Coaley Peak GEO Research Centre Engine guide
DeepSeek 深度求索 How China's open-weight AI answers, what it learned from, and how to be cited.
DeepSeek shook the AI industry in January 2025 with a free assistant and open models that matched far costlier rivals. Its weights are MIT-licensed and run inside products worldwide, and its app searches the web only when the user asks. This guide covers its timeline, the four layers behind its answers, the crawler question and how to opt out.
Independent guide: Coaley Peak is not associated with Hangzhou DeepSeek Artificial Intelligence Co., Ltd. Trademarks
- Owner
- Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
- Launched
- Company July 2023; first models and web chat November 2023; apps January 2025
- Home market
- China first; used worldwide
- Languages
- Strongest in Chinese and English; replies in the user's language
- Open weights
- Yes: DeepSeek-R1, V4 and V4.1-Flash under the MIT licence
- Crawlers
- None documented; Common Crawl named as a training source
Overview
What DeepSeek is, and who uses it
DeepSeek was founded in Hangzhou in July 2023 by Liang Wenfeng, whose quantitative hedge fund, High-Flyer, was its original backer. It released its first open models, DeepSeek LLM 7B and 67B, in November 2023, alongside a web chat. DeepSeek-V3 followed in December 2024, and the R1 reasoning model in January 2025, under the MIT licence.
R1 and the free DeepSeek app changed the market's view of what AI costs. In late January 2025 the app topped Apple's US App Store chart, and Nvidia's shares fell by more than 17% in a day. DeepSeek-V4, with a context window of a million tokens, arrived in April 2026, and V4.1-Flash, with native vision, in September 2026. In July 2026 Caixin reported DeepSeek's first outside funding round, at a valuation above 350 billion yuan.
DeepSeek publishes no user numbers. QuestMobile estimated 130 million monthly users of its app in China in May 2026, third behind Doubao and Qwen. Outside China its influence runs mostly through its models: because the weights are open, many products run DeepSeek models with their own search and citations, and DeepSeek's terms allow its outputs to be used to train other models.
For SEO teams DeepSeek is an unusual case. It has no search index, no crawler documentation and no webmaster tools of its own. Its app answers from training data unless the user turns Search on, when it queries third-party search engines it does not name and cites the pages it uses. Visibility depends on what was in its training data, and on how you rank in engines it does not name.
Timeline
DeepSeek, from launch to today
Liang Wenfeng founds Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. in Hangzhou.
DeepSeek releases its first general models, DeepSeek LLM 7B and 67B, as open weights; its web chat goes live the same month.
DeepSeek-V2 is released as an open model.
Web search goes live in DeepSeek's web chat; the API does not get it.
DeepSeek-V3 is released with open weights, pre-trained on 14.8 trillion tokens.
DeepSeek announces its free app on the App Store and Google Play, with web search and DeepThink reasoning.
DeepSeek-R1, a reasoning model, is released under the MIT licence.
DeepSeek tops Apple's US App Store chart, and Nvidia's shares fall more than 17% in a day.
Italy's data protection authority blocks DeepSeek over its handling of personal data.
Berlin's data protection commissioner asks Apple and Google to remove DeepSeek from German app stores over data transfers to China.
The US National Institute of Standards and Technology's Center for AI Standards and Innovation publishes an evaluation reporting security and accuracy shortcomings in DeepSeek models.
DeepSeek-V4 launches in preview, with open MIT weights and a one-million-token context across DeepSeek's services.
Caixin reports DeepSeek's first outside funding round: about 50 billion yuan at a valuation above 350 billion yuan, with Tencent and CATL among the investors.
DeepSeek-V4.1-Flash is released as open weights under MIT, with native vision.
Things to know
Things to know about DeepSeek
DeepSeek searches the web only when the user turns Search on. Otherwise answers come from training data that, for V4, was collected up to about May 2025, according to DeepSeek's EU training summary.
Its live search runs on third-party search APIs it does not name, which receive the user's keywords. DeepSeek has no index, crawler documentation or webmaster tools of its own.
DeepSeek's EU training summaries name Common Crawl and Stack Exchange among its pre-training sources, so your CCBot rule affects future DeepSeek training.
DeepSeek published its web-search prompt with R1: the model cites [citation:X] at the end of each supported sentence, draws on several pages rather than one, and keeps list answers to about ten points.
The weights are open under the MIT licence, and DeepSeek's terms allow its outputs to train other models. Many third-party products run DeepSeek models with their own retrieval and citations.
