Coaley Peak GEO Research Centre Engine guide
Kimi How Moonshot AI's assistant builds its own index, and how to get into it.
Kimi is the assistant from Moonshot AI, the Beijing lab known for very long context windows and open-weight models. Unlike most Chinese assistants, it runs its own search index, documents three crawlers and publishes their IP addresses. This guide covers its timeline, the four layers behind its answers, its crawlers and how to opt out of each.
Independent guide: Coaley Peak is not associated with Moonshot AI (月之暗面). Trademarks
- Owner
- Moonshot AI (月之暗面)
- Launched
- October 2023, in China
- Home market
- China (kimi.com) and international (kimi.ai)
- Languages
- Chinese and English; about 16 interface languages on kimi.ai
- Open weights
- Yes: Kimi K3 and the K2 series, under MIT-style licences
- Crawlers
- KimiBot, Kimi-SearchBot and Kimi-User
Overview
What Kimi is, and who uses it
Moonshot AI (月之暗面) was founded in Beijing in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin, and launched Kimi Chat that October as an assistant that could read 200,000 Chinese characters at once. Long context became its calling card: it tested two million characters from March 2024, and its current model, Kimi K3, handles a million tokens.
Moonshot's models are open-weight. Kimi K2, a trillion-parameter model released in July 2025, and K3, a 2.8-trillion-parameter model released in July 2026, can both be downloaded, under MIT-style licences with conditions for the largest commercial users. In May 2026 TechCrunch reported a funding round of about US$2 billion at a US$20 billion valuation, with annual recurring revenue above US$200 million.
Kimi runs as two services. kimi.com is the Chinese-language service, run by Beijing Moonshot with data stored in China. kimi.ai is the international service, run by a Singapore company, with English as the default and about 16 interface languages. Moonshot's last published user figure is more than 36 million monthly users, in October 2024.
For SEO teams, Kimi is one of the most transparent assistants to come out of China. It builds its own search index with a documented crawler, separates that from training and from user fetches in robots.txt, and gives sources it trusts 'authority badges'. Its search favours government, authoritative media, academic and official financial sources, and filters out aggregators and unverified self-media.
Timeline
Kimi, from launch to today
Moonshot AI is founded in Beijing by Yang Zhilin, Zhou Xinyu and Wu Yuxin.
Moonshot announces Kimi Chat, a closed beta that accepts 200,000 Chinese characters of input.
Moonshot's model enters China's register of filed generative AI services.
Kimi begins testing a two-million-character context window.
Kimi launches 'Explorer Search', combining its language model with real-time retrieval and source tracing.
Moonshot launches the global edition of the Kimi Open Platform API.
Kimi K2, a trillion-parameter mixture of experts with 32 billion active parameters, is released as open weights under a modified MIT licence.
Moonshot releases Kimi K2 Thinking, an open-weight reasoning and agent model.
Kimi K2.5 adds native vision and Agent Swarm.
Moonshot releases and open-sources Kimi K2.6, focused on long coding tasks and Agent Swarm.
TechCrunch reports that Moonshot raised about US$2 billion at a US$20 billion valuation.
Moonshot announces Kimi K3, a 2.8-trillion-parameter open model with a one-million-token context; its full weights follow by 27 July.
Things to know
Things to know about Kimi
Kimi separates training from search in robots.txt. Disallow KimiBot to refuse training and still allow Kimi-SearchBot, which builds Kimi's own search index; blocking Kimi-SearchBot removes a site from Kimi search results.
All three Kimi agents share four published IPv4 addresses in Alibaba Cloud address space (AS45102, according to RIPEstat), so an IP check confirms Kimi but cannot tell the agents apart. The user-agent tokens do that.
Moonshot says its IPs are dynamic and advises against blocking them, which can stop it reading your latest robots.txt.
Kimi's search filters out aggregators, unverified self-media and expired caches, and prioritises government, authoritative media, academic and official financial sources. Sources that pass get 'authority badges'.
For pages that load content dynamically, Kimi says it 'may only retrieve the initial HTML content'. Facts that appear only after client-side rendering may be invisible to it.
Deep Research plans about 74 keywords, finds about 206 URLs, keeps the top 3.2% of content and writes reports citing about 26 sources inline.
