Skip to content
Back to all articles
By TJ SoftUpdated:

Comparison of AI Tools for In-Depth Research – ChatGPT Deep Search, Gemini, Grok, and Perplexity

Porównanie narzędzi AI do głębokiego researchu – ChatGPT Deep Search, Gemini, Grok i Perplexity

A few days ago, OpenAI made a new feature called Deep Search available to users on the more affordable paid ChatGPT plans. The tool operates for 10 to 30 minutes, generating extensive, detailed, and advanced content.

How does this solution compare with other popular advanced research tools, such as Perplexity, Grok, or Gemini, which have also recently introduced free options? What is Deep Search best used for? Which types of prompts work best, and are there ways to achieve a similar result at a lower cost? You will learn all this from the text below.

According to many users, the results delivered by Deep Search are impressive, and for some people (myself included), they genuinely elicit a “wow” effect. Each use of this tool saves me 2 to 4 hours of tedious work, which probably makes me more enthusiastic about it than the average user. Nevertheless, in-depth research is something that can be useful to practically anyone.

Table of contents

Importantly, the links generated by Deep Search lead to reliable sources, and the conclusions presented proved accurate, even though I had only asked for a review of the sources, not an advanced analysis. Although the links and text still need to be checked manually, for some queries—for example, about events in 2025—deep search may provide links to events from earlier years.

Deep Search is not just about searching—it is also a powerful tool for writing valuable, well-styled content. After issuing a simple command, I received a detailed and useful 13,000-word article. Until now, I had not been particularly satisfied with the quality of texts generated by AI based on a single command. Usually, longer interaction was necessary, involving the use of multiple prompts and the separate generation of text fragments, followed by manually assembling them into a coherent whole. Even then, I still had to create 50–70% of the material myself.

I believe that the best results when describing AI tools can be achieved precisely by comparing them directly with other solutions. This approach provides a clear point of reference, enables a quick comparison of features, and helps to better understand how a given tool actually works.

What Is It Like to Use the Deep Search Feature in ChatGPT?

You can find access to the Deep Search feature at the very bottom of the chat window, next to the standard search mode. When you hover your mouse cursor over this button, you will see information about how many searches you have remaining. The Plus and Team plans offer 10 queries per month, while the Pro plan provides access to more than 100 advanced searches.

After asking a question, ChatGPT will ask you to clarify the scope of the research through additional questions. Once you answer these questions, a search process lasting 10 to 30 minutes will begin. As a result, the findings are highly detailed and thoroughly developed. A progress bar will appear on the screen, showing the tool’s current stage of operation.

On the right-hand side, you will find the activity and sources panel—it somewhat resembles the so-called “chain of thought,” known from models such as GPT-4o or GPT-3. This allows you to follow the search process in real time and provides a good overview of the entire process. Especially at the beginning, it is worth taking a look at this section, as it offers an interesting illustration of Deep Search’s capabilities and makes it easier to understand what is happening at any given moment.

You can find more information about the useful features of ChatGPT here. Unlike most texts that describe how the language model itself works, I focus on relatively simple features, such as content search, the ability to add files, etc.

Sometimes, after a search has finished, the window may “freeze.” In such a situation, simply refresh the page. However, if you are concerned about losing the results, a good solution is to open a new thread in a separate browser tab.

Testing the Deep Search Function in Various Language Models: Grok, Gemini, Perplexity

While waiting for the results, I will also discuss in detail other AI-powered tools for in-depth research. To this end, I asked several tools (ChatGPT, Grok, Gemini, and Perplexity) exactly the same question:

“Conduct a review of Polish companies, foundations, prominent individuals, and institutions involved in artificial intelligence.”

The screenshots illustrating the process of using Deep Search in ChatGPT concern a different query — unfortunately, I didn’t think to create them right away with the first question, and I don’t want to waste one of the 10 queries available per month repeating the same prompt.

How AI Models Work in the Context of Research

Different AI models vary in the time they take to analyze data, and in the case of the most popular research tools, a simple relationship can be observed—the longer a model processes a query, the more accurate the results we obtain. However, each has its own specific characteristics and differs in its approach to analyzing information.

