Metaso AI Search is an ad-free answer engine from Shanghai-based Metaso Technology that reads across the web and academic sources, then returns a structured answer with citations at three selectable depths. It suits researchers, students and analysts who need Chinese-language and cross-lingual literature more than a general chatbot.
Used this tool? Rate it
Used this tool? Rate it
Mita AI Search — known to its own users by its Chinese name 秘塔AI搜索 — is an answer engine rather than a link directory. You ask a question in natural language and receive a written, structured answer assembled from sources the system retrieves in real time, with citations attached so you can verify each claim. It is operated from Shanghai and is one of the more established products in the Chinese-language AI search category.
The product's entire positioning fits in the five characters printed under its logo on the homepage: 没有广告，直达结果 — no advertising, straight to the result. That is not a marketing flourish appended to a longer value proposition; it is the whole value proposition, and it is worth taking seriously because it defines what the product refuses to do. Conventional search monetises the results page. Sponsored listings sit above organic results, and the page is engineered to keep you on it. An answer engine that carries no advertising has no reason to delay your departure, which changes the shape of the interface: no sponsored block, no trending sidebar, no recommendation rail. You get an answer and its sources.
The operator is unusually transparent by the standards of this category. The about page states that 上海秘塔网络科技有限公司成立于2018年4月，是人工智能领域的一家新锐科技公司，致力于用算力换人力，让专业场景的生产力百倍提升 — Shanghai Metaso Network Technology Co., Ltd., founded in April 2018, describing itself as an AI company aiming to trade computing power for human effort and raise productivity in professional settings. The same page places the headquarters in the Xuhui district of Shanghai with research centres in Beijing and Chengdu.
Two further identity signals are worth noting because they are verifiable rather than self-asserted. The homepage footer carries the mainland Chinese ICP filing number 沪ICP备18046118号-1, linking to the Ministry of Industry and Information Technology registry — a licensing requirement for sites serving mainland users. And the company states that it developed its own large language model, MetaLLM, which completed China's generative model filing: 公司在AI搜索、AI写作、法律翻译等方向上开展研发与产品落地；自主研发了大语言模型MetaLLM，该模型于2023年通过《生成式大模型备案》（备案号：Shanghai-MetaLLM-20231106）. Search, writing and legal translation are the three stated product directions, which explains the product's noticeably document-oriented character.
It is not a general-purpose chatbot with search bolted on. The interaction model assumes you have a question with a factual answer that exists in documents somewhere, and the system's job is to find those documents and synthesise them. For open-ended creative work or long conversational exchanges, a general assistant fits better.
The most consequential control sits directly under the input box as three buttons: 简洁 (concise), 深入 (in-depth) and 深度研究 (deep research). This is a deliberate design choice that most competitors handle badly. Rather than guessing how much effort your question deserves, the product asks you before it starts. A factual lookup gets a short answer quickly; a literature question gets a long, sectioned report built from many more sources and considerably more processing time. Choosing correctly is the single biggest determinant of whether you find the tool fast or slow.
Beside the input field sits a scope selector defaulting to 全网 — the whole web — which can be narrowed to other source classes. This matters because the appropriate corpus differs sharply by question type. A question about current events wants recent web pages; a question about a mechanism of action wants peer-reviewed literature. The product also exposes an 互动网页 (interactive webpage) option that shapes the output format rather than the input scope.
Every substantive answer arrives with its sources attached. This is the feature that determines whether an answer engine is usable for serious work. A synthesised paragraph without provenance cannot be checked, and cannot be cited in anything that will be reviewed by someone else. With source links, the answer becomes a starting map: you read the synthesis to orient yourself, then open the two or three sources that actually matter to your question.
The homepage feature row shows the product has expanded well past retrieval. It lists 学点啥 (a learning/discovery mode), 视频生成 (video generation, currently flagged new), 幻灯片 (slides), 上传文件 (file upload) and API. The slide generation path is the most practically useful for knowledge work: a research question can go from query to a structured deck without an intermediate copy-and-paste stage. File upload inverts the direction, letting you query your own documents rather than the open web.
The input placeholder tells you that typing a forward slash opens custom skills — reusable instruction templates for recurring query patterns. For anyone who runs structurally similar research repeatedly (competitive scans, literature triage, regulatory checks), this converts a long typed prompt into a short command.