The DeepSeek API has no built-in web search, so 'DeepSeek-powered' products search however their developers choose.
The 'DeepSeekBot' user agent seen in bot directories is not documented by DeepSeek, so its log entries cannot be verified the way GPTBot or ClaudeBot entries can.
DeepSeek stores personal data in mainland China. Italy's regulator blocked it in January 2025; in the UK, the government said using it is a personal choice.
On the UK App Store, DeepSeek's official app appeared as 'DSeek - Your AI Assistant' when we checked on 6 October 2026.
DeepSeek's China service is filed with the Cyberspace Administration of China, and its answers follow Chinese content rules, including on politically sensitive topics.
The four layers
How DeepSeek 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 DeepSeek, and what it means for a brand that wants to be cited.
Layer 1
Pre-training: the base (foundation) model
DeepSeek-V4 (V4-Pro has 1.6 trillion parameters, 49 billion active) was released in April 2026 and V4.1-Flash in September 2026, both as MIT-licensed open weights. DeepSeek's EU summary says V4's data was collected up to about May 2025 and included more than 10 trillion tokens of public internet text, naming Common Crawl and Stack Exchange, plus licensed data, synthetic data from its own models and data written by an internal team. It says user conversations were not used to train V4.
What it means for you: With Search off, DeepSeek knows only what was public by around May 2025, filtered through Common Crawl-style datasets. Being present, consistent and crawlable in that kind of data is what gets a brand into its memory, and, because the weights are open, into every product that runs them.
Layer 2
Post-training: tuning and alignment
DeepSeek uses supervised fine-tuning and reinforcement learning with specialised safety data, and R1 introduced large-scale reinforcement learning for reasoning. Its China service is filed under the Cyberspace Administration of China's generative AI rules, which require providers to uphold 'core socialist values', use lawful training data and label AI output. NIST's 2025 evaluation found DeepSeek models echoed Chinese Communist Party narratives far more often than US reference models.
What it means for you: For most commercial topics this matters little, but answers on China-sensitive subjects follow Chinese rules. Brands in regulated or politically adjacent sectors should test answers carefully before relying on DeepSeek, or on products built on it, to describe them.
Layer 3
Retrieval: live search
Search is a mode the user switches on. DeepSeek's privacy policy says it integrates third-party search APIs and shares the user's keywords with them; the providers are not named. The model extracts several keywords and searches them in parallel. DeepSeek has no search crawler or webmaster tool of its own, and its API has no web search tool.
What it means for you: You cannot submit pages to DeepSeek. Visibility in its cited answers depends on ranking in the unnamed engines behind it, so strong organic visibility in the main engines of your target markets, in Chinese and English, is the practical lever.
Layer 4
The answer: what people read
With Search on, answers carry inline numbered citations. DeepSeek's published prompt asks for each citation at the end of the sentence it supports, drawn from several pages rather than one, with the reply in the user's language. Outputs are labelled as AI-generated, and shared conversations are kept from crawlers with robots rules.
What it means for you: Short, self-contained sentences that state one fact each are the easiest to cite under DeepSeek's own prompt rules. A page that answers the question in its opening lines, with your brand named, has the best chance of being among the sources.
Crawlers
DeepSeek'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 |
|---|---|---|
| DeepSeekBot | Reported by a third-party bot directory as DeepSeek's training-data crawler, with the user agent Mozilla/5.0 (compatible; DeepSeekBot/1.0; +https://www.deepseek.com/). DeepSeek has not confirmed it. | Unverified. DeepSeek's EU summaries say its unnamed crawler is designed to respect robots.txt. |
| CCBot Documentation | Common Crawl's crawler. DeepSeek names Common Crawl as a pre-training source, so CCBot is the documented, indirect route into DeepSeek's training data. | Follows robots.txt, Common Crawl says |
DeepSeek documents no crawler of its own: deepseek.com/bot returns a 404. Its live search uses third-party search providers it does not name, so their crawlers, not DeepSeek's, decide which pages can be cited.
Crawl sources
Where DeepSeek crawls from, and how to verify it
Not published. DeepSeek publishes no IP ranges, reverse DNS hostnames or networks for any crawler. A third-party directory reports that DeepSeekBot usually visits from the United States, but there is no official list to check a request against, so a user-agent match in your logs proves nothing.
The routes you can verify are indirect: CCBot, which publishes its IP ranges and a reverse DNS scheme, for training data; and the crawlers of the main search engines in your markets, for the pages DeepSeek's search providers can return.