Deep Search Feature in Grok – AI on X (formerly Twitter)

One of Grok’s greatest advantages is the Deep Research feature, which allows users to conduct advanced analysis completely free of charge. Users have access to 15–20 queries per day. Compared to its competitors, Grok also stands out for its near-total lack of censorship when generating text, as well as fewer restrictions in the case of images. Additionally, thanks to its integration with Twitter, the model is excellent at searching for information about current events. If we use Twitter, Grok is also useful for summarizing an account’s history and adding additional information about the posts and comments we see.

How does research work in Grok?

The query processing time usually ranges from 20 to 50 seconds, although in some cases it may take several minutes. As a result, we obtain a well-considered text based on an analysis of various sources.

Grok Performance Results

Deep Research in Grok can analyze up to 80 websites, from which the most relevant content is then selected. By default, the generated text is concise and often based on lists or tables, making it easier to quickly absorb the information. Another important advantage is its ability to add paragraphs explaining the subject matter, which addresses a problem known from the GPT-4o model, which tends to present results primarily in list form.

Although Grok generates useful summaries, they require additional verification and refinement. The model does not always demonstrate a complete understanding of the context, which may lead to oversimplifications or incorrect conclusions.

Errors and Shortcomings

One of the noticeable problems is the omission of important companies from research results. An example is Eleven Labs, formally registered in the USA but founded by Poles – Piotr Dąbkowski and Mati Staniszewski. The company emphasizes its Polish roots, and in November 2024 invested 11 million dollars in developing Poland’s AI ecosystem by establishing a small research center in Warsaw. Despite this, Grok did not include it in the analysis.

Another problem is incorrect conclusions. In one case, Grok stated that in 2025, there were “at least 35 AI startups” operating in Poland. However, this figure did not result from an analysis of the available data; it came from an article describing the 35 best AI startups in Poland—it was therefore not the actual number of all companies on the market. In reality, there are approximately 1,000 companies associated with AI in Poland, and approximately 200–400 companies actively engaged in artificial intelligence.

Another drawback is the so-called hallucinations. In one of the tests, when I asked for links to YouTube channels, Grok generated fictional addresses that did not exist. This means that the model may create incorrect information when the data is not readily available.

How Does Grok Handle Estimating the AI Market in Poland?

Asking the question “How many AI-related startups are there in Poland?” made it possible to obtain more realistic estimates—from 50 to 400 companies—and an investment value of PLN 2–4 billion, depending on the criteria adopted in the analysis. In this case, Grok did not provide one specific number as a certainty, but instead presented various assessment criteria and highlighted the uncertainty in the available data.

An interesting case was one of the sources received—Grok provided a map depicting the development of artificial intelligence research in Poland. Unfortunately, it turned out to date from 2018, and the website itself, sztucznainteligencja.org.pl, although excellently designed, stopped being updated in 2020. This shows that despite its excellent ability to search for the latest information, when conducting Deep Research, the model may turn to older data if the user does not specify that they are interested in the latest results.

“Think in Grok” Function

Initially, I assumed that advanced language models such as GPT-4o did not have access to the internet, so it took me a while to discover that Grok’s Think function can be used for research. It turned out to work even better than Deep Search. The model analyzes a smaller number of sources, but examines those it considers valuable in greater detail.

The results were surprising—the answers were accurate, and their content resembled the opinion of an expert familiar with the industry. Grok was able to effectively find and compile valuable opinions from specialists, highlighting, for example, the key challenges associated with the development of AI in Poland. It paid particular attention to the problem of a lack of funding in the later stages of startup development. Companies receive funding for initial research and technology development, but later there is a shortage of funds to implement completed solutions on a market scale, which often makes their commercialization impossible.

Summary

Grok’s greatest advantage in the context of research is its optimal processing time. Depending on the query, the analysis takes anywhere from several seconds to a maximum of 2–3 minutes. This is a much better solution than a standard AI model with internet access, while also enabling active research—the user can ask additional questions and refine the topic as they go. Grok can also be a useful tool for testing effective prompts before conducting more advanced searches in ChatGPT.

