm filing.
The Doubao iOS app is first released in China's App Store.
ByteDance launches its Doubao models, formerly Skylark, on Volcano Engine. Caixin reports the Doubao app at more than 26 million monthly users.
Fortune reports data from the security firm Kasada saying Bytespider crawled about 25 times faster than OpenAI's GPTBot and did not respect robots.txt.
ByteDance releases Doubao-1.5-pro, a mixture-of-experts model.
ByteDance Seed releases the open-weight Seed-OSS-36B models under Apache 2.0, separate from the models behind Doubao.
ByteDance launches the Seed2.0 (Doubao 2.0) series of models.
Seed2.1 (Pro and Turbo) is released and becomes available to Doubao users.
QuestMobile reports Doubao reached 382 million monthly active users in May 2026, the most of any AI-native app in China.
Doubao takes its user-built agents offline as China's rules on anthropomorphic AI services take effect.
Volcano Engine launches a 'Doubao Search' API for AI agents, which grades sources by site and creator authority.
ByteDance's EU training summary for Seed 2.0 Pro names Bytespider as its AI training crawler and says it respects robots.txt.
Things to know
Things to know about Doubao
Doubao is China's largest consumer AI app: 382 million monthly users in May 2026, according to QuestMobile. ByteDance publishes no user count of its own.
When Doubao detects search intent, or the user chooses AI search, it searches public third-party web pages and shows its sources above the answer.
One crawler, Bytespider, serves both Toutiao Search indexing and ByteDance's AI training. There is no training-only token, so blocking it is all or nothing.
ByteDance documents no Doubao-specific search or fetch bot, and does not say which index Doubao's live search uses.
Bytespider can be verified: ByteDance publishes ten IPv4 ranges and says genuine hostnames end in .bytedance.com.
Third-party monitors have reported Bytespider as very high-volume and sometimes ignoring robots.txt. ByteDance says its crawler is designed to respect it.
ByteDance's 'Doubao Search' API for AI agents, launched in July 2026, ranks sources by site and creator authority and explicitly targets low-quality 'SEO filler'.
Doubao draws on ByteDance's platforms: a bound Douyin account brings Douyin search results into its answers, and Doubao runs inside Toutiao.
UK users meet ByteDance's assistant as Dola, run by a Singapore company, not as Doubao. Dola's privacy policy names a UK GDPR representative.
Doubao withdrew user-built agents on 15 July 2026 under China's new rules on anthropomorphic AI, so brand agents inside Doubao are no longer a channel.
The four layers
How Doubao 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 Doubao, and what it means for a brand that wants to be cited.
Layer 1
Pre-training: the base (foundation) model
Doubao runs on ByteDance's proprietary Seed models; Seed2.1 has been available in Doubao since June 2026, after Seed2.0 in February. The weights are not public (ByteDance's open-weight Seed-OSS-36B is a separate model). ByteDance's EU summary for Seed 2.0 Pro gives February 2026 as its latest data collection date and lists public datasets, licensed data, pages crawled by Bytespider between about January 2022 and June 2024, data from users of other ByteDance services, and synthetic data: more than 10 trillion text tokens in all.
What it means for you: Doubao's memory of your brand comes largely from Chinese web content Bytespider could reach, and from ByteDance's own platforms. A brand with accurate, substantial Chinese pages and an official presence on Douyin and Toutiao gives the model consistent facts to learn.
Layer 2
Post-training: tuning and alignment
Doubao's filing notice describes unsupervised pre-training, then supervised fine-tuning, then reinforcement learning from human feedback, and its model cards say tuning prioritises instruction-following, long-tail knowledge and long-context stability. The service is filed under China's generative AI rules, which require 'core socialist values', lawful training data, security assessments and algorithm filing. Users' de-identified content may be used for training unless they switch it off.
What it means for you: Answers follow Chinese content rules, and long-tail knowledge is a stated tuning goal, which rewards brands with clear, specific information about niche products. Claims that would not pass Chinese advertising or content rules are unlikely to be repeated.
Layer 3
Retrieval: live search
Doubao's user agreement says that when it detects search intent, or the user picks AI search, it automatically searches public sources on third-party web pages and answers from them; its filing describes 'whole-web search'. With a bound Douyin account, Douyin returns personalised results inside Doubao. It uses location for local results and has a browser plug-in that answers questions about the page in view. The index or provider behind the app is not published.
What it means for you: Being in the pages Doubao finds is the main lever, and those are Chinese pages. Ranking for the Chinese questions your customers ask, in the search engines and on the platforms ByteDance can draw on, including Douyin and Toutiao, is how you get into its sources.
Layer 4
The answer: what people read
Doubao's agreement says 'search sources and related content are displayed above the reply', and warns that answers may still be wrong. Its modes include search, deep search and academic search, reading web pages, image and video generation, voice calls with real-time interpretation, memory, document and slide generation, and Douyin shopping with a bound account.
What it means for you: Sources sit at the top of the answer, where users see them first. Pages that answer the question directly in Chinese, from a site or creator Doubao treats as authoritative, are the ones most likely to be shown.
Crawlers
Doubao'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 |
|---|---|---|
| Bytespider Documentation | ByteDance's crawler for Toutiao Search and AI training, according to its webmaster documentation and EU training summaries. Whether Doubao's live search uses it is not published. User agent: Mozilla/5.0 (compatible; Bytespider; https://zhanzhang.toutiao.com/) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/70.0.0.0 Safari/537.36. | ByteDance says it is designed to respect robots.txt; third-party monitors have reported otherwise. |
| Dola-SearchBot | Listed by a third-party directory as a search crawler for ByteDance's Dola assistant, usually visiting from Singapore. ByteDance does not document it. | Not published |
| Doubaobot | Listed by a third-party directory as supporting Doubao's training and answers. ByteDance does not document it, and no user agent is published. | Not published |
Bytespider also has Android and iOS user agents ending '(compatible; Bytespider; https://zhanzhang.toutiao.com/)'. Only Bytespider is documented by ByteDance; treat the other two as unverified.