The left navigation carries 专题 (topics), 今天学点啥 (today's learning), 书架 (bookshelf) and 历史记录 (history). The bookshelf and topic collections indicate the product treats research as something accumulated across sessions rather than discarded after each query — closer to a research workspace than a search box.
The API page publishes three endpoint families. Read Webpage takes a URL and, per the documentation, 根据输入的网址，抓取对应网页的全文信息，以markdown格式返回 — fetches the full text of that page and returns it as markdown. Search takes a query and returns live results, and a Q&A endpoint returns a synthesised answer. Parameters include scope, size, page, model and format, with streaming supported. The page also notes 已支持MCP协议 — Model Context Protocol support — which matters for anyone wiring the service into an agent framework rather than calling it directly.
The strongest use case. When you need to understand a topic you have not worked in, the deep research mode produces a sectioned overview with citations, which functions as a reading list ordered by relevance. This is the difference between an afternoon of link-following and twenty minutes of orientation followed by targeted reading.
This is where the product has a genuine structural advantage over Western answer engines. Chinese-language sources — academic databases, industry publications, regulatory documents, local media — are unevenly indexed by tools built primarily for English. If your question concerns the Chinese market, Chinese policy or research published in Chinese, a product built for that corpus retrieves material that English-first tools miss.
Search with citations attached fits the workflow of anyone who must eventually produce a bibliography. The caveat, discussed at length below, is that the terms of service restrict how the output itself may be used — the citations are a route to primary sources, not a substitute for reading them.
Market scans, competitor overviews and regulatory summaries all benefit from a structured answer that names its sources. The slide generation feature closes the loop from question to deliverable.
File upload turns the engine on a private corpus — contracts, reports, papers — combining retrieval quality with material the open web does not contain.
The API exists for developers who want retrieval and answer synthesis inside their own products or agents, particularly where Chinese-language index coverage is the requirement.
There is no barrier to first use. The homepage input box accepts a question immediately; a login prompt exists in the corner but basic search and AI answers work without an account. Account creation becomes relevant for history, saved collections and higher usage allowances.
Set 简洁, 深入 or 深度研究 first. The most common new-user mistake is leaving everything on the default and concluding the tool is either shallow or slow. Concise mode is for lookups. In-depth is the sensible default for a real question. Deep research is for the case where you intend to read the output carefully and follow its sources — it is slower and it earns that time only if the question deserves it.
Leave scope at 全网 for general questions. Narrow it when the question belongs to a specific literature, since a narrower corpus keeps general-web noise out of an academic answer.
The system parses natural language, so keyword stacking discards information. "What are the main mechanisms proposed for X, and where do they disagree?" gives the retriever far more to work with than "X mechanism." State constraints explicitly — a time window, a geography, a source type — because these are used during retrieval rather than merely as filters afterwards.
Treat the synthesis as orientation. Identify the two or three citations that bear directly on your question, and read those. This is the workflow the citation design supports, and it is also the workflow the vendor's own disclaimer requires, since the platform does not warrant that the generated text is correct.
The first answer's value is often that it teaches you the right vocabulary. Follow up using the terminology the sources actually use, which retrieves much more precisely than your initial phrasing did.
Where the outcome needs to be presented, use the slide generation path rather than exporting manually — subject to the usage restrictions in the user agreement, which you should read before building any commercial process on the output.
Deep research on a factual lookup wastes minutes; concise mode on a literature question produces a superficial answer you will then wrongly attribute to the tool's capability. This one habit determines most of the perceived quality difference between users.
Not as a formality. The user agreement is explicit that facts in generated text are not warranted, and numbers and dates are named specifically as unreliable. One click of verification on the load-bearing claim is the minimum defensible practice.
Play to the structural advantage. For questions whose best sources are in English, the comparative benefit over English-first tools narrows considerably.
If you run the same shape of query weekly, encode it once as a skill. The consistency also makes results across weeks comparable in a way that ad-hoc prompting does not.
This is the most important practical caveat in this entire guide, and it is unusual enough to catch experienced users off guard. See the Limitations section.