Kimi's search API labels every result with an 'authority' level for its source; the scale is not published.
kimi.ai's own robots.txt blocks Baiduspider, YisouSpider, 360Spider and Sogouspider.
Every search-based answer includes source links, according to Kimi's help centre.
The China service is filed with the Cyberspace Administration of China and screens inputs and outputs for illegal and harmful information; the same is not documented for kimi.ai.
The four layers
How Kimi 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 Kimi, and what it means for a brand that wants to be cited.
Layer 1
Pre-training: the base (foundation) model
Kimi K3, announced in July 2026, is a 2.8-trillion-parameter mixture of experts with 104 billion active parameters, native vision and a million-token context; the app also offers K3 Cluster and K2.6. K3 was trained on curated web text, code, maths and knowledge corpora plus vision data, with no token count published; K2 used 15.5 trillion tokens. kimi.ai's privacy policy says public online information may be used for training, and KimiBot is the crawler that collects it. No cut-off is published.
What it means for you: KimiBot is the documented route into Kimi's training data, and K3's open weights carry what it learned into other products. Allowing KimiBot, with substantial, server-rendered pages about your brand, is how you shape what future Kimi models know.
Layer 2
Post-training: tuning and alignment
K2 was tuned with large-scale synthesised data on agentic tool use and reinforcement learning with a self-judging critic. K3 used reinforcement learning across general, agent and coding tasks, distillation from several teacher models, and a generative reward model for agentic work. Safety methods are not published. The China service is filed with the Cyberspace Administration of China and screens content; kimi.ai's rules are not published.
What it means for you: Kimi is tuned for long, agentic research tasks, so it reads and compares many sources before answering. Pages that hold up under comparison, with specific, checkable facts, fare better than vague marketing copy.
Layer 3
Retrieval: live search
Kimi runs its own search index, built by Kimi-SearchBot, and decides for itself when to search; users can set search to Auto or Off. It also calls vertical databases, including finance, academic and World Bank data, and screens sources for relevance, authority and freshness. Pages users paste in are fetched live by Kimi-User, and for dynamic pages Kimi may read only the initial HTML.
What it means for you: This is the layer you can work on most directly. Allow Kimi-SearchBot, serve key content in the initial HTML, and build the kind of authority Kimi's filters reward: coverage by media and institutions, not aggregator listings.
Layer 4
The answer: what people read
Kimi's help centre says all search-based answers include reference source links, and vetted sources carry authority badges for official, academic and media sites. Deep Research writes reports of 10,000 words or more with about 26 inline citations. Agent mode builds slides, documents and spreadsheets, and Agent Swarm runs up to 300 sub-agents.
What it means for you: Authority badges make some citations more prominent than others. Earning recognition as an official or authoritative source in your field, with facts kept consistent across your site, is what turns a citation into a trusted one.
Crawlers
Kimi'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 |
|---|---|---|
| KimiBot Documentation | Crawls content that may be used to train Kimi's foundation models. User agent: Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; KimiBot/1.0; +https://www.kimi.com/policies/kimi-crawlers. | Follows robots.txt, including Disallow and Crawl-delay. |
| Kimi-SearchBot Documentation | Builds Kimi's own search index and analyses pages for relevance. Disallowing it keeps a site out of Kimi search results. | Follows robots.txt, including Disallow and Crawl-delay. |
| Kimi-User Documentation | Fetches pages when a user asks, for example to summarise an article. Moonshot says it is not used for bulk crawling. | Not guaranteed: Moonshot says robots.txt 'may not directly apply' to user-triggered fetches. |
Kimi-SearchBot and Kimi-User use the same user-agent pattern as KimiBot, with their own token and version 1.0. Moonshot asks for crawler problems to be reported to kimibot@moonshot.ai, with the domain and logs.
Crawl sources
Where Kimi crawls from, and how to verify it
Moonshot publishes three JSON IP lists, at kimi.ai/policies/kimibot.json, kimi-searchbot.json and kimi-user.json. When we checked, all three were identical, dated 18 May 2026, and listed four IPv4 addresses: 47.79.253.42, 47.79.253.91, 47.79.225.241 and 47.79.230.254. There is no IPv6.