An additional advantage is its availability. The limitations system is quite flexible, but in practice it provides free access to 15–20 queries per day. For a free tool, that is a very fair offer. What’s more, you can often achieve better results by using the Think function, which also has internet access and can generate lists of valuable links.

Grok and Other AI Tools (Better Than Gemini, Faster Than Perplexity)

In terms of research efficiency, Grok often delivers results comparable to those generated by Gemini. Its advantage, however, is that depending on the query, it can significantly extend the search time—even up to 5 minutes. This allows it to analyze more information within a single query than Perplexity, which in many cases results in more comprehensive and well-considered answers.

Is Grok an Effective Research Tool?

Grok is a very useful research tool, but it is not free from errors. It can provide valuable information, often in a very clear manner; however, users should exercise caution and verify the sources. Its ability to find expert opinions and analyze them in context is impressive, but it can also lead to inaccuracies, particularly when it comes to specific numerical data or web addresses.

Its greatest advantage is access to up-to-date information and the absence of restrictive censorship, making it an excellent complement to other research tools, especially for topics related to trends and new developments.

How to Use Grok’s Deep Search in Practice?

In my opinion, the most convenient way to think of this tool is as an enhanced browser. Its main advantage is quickly finding links and summarizing them, whereas the text it generates and the conclusions it draws are not the most useful, especially when we compare Grok with other tools. That said, the text itself is not always bad; it also has some value, but there are better tools than this one.

Perplexity Deep Search – more precise than Grok

Various Search Options

Perplexity offers several search modes that allow users to adjust the level of analysis sophistication. Users can choose from Auto, Pro, and Deep Search, the latest and most advanced solution. Additionally, the o3 Mini and Deep Seek R1 models are available, offering alternative approaches to data analysis.

In my view, Deep Research mode works best, especially when using standard prompts. The search process takes longer here than in Grok—usually between two and five minutes—which results in greater accuracy and better-structured responses.

Perplexity’s greatest strength is the conciseness and precision of its information. Right at the beginning of the results, we receive key data, such as the fact that Polish companies have invested PLN 1.8 billion in the development of artificial intelligence and that 75% of large enterprises plan to increase spending in this area. Such specific data is particularly useful because it makes it possible to quickly understand the scale of a given issue without having to search through lengthy reports.

An additional advantage is the ability to use Perplexity for free, although users are limited to three advanced queries per day.

Quality of the Generated Text

Texts generated by Perplexity are written well enough that, in theory, they could be published without major revisions. They do not give the impression of being AI-generated content, which sets this model apart from many competing tools.

Compared to Grok, Perplexity performs better when searching for information about institutions, government programs, and foundations, providing more detailed context and better descriptions. Grok mentioned only the Digital Poland Foundation and the Polonium Foundation, whereas Perplexity also added AI Embassy, AI Poland, and Warsaw.AI to the list.

Company Analysis

In my query, Perplexity prioritizes large companies associated with AI, even if they use it only as part of their products rather than developing it themselves. As a result, companies such as CD Projekt Red, Allegro, and Żabka appear in the results in the context of using artificial intelligence to automate processes.

Unfortunately, the model made an error in this regard. Allegro.pl does indeed use AI on its platform, but in a manner typical of large e-commerce websites. Allegro.pl does not conduct or publish groundbreaking research in this field, and the only materials published by the company concern basic applications of AI in sales.

The error stems from confusing Allegro.pl with the Israeli company Allegro.ai, which is indeed involved in developing tools for professionals working on artificial intelligence. Its best-known product is ClearML – a platform for managing AI models. ChatGPT makes this error as well. Search results often include pages featuring AI books sold on Allegro.pl, which can be mistakenly interpreted as evidence of this company’s activity in the AI sector.

On the other hand, it is positive that Perplexity mentioned Eleven Labs and MiM Solutions, although the results lacked a longer list of companies operating in Poland’s AI market.