The privacy policy commits to keeping data only as long as necessary. Export anything you want to keep rather than relying on server-side history.
The privacy policy does not state that inputs are excluded from model development. Where confidentiality matters, do not paste sensitive material.
The user agreement explicitly forbids crawlers and automated download programs. The supported route for programmatic access is the official API, which exists precisely for this purpose.
The best fit. Citation-anchored answers, academic source coverage and a deep research mode map directly onto literature work, particularly where Chinese-language scholarship is relevant.
Anyone who produces written syntheses from many sources under time pressure, and who needs to be able to point at where a claim came from.
The operator's stated product lines include legal translation, and the document-oriented design shows. These are also fields where the verification discipline described above is not optional.
Anyone whose primary working language or primary source corpus is Chinese gets the clearest advantage, since this is the corpus the product was built to index well.
The API and its MCP support serve teams building applications or agents that must retrieve from Chinese sources.
Users who need an assistant for open-ended creative writing or extended conversation — a general chatbot suits better. Teams that need broad commercial redistribution rights over generated output, given the usage restrictions. And anyone whose research is overwhelmingly English-language, where the comparative advantage is smallest.
The primary product is a web application at metaso.cn, and it is a client-rendered single-page application in the strict sense: fetching the page without executing JavaScript returns a shell of under a hundred characters of text. This has a practical consequence for anyone investigating the product — guessed URLs are not informative. Plausible paths such as /price and /about return HTTP 404 outright, while the genuine company page lives at /about-us. Rely on in-app navigation rather than assuming conventional paths.
Mobile access is supported. The homepage offers a 手机端 menu item with a QR code for the mobile app, and the application is distributed through mainland Chinese Android stores under the package name com.metaso, including Tencent's Yingyongbao and the Xiaomi store. Because the core service is a web application, everyday use needs only a modern browser and a connection; retrieval and generation happen on the provider's infrastructure.
One point of transparency about international access: the service is built for and filed to serve mainland China. Users elsewhere should expect the interface and the majority of retrieved sources to be Chinese, and should verify current availability from their own location.
Basic search and AI answers are free and usable without an account — you can open the homepage and ask a question immediately. This is a real free tier rather than a trial, and it is enough to evaluate answer quality on questions you already know the answer to, which is the honest way to assess any answer engine.
Beyond the free tier the product operates a membership and credits system that unlocks higher usage volumes and more advanced capabilities. The specific tiers and prices are not published on the open site; the purchase pages sit behind login. Various third-party write-ups quote figures for membership pricing, daily free credit allowances and per-query API rates, but these sources disagree with each other and cannot be verified against the official pages without an account, so no numbers are reproduced here. Check the current rates after signing in. (To be verified)
Two commercial terms are documented in the user agreement and are worth knowing before you buy. Long-term credits carry a twelve-month validity period, and automatic renewal requires five days' advance notice to the user. The refund provision is narrow: it applies to technical failure that makes the service unusable for more than ten consecutive days or more than twenty days cumulatively. That is a service-availability remedy, not a satisfaction guarantee — assume you cannot get your money back because you did not like the output, and evaluate thoroughly on the free tier first.
The API is publicly documented, and its pricing is likewise not shown on the open documentation page; API keys and the playground require an account.
Western answer engines. Perplexity and similar citation-first search products occupy the same conceptual space and are stronger on English-language sources and international coverage. If your work is primarily in English, the comparison favours them; if it is in Chinese, coverage is the deciding factor and it favours Metaso.
General assistants with browsing. ChatGPT, Claude and Gemini all retrieve from the web now. They are more flexible conversationally and weaker as dedicated research instruments — less consistent about citation discipline and less specialised in academic corpora.
Other Chinese AI search products. Several domestic competitors exist, backed by the major Chinese platforms. Compare them on index coverage in your specific field, on whether answers are consistently citation-anchored, and on whether the results page carries advertising, since that is the axis this product competes on.
Academic databases directly. For systematic reviews and formal literature searches, a proper academic database with reproducible query syntax remains the correct instrument. An answer engine is for orientation and discovery, not for a search you must be able to reproduce exactly.
