100 things you must know to master deep research | Build your deep research personal knowledge system in one go

• Deep research is likely (probable, 60%-80% probability) to be the second revolutionary AI technology after chatgpt , because it truly changes the productivity of intellectual workers ;
• Your attitude towards deep research essentially depends on your judgment of the nature of technology , your sensitivity and insight into technology;
• My current confidence in "deep research is a revolutionary technology" is 80%. If contrary evidence appears, I will naturally update my inference using the Bayesian method. Before contrary evidence appears, I suggest that you pay enough attention to this technology. If you can spend 100 hours on it, don't just spend 99 hours .
• The more comprehensive your personal knowledge system is about deep research, which has changed the productivity landscape , the deeper your understanding will be, and the better you will be at using it. Through this article, I have written down all the knowledge and experience I have about deep research in my mind, hoping to help you understand more comprehensively and deeply what deep research is , why it is important, how to use it, and how it can change your life;

why

• 2025 is the year of agent. According to OpenAI's AGI classification framework , chatbot/knowledge model is L1 AGI, reasoning model/reasoning model is L2 AGI, and agent/task model is L3 AGI;

• This year, OpenAI has launched three agents : ChatGPT Tasks, Operator, and Deep Research . Tasks are basic task reminders, Operator is to directly operate the browser and interact with the real world, and Deep Research is to help professionals do research;
• Almost no one is interested in the first two. However, for deep research, if you are not doing manual labor , there is really no reason not to pay attention to it, unless you have a stable job , and it is the most stable one?;

what

• Deep research is a function, a model, and an AI agent that specializes in research . OpenAI’s official slogan is: your personal research assistant.
• The first to launch deep research capabilities was Google Gemini, released in  mid- December 2024 ; followed by OpenAI (February 2025), followed by Perplexity;
• Grok3 launched  deep search (AI search, closer to the inference model version of chatgpt search), not deep research (essentially AI agent);
• The underlying model of Google Deep Research  is Gemini 1.5 Pro, whose full name is "Gemini 1.5 Pro with Deep Research". Although Google has accumulated nearly 30 years of technology in search, the basic performance of the model is poor, resulting in an order of magnitude difference between the output results and OpenAI Deep Research;
• Google Deep Research will automatically develop a multi-step research plan after the user asks a question . The system first generates a step-by-step research plan for the user to review, which can be modified as needed, and then the user clicks "Start Research" to begin execution. After obtaining authorization, Gemini will repeatedly perform  the "search-read-analyze" cycle like a human researcher : using Google Search to find relevant content, reading web pages to obtain information, and then adjusting search strategies based on new discoveries. This process will last for several minutes, during which Gemini continues to improve its understanding of the topic and initiates multiple rounds of retrieval and reasoning to ensure that all aspects of the topic are covered. The entire browsing and thinking process is completed automatically in the background without user intervention. Basically, this is the process for deep research from various companies.
• The characteristic of Google Deep Research is that it searches hard (but the results are not miraculous): it usually browses hundreds of websites, reads thousands of web pages, and provides research reports with hundreds of references. It is very impressive, but it is impressive without knowing it. However, after the release of OpenAI Deep Research, this method no longer works, and it cannot withstand the sharp contrast in content quality;
• The deep research launched by perplexity is basically to take advantage of the popularity and seek attention, which is equivalent to "selling dog meat under the guise of sheep meat". The quality is so poor that no one uses it. It is not recommended to waste time .
• Google AI membership is $20 per month, supports 6-person family sharing, and has unlimited deep research; Perplexity AI membership is $20 per month, also unlimited; OpenAI Plus membership is $20, with a limit of 10 times per month. The price difference is due to the huge cost difference on the one hand, and the quality difference on the other hand.
• Deep research is essentially doing one thing: topic research . This is something that is relevant to undergraduates and a key skill for graduate students, but most people can't do it well. The strength that AI has demonstrated in this matter has shocked and impressed human users, and made them think about the value of their own intellectual activities (including metaphysical meaning and value, and even more so, economic value) (human intellectual activities are of course valuable, but the positioning of their own intellectual activities must be readjusted. If they can only achieve what AI can do, that would be dangerous);
• What determines the effect of deep research is not the steps, methods, processes and techniques of "how to do research" (the system instructions of the model are very simple, and O3's basic intelligence is very strong, so you don't need to teach it how to do research), but the underlying reasoning model . This is why the deep research driven by O3 is amazing;