Problem Generating Lists

One of the noticeable limitations of Perplexity is the model’s reluctance to present data in list form. Even if we request such a structure, the result can be unpredictable—the standard instruction “use a list” does not always work. Of course, with the right prompt wording, it is possible to force the generation of lists, but this requires experimentation.

The absence of lists is not an accidental feature but rather a result of the model’s internal configuration. However, if the user explicitly requests a numbered list, the model can provide a detailed and well-structured compilation.

One of Perplexity’s unique advantages is its image search capability, which works exceptionally well compared to competing models. Unlike Gemini, which often returns random images, Perplexity focuses on informative visuals, making it a more useful research tool.

Summary

Deep Search in Perplexity works more precisely than its counterpart in Grok. The model analyzes data more effectively and provides more accurate conclusions, although it sometimes makes mistakes, particularly when distinguishing between companies with similar names. In standard searches, it can also provide nonexistent links, although I did not notice this problem while testing Deep Search mode.

Overall, Perplexity “thinks” better than Grok, delivering more useful summaries and better context. If we care about more accurate results and more professional research, it is definitely worth spending those extra minutes on Deep Search in Perplexity.

A significant advantage of Perplexity is its ability to create fairly good-looking, fairly long, and coherent text.

Google Gemini – the fastest and simplest model (April 2025 update: Google has significantly improved this feature)

Speed of operation

Gemini stands out for its lightning-fast result generation, which is both an advantage and a disadvantage. Its speed makes it excellent for testing different prompts and working dynamically with the model. At the same time, however, this results in limited analytical depth in Deep Search mode, which is not as comprehensive as in Perplexity or Grok.

Unlike other models, Gemini does not offer a dedicated button for advanced search. The model independently recognizes whether the user wants a quick summary or more in-depth research. To use more advanced features, you need to select the 2.0 Flash Thinking Experiment with Apps model.

Search scope

Gemini is not limited to searching standard websites. The model can also analyze YouTube and Google Maps, making it a useful tool for searches related to locations and multimedia.

The speed of Deep Search in Gemini results from three factors:

  • instant text generation,

  • direct access to their own databases,

  • a simplified search process.

Analysis Results

Gemini Deep Search provides a quick overview of key companies, individuals, and foundations associated with AI, but unlike Perplexity or Grok, it does not add its own thoughts or interpretations. Like Perplexity, it correctly states the amount of investment in artificial intelligence in Poland.

Its greatest advantage, however, is that it mentions the PLLum project, which was overlooked by other models.

Deep Thinking Function

One of the biggest surprises for me was the Deep Thinking feature, which generates more complex texts and additionally weaves graphics into them. Interestingly, these images serve as links to specific pages, which is a unique solution among AI models.

However, a problem arises when providing the links themselves. If we ask Gemini to generate them, it may provide them, but not all of them work. In most cases, the links are based on real websites, and searching for them on Google reveals related articles or locations. It therefore seems that the model stores link data in its database but does not use it in real time, which can lead to errors in the results.

Summary of the deep search feature in Gemini പ്രവർത്തിക്കുന്നു

Gemini is the fastest model for research, making it useful when we need a quick summary or want to rapidly test different queries. However, its biggest limitation is the superficiality of its analyses—it does not delve into details as Perplexity or Grok do. In contrast, the Deep Thinking feature stands out for its ability to add graphics, which can be useful, although problematic when verifying sources. Gemini follows different principles than Perplexity when selecting graphics.

Google Deep Search – April 2025 Update

First iteration – Gemini presents the research plan and asks whether anything should be modified in it.

Then Gemini searches several hundred pages, analyzes the results, and searches for information again. This process is repeated many times. We receive the result after several to a dozen minutes.

Performance? Very satisfactory. Although there are some drawbacks. The advantages include a very good overview of the entire market. In this regard, it performs better than the other tools in this article (although it should be noted that I am conducting this test later and comparing a newer tool with older ones). In this update, I don’t feel like checking how the other models perform all over again, but I’ll make such a comparison once more someday.

The resulting text is 6,500 words long. However, a significant portion of it consists of lists. In another text, when I asked for a 5,000-word text, Gemini wrote 3,500 words, which is still a good result. Therefore, when it comes to the length of the text produced, Gemini ranks second after ChatGPT and creates longer texts than Perplexity. However, ChatGPT’s Deep Search can generate texts two or three times longer.