The user agreement is blunt: 秘塔科技对本平台任何服务所生成内容的准确性、完整性和功能性不做任何保证；其中包含的数字、时间以及各类事实性描述等内容，不保证正确 — no guarantee of accuracy, completeness or functionality, with numbers, dates and factual descriptions specifically named. This is standard for the industry but unusually explicit, and it converts source-checking from good practice into a requirement.
This is the limitation most likely to matter commercially, and it runs opposite to most generative tools. The user agreement states 平台内容仅限于您以个人援引及参考为目的使用；不得对平台全部或部分资料进行再发行、转发、出版、转让、商业或其他开发 — content is for personal citation and reference only, and redistribution, republication, transfer and commercial development are prohibited. Many generative products assign output rights to the user; this one does not. Before embedding results in a client deliverable, a published report or a commercial product, read the agreement and, if the stakes are material, seek clarification from the vendor.
The agreement forbids crawlers and automated download tools, and forbids attempts to access the company's computer systems. Programmatic use belongs on the official API.
The privacy policy does not clearly state whether queries and uploaded content feed model development. Its closest language covers using personal information 为新产品、新服务的研发、已有功能、服务的完善提供数据支撑 — providing data support for R&D of new products and improvement of existing services. That phrasing neither confirms nor excludes training use. Absent an explicit exclusion, treat confidential material as unsuitable for upload.
The product is designed for Chinese users and Chinese sources. Non-Chinese speakers face an interface barrier, and the comparative retrieval advantage disappears for English-only questions.
The privacy policy discloses integrated SDKs from WeChat, Tencent, iFlytek, Umeng and Alipay among others. This is normal for a Chinese consumer application and it does mean data touches multiple processors; the policy links each provider's own terms.
Retention is committed only to the minimum necessary period, with deletion or anonymisation afterwards, apart from network logs kept at least six months by law. Export what you need.
The remedy covers extended technical unavailability, not dissatisfaction. Combined with credits that expire after twelve months, this argues for buying conservatively.
Basic search and AI answers are free and work without signing in — you can ask a question on the homepage immediately. An account unlocks history, saved collections and higher allowances, and paid membership and credits exist above that for heavier use and advanced features.
No advertising is the product's defining claim and its entire homepage slogan. The results interface carries no sponsored block, no trending list and no recommendation rail — the page is the answer and its sources.
They set how much retrieval and synthesis effort the system spends. Concise returns a short direct answer, in-depth returns a fuller treatment, and deep research produces a long structured report drawing on many more sources and taking considerably longer. Choose before asking rather than re-running afterwards.
Trust the citations, not the prose. The operator explicitly disclaims any guarantee of accuracy and singles out numbers, dates and factual statements as not warranted. Use the synthesis to orient yourself and verify load-bearing claims in the linked sources — that is what the citation design is for.
Read the user agreement carefully first, because this product is more restrictive than most. It limits platform content to personal citation and reference and prohibits redistribution, republication, transfer and commercial development. If you plan to use output in client deliverables or products, verify your rights with the vendor rather than assuming the permissive norms of other generative tools.
Yes. Three endpoint families are documented: fetching a page's full text as markdown, live search, and question answering, with parameters for scope, size, pagination, model and output format, plus streaming. Model Context Protocol is supported, which helps when integrating into agent frameworks. Keys and the playground require an account.
Scope is selectable and defaults to the whole web, with narrower source classes available including academic material. Its comparative strength is Chinese-language and cross-lingual coverage; for questions whose best sources are English, an English-first competitor is likely to match or beat it.
The privacy policy does not say. It does not contain an explicit exclusion, and its nearest statement refers to supporting research and product improvement in general terms. Because there is no clear commitment either way, avoid submitting confidential material.
Yes. The homepage provides a QR code entry point for mobile, and the app is published to mainland Chinese Android stores under the package name com.metaso, including Yingyongbao and the Xiaomi store. The web version runs in any modern mobile browser as well.
Shanghai Metaso Network Technology Co., Ltd., founded in April 2018, headquartered in Xuhui, Shanghai, with research centres in Beijing and Chengdu. The claim is checkable rather than merely asserted: the site publishes ICP filing 沪ICP备18046118号-1 linking to the government registry, and the company states its in-house MetaLLM model completed China's generative model filing under registration Shanghai-MetaLLM-20231106.