OpenAI Deep Research

• OpenAI Deep Research was launched on February 3, 2025, and was initially available exclusively to Pro users who paid $200 per month. After three weeks of exclusive use by Pro users, OpenAI expanded this feature to Plus users, and will later expand it to Free users .
• After Deep Research was released, I thought about it for 5 minutes and then upgraded to a $200 pro membership because it was exclusive to pro. Today, after using it for nearly a month, I can say this: unless my child has no money to eat, I will always use pro (I will also use pro if I have no money to eat);
• The difficulty of using ChatGPT Pro is not low . It is not a matter of $200 (people who are reluctant to spend money on "software" may be generous in spending money on other things), but a matter of perspective (is it software? Is it a tool? Or is it an intelligent entity that you integrate into your workflow and creates value for you, even monetary value?), and network technology issues (OpenAI unscrupulously reduces intelligence);
• To use Pro, you need a high level of  agency : you must have the ability and willingness to compete with OpenAI and never compromise, and you must be open-minded enough to try out the new work and learning experience in the intelligent era for at least a month;
• The monthly limit for pro users  is 120 times, for plus users it is 10 times, and for free users it is 2 times (tentative estimate);
• The deep research functions used by pro, plus and free users are the same, with only the number of times limited being different, and no performance difference ;
• The monthly limit is not divided by the calendar month, but by the specific date when you upgrade your membership;
• The underlying model of OpenAI Deep Research is the yet-to-be-released full-blooded version of O3, a special version fine-tuned for Internet browsing;
• My summary of deep research is: o3  reasoning model + search  = deep research magic. Unlike conventional AI search, deep research focuses on depth rather than timeliness. When you put forward a research requirement, deep research will help you turn the Internet upside down , thoroughly understand these materials, and customize a research report for you.
• Tips: For quick questions and quick answers , use AI search; for systematic research, use deep research;
• I divide AI search into three levels. L1 is  gpt-4o+search , which searches for information, integrates multiple information sources, and gives an overall answer; level 2  o3-mini+search , which adds reasoning ability. If the problem requires thinking and a multi-step reasoning process, and finally gives a weighed and analyzed result, use o3-mini; level 3  o3+deep research , the difficulty is increased to the research level, and the output result is a research report of tens of thousands of words;
• From a time perspective: if a problem can be solved manually in a few minutes , gpt-4o+search is enough; if a problem can be solved manually in dozens of minutes , use o3-mini+search; if a problem can be solved manually in several hours or even days , use deep research;
• In terms of information sources , Deep Research can access all public web pages, including online pictures, PDFs, and documents, and also supports you to upload your own local pictures and documents;

how

• Since the limit is 20 or 120 times/month, how is it counted as "one time" ? As long as the research progress bar starts to move, the deep research stage is officially started, and it is counted as one research;
• A deep research process is divided into two stages: the demand alignment stage and the formal research stage;
• You put forward your research requirements, and ChatGPT will restate its understanding and ask clarifying questions for unclear, vague, and unexpected areas. You need to confirm them one by one. This " demand alignment phase " may be repeated one or more times.
• Unfortunately , even if we don’t compare the research stage, but only the demand alignment stage, humans often have a gap with AI such as deep research;
• The model you select in the model picker is not the o3 model used in the requirements alignment phase ; in general, it is recommended that you select the o1 model, followed by the gpt-4o model; the actual difference is not big, because the formal research is all based on the o3 model, and there is not much difference in the simple task of understanding requirements;
• You can upload your own materials when making a request : word, PDF, md, pictures, as reference materials for deep research;
• After the formal research begins, it usually takes  5 to 30 minutes, and the length varies according to the difficulty of the task. The longest research I have done so far is to have Deep Research interpret the Munger 100 model in a report, which took 36 minutes and the report was 57,000 words, of amazing quality.
• In the deep research session, there is a  sidebar on the right side of the page similar to CoT , which shows all the information sources of this research and the specific step-by-step research process. It is like you are standing behind a real researcher and watching him do research: think - search - read - think - search...
• Deep research is essentially an agent, and deep research tasks are asynchronous tasks (in contrast, chat is a synchronous activity); after you give instructions, the model starts working, and you can do whatever you need to do. You can close the webpage and exit the app. The research runs remotely on the server and will be pushed to you when it is completed.
• After the report is generated, you can continue to raise new research requirements in the original conversation . The model will keep the memory of the previous research, but it is essentially a new research (it will not modify the original report, but will regenerate a new research report);
• In addition to text, research reports will also include tables, pictures, charts and other forms of content that facilitate understanding. In the future, you will see data visualizations and schematic diagrams automatically generated by AI.