Let’s move on to the drawbacks: a significant portion of the text consists of lists, which are not the most useful format when learning about a given topic. I feel that a deeper discussion of some of the most interesting companies is missing here. Especially since Gemini additionally created a table of companies with slightly shorter descriptions. This problem can probably be solved with better prompts, but here I am comparing the approach requiring the least effort. And if someone is looking for lists, this is a good solution.

Bonus: another prompt, this time about automation using AI. Gemini was a bit long-winded, but it managed to provide very useful information. So I see that it is a good idea to ask it to write concisely and provide facts, keywords, abbreviations, etc.

ChatGPT Deep Search – a true game changer

Quality of the Generated Text

With just one query, I received a comprehensive, well-written article that was a pleasure to read. The text not only contained unique information but also had a well-structured introduction that provided an overview of the topic and presented the key contexts. ChatGPT began by describing AI departments and universities, then smoothly transitioned to the topic of startups.

The model also provided information on the largest investments in Polish companies involved in artificial intelligence. It pointed to PLN 120 million invested in Intermedica and PLN 109 million in Tidio, although there was no comprehensive summary of all investments. The data came from 2021–2022, meaning it was not the most up to date, but it still constituted a valuable source of information. Additionally, the results included a map of Polish AI startups, which proved to be particularly useful material.

Content Structure

The article contained a list of 20 companies, divided by industry: medicine and biotechnology, finance and cybersecurity, industry, robotics and IoT, retail, marketing, and customer service. Each company was described in a unique way, without a repetitive pattern, which is a common shortcoming of AI-generated content. ChatGPT was able to identify the key distinguishing features of each company.

For example, in the case of Nomatic, the model focused on its specialization – artificial intelligence for robots involved in product arrangement. If a company stood out financially, ChatGPT highlighted that as well, for example by pointing out that Tidio is one of the best-funded Polish AI companies. As a result, the text did not have a monotonous structure, and each discussion was unique.

Among the more interesting pieces of information was also the mention that Addepto worked on the control algorithms for the famous spherical screen in Las Vegas. Naturally, there was also information about Eleven Labs, which plays a key role in Poland’s AI ecosystem.

Trurl – an underrated Polish language model

One of the more interesting threads was the Trurl model, one of the first Polish language models based on LLaMA LLM, developed in 2023. ChatGPT identified this name as a key term for further research, which turned out to be a very accurate lead.

After manually searching, I came across an article describing Polish language models, available on Medium: Landscape of Polish LLMs. Trurl itself is available for download on GitHub and Hugging Face (it weighs over 60 GB), but I could not find an online version of it. The official Voice Labs website allows users to log in, but attempts to run the model ended in an error.

The lack of reviews suggests that, although Trurl had potential, it did not gain widespread adoption. The creators considered it better than GPT for business applications, but actual interest appears limited—around 1,000–2,000 people download it from GitHub each month.

I mention this because it demonstrates the power of Deep Search in ChatGPT – the model provided me with useful search terms that significantly facilitated further independent research.

A Solid Summary of Events from the Past Few Years

Deep Search in ChatGPT effectively covers the latest events, while also not overlooking key facts from several years ago, which helps explain the context of current trends. An example is the information about IKO Voice Assistant at PKO, which has been operational since 2020, i.e. three years before the global AI boom sparked by ChatGPT.

The later part of the article also included a detailed description of foundations, institutions, and figures associated with AI. A significant advantage was that the model not only listed well-known names but also connected them with the companies and organizations it had described earlier, making the whole piece more coherent and logical.

The article was 3,600 words long, but it is possible to generate even longer content—up to 13,000 words. With such extensive texts, some sections begin to repeat, so the actual new content amounts to approximately 6,000 words.

Weak Points

Despite the high quality of the research, there was no mention of the Pllumm model, which appeared in the results from other tools. The Bielik model was also not mentioned, which, interestingly, was overlooked by all the tested models. To a large extent, however, this was because I did not specify that I was interested exclusively in the latest solutions.