Opting out
How to opt out of DeepSeek, or limit what it uses
Training: only a wildcard rule is certain
DeepSeek documents no crawler token. Its EU summaries say its crawler respects robots.txt and text-and-data-mining opt-outs. A rule for DeepSeekBot targets the third-party-reported token; only a wildcard rule is certain to apply, and that blocks every compliant crawler.
Official pageBlock Common Crawl
DeepSeek names Common Crawl as a pre-training source. User-agent: CCBot with Disallow: / keeps your pages out of future snapshots, but not out of models already trained.
Official pageSearch answers: no DeepSeek-side control
Search runs on unnamed third-party APIs, so inclusion in DeepSeek's cited answers depends on those providers' indexes. There is no URL submission and no webmaster tool.
Correct or remove personal data in answers
Email privacy@deepseek.com to ask for a correction or removal of inaccurate personal data in outputs. DeepSeek says it will consider requests based on applicable law and the technical capabilities of its models.
Official pageUK and EU requests through its representative
Prighter is DeepSeek's UK and EU privacy representative: email rep_deepseek@prighter.com with 'Data Subject Request' in the subject line, or use its portal.
Official pageIntellectual property or false information
Use the in-product 'Contact us' button or email service@deepseek.com. DeepSeek's terms are governed by the law of mainland China.
Official page
Opting out of DeepSeek is mostly indirect. DeepSeek's open weights are already downloaded and keep what they learned, and blocking CCBot affects every model trained on Common Crawl, not only DeepSeek. For most brands the better goal is accurate representation, not absence.
How we help
How we get brands cited in DeepSeek
The Coaley Peak GEO Research Centre researches and implements GEO for SEO agencies and in-house teams, with foreign markets as our speciality.
DeepSeek visibility audit
We ask DeepSeek about your brand and category with Search on and off, in Chinese and English, and record what it knows, what it cites and what it gets wrong.
Pages written for DeepSeek's citation rules
DeepSeek's own prompt cites sentence by sentence. We rewrite key pages so each fact stands in one clear sentence with your brand named, in simplified Chinese and English.
Ranking where DeepSeek searches
DeepSeek's search providers are unnamed, so we build strong organic visibility in the main engines of your target markets, which is where its search results are drawn from.
Base model targeting for DeepSeek's open weights
DeepSeek models run inside products worldwide. We probe the open weights directly to see what they learned about you, and work on the training sources future versions draw on.
China market entry
For brands entering China, we plan DeepSeek alongside Baidu, Doubao, Qwen, Kimi and Yuanbao, so you see the whole Chinese AI picture, not one assistant.
More on our research, services and team at the Coaley Peak GEO Research Centre.
FAQ
DeepSeek GEO: questions
Does DeepSeek search the web?
Only when the user turns Search on. DeepSeek then queries third-party search APIs it does not name and cites the pages it uses with inline numbers. With Search off, it answers from training data collected up to about May 2025 for V4.
Does DeepSeek have a crawler I can block?
DeepSeek documents none. A 'DeepSeekBot' user agent appears in third-party bot directories, but DeepSeek has not confirmed it. DeepSeek does name Common Crawl as a training source, so blocking CCBot is the documented, indirect route.
What is DeepSeek trained on?
DeepSeek's EU summary for V4 describes more than 10 trillion tokens of public internet text, naming Common Crawl and Stack Exchange, plus licensed data, synthetic data from its own models and data written by an internal team, collected up to about May 2025.
Are DeepSeek's models open source?
The weights of DeepSeek-R1, V4 and V4.1-Flash are open under the MIT licence, which allows commercial use and distillation. That is why so many third-party products run DeepSeek models with their own search.
Is DeepSeek available in the UK?
Yes. There is no public ban; the UK government said using it is a personal choice. On the UK App Store the official app appeared as 'DSeek - Your AI Assistant' when we checked. DeepSeek stores personal data in mainland China.
How do I get my brand cited by DeepSeek?
Rank well organically in the engines of your target markets, because DeepSeek's search draws on third-party engines, and write pages whose key facts stand alone in single sentences, which is how DeepSeek's prompt tells it to cite.
Want a free chat about DeepSeek?
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. DeepSeek and 深度求索 are trademarks claimed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd. or its affiliates. Coaley Peak is independent: we are not associated with, endorsed by or a partner of Hangzhou DeepSeek Artificial Intelligence Co., Ltd. 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_deepseek_v1
Last modified: 7 October 2026
GEO Research Centre·Engine guides·DeepSeek