• Deep research cites references, accurate to the "line" . Click on the reference link in the report, and the original web page is actually highlighted by the model. Currently, due to browser limitations, most people cannot see this precise citation;

• Deep research accurately cites this point, which is a killer in occasions such as paper writing ;

• OpenAI officials have repeatedly emphasized that even though the O3 model is very powerful, deep research  may still be an illusion , because the model currently does not actively distinguish the authenticity of information sources on the Internet and is still subject to the "garbage in, garbage out" law;
• However, actively identifying the quality of information sources and the authenticity of information is not difficult for the reasoning model (truth- grounding technology ), and will be solved in the functional iteration;
• Because of this, you may want to limit the language and scope of the model search : only search for English keywords, and only accept English materials. You should do this if the topic is of international interest;
• One major limitation currently is the lack of access to non-public information such as paid resources (databases, academic journals) and private knowledge bases, but this can be solved;
• However, if your research topic only has Chinese materials , you should probably limit your model to search only with Chinese keywords and only take Chinese materials;
• The research report is tens of thousands of words long, and the reading experience directly on the chatgpt page is not the best. You cannot highlight or take notes. The easiest way is to save it to a reader software such as Readwise Reader. Another way is to use the "chatgpt to markdown" chrome plug-in to export it as an md document, and then use Typora to convert it to any format, such as epub that can be imported into the WeChat Reading app;
• I have done a lot of deep research tests, and my conclusion is that the quality of research reports (content richness, research breadth, information quality, report structure and language clarity, references, etc.) exceeds 99% of human output. Some people surveyed industry experts and their feedback was that it takes at least 10 hours for each report to be done by an expert in person, but this number may be conservative; I think a closer statement is "like having a professional researcher work for you for a week and then write a complete analysis report";
• Compared with traditional knowledge media produced by humans (articles, books, podcasts and videos), the information density, structured degree, quality, richness and personalization of deep research reports are an order of magnitude higher than traditional media. A logical conclusion is: read more deep research reports, as many as possible;
• For example, I did a deep research yesterday and found that the quality of the research on a topic was higher than that of Steven Pinker's book "Reasoning". It's not that Steven Pinker is not good, but he is also a human being. The limitations of human cognition make him unable to see things that only deep research can see;
• A typical application scenario of deep research : generating reading guide reports for books; it does not replace reading the entire book, but it will definitely help you read those good books worth reading 5678 times faster and better;
• From now on, after reading any book worth reading , you must do deep research! For a 200,000-word book, assuming you spend 6 hours reading it (1 hour a day for a week), then do some deep research, spend half a day studying the reading report, organizing logseq notes, and then have multiple rounds of conversations with chatgpt... The effect of 10 hours may be equivalent to dozens or even hundreds of hours before... All in all, your life has been extended a lot;
• My test on book interpretation, using English information sources vs Chinese information sources, is garbage in garbage out in principle, but in terms of results, "a slight difference can lead to a great error". Take  the in-depth research report on "One Hundred Years of Solitude"  as an example, with the same prompt, the same o3 model, and the same research topic, but only one difference: one study specified that all English information sources were used, and the other study specified that all Chinese information sources were used. One of them was unbearable to read;
• Deep research operates at the information integration level of the human cognitive ability pyramid : it does not require creativity, innovation, or creative writing. It is very simple to find valuable information, read it all, and then integrate the massive amount of high-value information into a high-quality research report in a structured, clear and orderly manner.

• This is not a difficult task. From the perspective of the "information integration" requirement, it is reasonable for a graduate student to be able to do it. However, it fully demonstrates how terrifying it will be when AI reaches a certain intelligence level and exerts its computing power advantages such as information acquisition and processing that humans cannot achieve. From this perspective, we can see that as long as deep research exceeds a critical point in the information integration level, the value of human intellectual labor at the same level (which does not reach the creative level) will plummet.
• The system prompt of deep research   is actually very simple, with only two tools: browser and Python. The browser only does three things: search, read, and cite. Python only does data processing and table presentation. However, when the language recording and logical reasoning capabilities of the underlying model are strong enough (such as o3), only minimal cognitive activities (search-read-cite, without explicitly defining complex topic research processes) are required to produce outstanding results that humans cannot achieve;
• To unleash the power of deep research, you need to fully utilize your imagination and your expertise in your field. I found that as your imagination is opened up and first-hand experience is accumulated, the power of deep research becomes more and more significant;
• 19 experts in the field evaluated the OpenAI and Google Deep Research reports: 7 (37%) believed that OpenAI Deep Research had reached the level of "experienced professionals"; 10 (52%) believed that the reports produced by OpenAI Deep Research would take them at least 10 hours to complete;

how good

• Deep research is the new search . Like Google, deep research will change from a noun to a verb: “Have a question? Do some deep research!” The essence behind this is that the best intelligence will become a cheap resource that is readily available to everyone (the essence of the intelligent era);
• Deep research, as a killer application of AI agents, directly provides outputs and directly targets human productivity and value creation activities. It is the most powerful productivity technology that ordinary people should pay strategic attention to and fully master. This is also the first time that OpenAI has used expected economic value and human expert hours to measure new functions in the blog. This change is very telling;
• OpenAI's vision for agents is to create  a "super assistant" that can perform analytical research tasks that only human experts can complete. It is very likely that this will happen by the end of 2025 (don't say I didn't warn you);
• So, do some deep research every day . Read a deep research report or content on the same topic (such as this article) every day.