The Snowball Effect in Research

One of the biggest advantages of Deep Search in ChatGPT is that after just one query, I received a fully-fledged article that I could read in its entirety and focus on analyzing. With other tools, such as Perplexity or Grok, I often immediately noticed the lack of key information and had to ask additional questions.

Because the content was well structured, I could read it calmly, highlight the most interesting sections, and focus on further research in selected areas. With other tools, I often immediately felt that something was missing, which meant that instead of reading the whole thing, I moved on to subsequent queries.

Of course, fast writing with Grok, Perplexity, or Gemini also has its advantages—it allows you to adjust subsequent queries on the fly. However, Deep Search in ChatGPT has a huge advantage, because the very first text generation provides more polished and complete material.

ChatGPT vs. Other Models – Time Efficiency

Even if we devote the same amount of time to each model, asking more questions to models that operate faster, the final answer in ChatGPT will still be better. There is no need to repeatedly revise and supplement the results—the very first version is already complete and well-written.

For comparison, using Grok AI in conjunction with Google for 10–20 minutes does not yield results as good as one well-executed Deep Search in ChatGPT. This makes the model not only more effective but also more time-efficient, making it the best tool for advanced research.

Technical problems with deep search by ChatGPT:

We cannot close the tab in which a deep search is running. During one test, I submitted a query and then turned off the computer. I did receive answers, but they were generated by the regular model, not by the deep search function. Fortunately, GPT did not count it as a deep search query.

Summary of ChatGPT’s performance in deep research, main substantive issues

Although the research conducted by ChatGPT is impressive, it does not, in fact, cover the key aspects of Poland’s AI scene. First and foremost, it fails to mention reports on artificial intelligence in Poland. For example, the State of Warsaw AI 2024 report, prepared by the Digital Poland Foundation, significantly expands our knowledge of the development of artificial intelligence in Poland. From this report, we learn, for instance, that there are approximately 1,000 companies related to artificial intelligence in Poland, with 46% of them located in Warsaw. Other important cities include Poznań and Kraków, where the Helios supercomputer is located. It is managed by the Cyfronet AGH center and is listed among the 100 most powerful computers in the world.

ChatGPT did mention the Digital Poland Foundation and the fact that they produce reports, but it did not elaborate on the topic. The subject of reports came up when I later asked the regular model about AI in finance in Poland.

Since reports are not available directly on websites, they usually need to be downloaded as PDFs, and often we first have to provide our own email address, etc. As a result, language models generally do not “think” about specific reports and do not have access to their contents. However, if we ask them to find reports, they are capable of doing so.

Main application of the deep search function

The main use of the deep search function in thorough research is to initially organize the subject, help gain a preliminary understanding of the topic, and create a text that can serve as a practical guide to the steps that need to be taken. In this case, I did not ask it to create a research plan; however, the resulting text can be used in this way. Thanks to it, I can read each section in sequence and then search for aspects that I consider worth expanding upon, which helps avoid the chaos that occurs during conventional research.

Additionally, it is very helpful to search using all available tools (Grok, Perplexity, Gemini); although they are much simpler, they offer “a different point of view.”

Practical Tips for Cheaper Tools

Using Numbered Lists

If you want to obtain as many results as possible, it is worth asking the model to create a long numbered list. You can also specify the number of items it should contain, which will increase the precision of the response.

Creating a Search Plan

Instead of asking for research right away, it is worth first generating a search plan. This can be done using the command:

“Write a research plan, analyze it, search for content online, analyze it, and search again.”

This approach often produces better results than a single query because it allows the model to process information in multiple stages. Moreover, the “List of steps” command is also useful in many other contexts, not just during research.

Strategy First, Analysis Second

One of the most effective methods is to divide the process into two stages. In the first prompt, we ask the model to develop a search strategy, analyze key issues, and identify the best sources. We then use this response as the basis for a subsequent prompt, in which we ask it to actually search for the information.