Crawl sources
Where Doubao crawls from, and how to verify it
ByteDance publishes Bytespider's addresses on its Toutiao Search webmaster pages as ten IPv4 /24 ranges: 110.249.201.0/24, 110.249.202.0/24, 111.225.148.0/24, 111.225.149.0/24, 220.243.135.0/24, 220.243.136.0/24, 220.243.188.0/24, 220.243.189.0/24, 60.8.123.0/24 and 60.8.151.0/24. There is no machine-readable feed and no IPv6, and the page was last updated in April 2024.
To verify a request, run a reverse DNS lookup. ByteDance says genuine Bytespider hostnames end in .bytedance.com, such as bytespider-111-225-148-148.crawl.bytedance.com, and that anything else is an impostor. ByteDance publishes no networks, and no IPs or hostnames for any of its other agents.
Opting out
How to opt out of Doubao, or limit what it uses
Block or limit Bytespider in robots.txt
User-agent: Bytespider with Disallow: / or path rules; Toutiao Search supports the * and $ wildcards. One token covers both Toutiao Search and AI training, and no document says whether blocking it affects Doubao's cited sources.
Official pageEnforce it at the firewall
Because third parties report Bytespider fetching disallowed URLs, match requests against the ten published ranges and the .bytedance.com reverse DNS rule, and block at your CDN if you need a hard stop.
Official pageLower the crawl rate
Verified site owners can cut Bytespider's crawl rate for each site in the Toutiao Search webmaster platform. Subdomains are set separately, and changes apply from the next day.
Official pageRemove dead URLs from Toutiao Search
Submit dead links to Toutiao Search's removal tool; the pages must return 404. This applies to Toutiao Search, not explicitly to Doubao's answers.
Official pageReport a wrong or infringing answer
Use in-app feedback on the answer, or email feedback@mail.doubao.com. Rights holders can send a signed notice with proof of rights to jubao@mail.doubao.com.
Official pagePersonal data requests
For Doubao, email Dataprivacy@bytedance.com; replies follow identity checks. For Dola, which is what UK users use, email privacy@mail.dola.com; UK users can also complain to the ICO.
Official page
There is no training-only opt-out for publishers: ByteDance offers no Google-Extended-style token, meta tag or form, and its EU summary cites only robots.txt. Blocking Bytespider removes you from Toutiao Search and ByteDance's training together, and may or may not affect Doubao's live answers.
How we help
How we get brands cited in Doubao
The Coaley Peak GEO Research Centre researches and implements GEO for SEO agencies and in-house teams, with foreign markets as our speciality.
Doubao visibility audit
We ask Doubao the Chinese questions your customers ask and record whether you appear among the sources above the answer, what it says about you, and which sites it trusts instead.
Chinese content built for Doubao's sources
With native Chinese writers, we create pages that answer questions directly in simplified Chinese, on a site Doubao's search can recognise as an authority in your field.
ByteDance platform presence
Doubao draws on Douyin and Toutiao. We help you build an official, consistent presence there, so the platforms Doubao searches describe you the way your website does.
Bytespider policy and verification
We set robots.txt and firewall rules for Bytespider that match your goals, verified against ByteDance's published ranges, and monitor what it actually fetches.
China market entry
We put Doubao alongside Baidu, DeepSeek, Qwen, Kimi and Yuanbao in one plan, so your entry into China covers the assistants people there really use.
More on our research, services and team at the Coaley Peak GEO Research Centre.
FAQ
Doubao GEO: questions
Can people in the UK use Doubao?
Doubao is a mainland-China, Chinese-language service, and its official app was not in the UK App Store when we checked in October 2026. ByteDance's assistant for UK users is Dola, run by a Singapore company. Apps called 'Doubao AI' in the UK store are not ByteDance's.
Does Doubao show its sources?
Yes. When Doubao searches the web, its user agreement says search sources and related content are shown above the reply. It also warns that answers may still be inaccurate.
Which crawler does Doubao use?
ByteDance documents only Bytespider, which serves Toutiao Search and AI training, and does not say which index Doubao's live search uses. Bots called Doubaobot and Dola-SearchBot appear in third-party directories, unconfirmed by ByteDance.
Can I stop ByteDance training on my site but stay in its search?
Not with a documented control. Bytespider is one token for both Toutiao Search and AI training, and ByteDance offers no training-only token, meta tag or form.
How many people use Doubao?
ByteDance publishes no figure. QuestMobile estimated 382 million monthly active users in May 2026, the most of any AI-native app in China, ahead of Qwen and DeepSeek.
How do I get my brand into Doubao's answers?
Publish clear, authoritative Chinese pages that answer the questions your customers ask, make sure Bytespider and China's search engines can reach them, and build a consistent official presence on Douyin and Toutiao, which Doubao draws on.
Want a free chat about Doubao?
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. Doubao (豆包) and Dola are trademarks of ByteDance or its affiliates. Coaley Peak is independent: we are not associated with, endorsed by or a partner of ByteDance. 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_doubao_v1
Last modified: 7 October 2026
GEO Research Centre·Engine guides·Doubao