Moonshot publishes no reverse DNS scheme or network. ARIN lists the addresses as Alibaba Cloud's, and RIPEstat shows them announced by AS45102. Because one set of addresses serves all three agents, an IP match confirms the traffic is Kimi's, and the user-agent token tells you which agent it is.
Moonshot says the addresses are dynamic and recommends robots.txt as the main control, warning that blocking its IPs may stop it reading your latest robots.txt. Use the lists to verify, not to block.
Opting out
How to opt out of Kimi, or limit what it uses
Refuse training, stay searchable
User-agent: KimiBot with Disallow: / signals that your content should not be used for training. It is independent of search visibility, so leave Kimi-SearchBot allowed.
Official pageLeave Kimi search
User-agent: Kimi-SearchBot with Disallow: / keeps your site out of Kimi search results.
Official pageLimit user-triggered fetches
You can disallow Kimi-User, but Moonshot says robots.txt may not apply to user-triggered fetches, and blocking it may reduce your visibility when users ask about your pages.
Official pageSlow it down or limit paths
Kimi supports Crawl-delay and path-level Disallow rules. Each subdomain needs its own robots.txt.
Official pageCorrect inaccurate answers about a person
On kimi.ai, email membership@moonshot.ai to correct or remove generated content with inaccurate information about you; identity checks may apply. For kimi.com, email support@moonshot.cn.
Official pageCopyright complaints
Email support@moonshot.ai with a notice that identifies the work and the infringing material, with your contact details, signature and a good-faith statement.
Official page
Opting out applies going forward. Kimi K2 and K3 open weights already downloaded keep what they learned, and a training opt-out for your own conversations does not remove content already used. Blocking KimiBot costs no search visibility; blocking Kimi-SearchBot does.
How we help
How we get brands cited in Kimi
The Coaley Peak GEO Research Centre researches and implements GEO for SEO agencies and in-house teams, with foreign markets as our speciality.
Crawler set-up for Kimi
We configure robots.txt so Kimi-SearchBot can index you, with KimiBot allowed or refused as you choose, and make sure key content is in the initial HTML Kimi reads.
Kimi visibility audit
We ask Kimi your customers' questions in Chinese and English, on kimi.com and kimi.ai, and record whether you are cited, whether you carry an authority badge, and what it says about you.
Authority Kimi recognises
Kimi favours official, academic and authoritative media sources. We build the coverage and citations that put your brand among the sources its filters keep.
Base model targeting for Kimi's open weights
Kimi K3 is open-weight. We probe it with retrieval switched off to see what it learned about you, and work on the training data future versions draw on.
China market entry
We plan Kimi alongside Baidu, Doubao, DeepSeek, Qwen and Yuanbao, so your China strategy reaches every assistant your buyers use.
More on our research, services and team at the Coaley Peak GEO Research Centre.
FAQ
Kimi GEO: questions
What crawlers does Kimi use?
Three, all documented by Moonshot AI: KimiBot for training data, Kimi-SearchBot for Kimi's own search index, and Kimi-User for pages users ask about. Each has its own robots.txt token.
Can I stop Kimi training on my site but stay in its search?
Yes. Disallow KimiBot and allow Kimi-SearchBot. Moonshot documents the two as independent: blocking KimiBot signals no training, while Kimi-SearchBot controls whether you appear in Kimi search results.
Does Kimi cite its sources?
Yes. Kimi's help centre says all search-based answers include reference source links, and sources it has vetted carry authority badges for official, academic and media sites.
Where does Kimi crawl from?
Moonshot publishes four IPv4 addresses, shared by all three agents, in Alibaba Cloud address space. It says the addresses are dynamic and recommends robots.txt over IP blocking.
Is Kimi available outside China?
Yes. kimi.ai is the international service, run by a Singapore company with English as the default, and the Kimi app is on the UK App Store. kimi.com is the Chinese-language service, with data stored in China.
Are Kimi's models open source?
They are open-weight. K3 uses the Kimi K3 License, which is MIT-style with conditions for the largest commercial users, and the K2 series uses a modified MIT licence.
Want a free chat about Kimi?
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. Kimi is a trademark of Moonshot AI or its affiliates. Coaley Peak is independent: we are not associated with, endorsed by or a partner of Moonshot AI (月之暗面). 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_kimi_v1
Last modified: 7 October 2026