The “Act Like the Deep Search Model” Command

Although entering this command will not suddenly give us full access to advanced models, it can significantly improve the results. In many cases, especially with models equipped with multi-step reasoning, AI initially creates its own prompt based on our command. As a result, if we tell the model that we want functionality similar to Deep Search, the initial analysis may be more detailed.

Research Automation

Contrary to appearances, the Deep Search function is not as difficult to replicate as it might seem. Various tools enabling advanced information-search automation have existed for a long time. Many users have created them, for example, with the help of Relevance AI, or connected ChatGPT with the Make.com and Zapier platforms.

The simplest form of research automation looks as follows:

  1. ChatGPT generates a list of search terms

  2. The automation tool searches for them online and retrieves content from articles

  3. ChatGPT analyzes the collected data and creates an outline for the text

  4. The AI model generates subsequent sections of the article based on the conclusions

Differences Between Standard Deep Search and an Automated System

Deep Search in ChatGPT operates at a more advanced level because it can assess the quality of sources, determine whether it is worth looking for new information, and establish research objectives, e.g., “Check whether this is true.”

The process works more or less as follows:

A ready-made tool that enables an effect similar to Deep Search is described in this video: YouTube.

This system uses the GPT-4o model to analyze and process information. It can be accessed via an API, which means that we pay for each query—the cost of a single search ranges from PLN 2 to 5. Importantly, there is no need to purchase the full Pro version ($200 per month) to get more queries. By using less expensive models, such as DeepSeek, costs can be reduced even further.

This solution is based on ChatGPT, automation, and the Firefly tool, which transforms websites into a format more accessible to language models.

How to Achieve a Deep Search Effect Without Technical Skills?

  1. Search for terms on Google and save the content of the articles you consider valuable.

  2. Paste the texts into the language model and ask it to analyze them.

  3. Ask another question to obtain new related keywords.

  4. Repeat the search process until you have collected enough data.

  5. Based on the information gathered, create an outline of the article and write the individual sections.

How to Conduct Effective Research?

Using AI to Develop Skills

For most applications of artificial intelligence, the key is for it to support our capabilities rather than replace them. AI should help develop analytical skills, not weaken them.

Deep Search is a tremendous help for people who previously lacked the patience to conduct in-depth analyses. Knowing that thorough research will take two hours instead of six makes it much easier to take on more demanding topics.

Basic Principles of Effective Research

The most important thing is the willingness to understand the subject. This means that the user should be able to ask the right questions and notice relationships between pieces of information. It is also crucial to remember the most important facts.

There are many methods of verifying information, but the foundation is taking notes. Many people do not do this, causing them to waste time or avoid more difficult topics that could be valuable to them.

Benefits of Taking Notes

Creating paper notes helps with remembering and organizing knowledge, as well as making it easier to identify connections. In turn, digital notes are more convenient for subsequent analysis and processing.

The most popular tools for organizing information are:

  • Notion

  • Obsidian

  • Canva (Whiteboards)

In many cases, regular browser bookmarks and a folder for text documents (e.g., on Google Drive) are an effective enough solution.

Artificial intelligence can also help with processing notes. We can upload them to Google’s NotebookLM or analyze them in ChatGPT.

After completing the research, it is worth consolidating the information gathered. The most effective methods are:

  • writing the text independently

  • sharing acquired knowledge with another person

  • an attempt to rephrase the conclusions in one’s mind

Consensus AI – a tool for analyzing academic papers

One of the more interesting research tools is Consensus AI, which specializes in searching for and analyzing scientific papers. If you’re looking for verified research, it’s worth trying as a supplement to standard search methods.

Deep Search and Scientific Research

Creating Scientific Texts with AI

Sam Altman claims that OpenAI’s Deep Search tool can generate PhD-level texts and that a query costing 50 cents can generate content worth 500 dollars.

According to OpenAI, Deep Search outperformed a PhD holder with internet access in the Last Humanity Exam test. However, it is worth noting that this test examined only knowledge and the ability to reason based on existing data, without assessing the actual ability to conduct scientific research.

Practical Test – Does Deep Search Really Create Valuable Analyses?

To test Deep Search’s capabilities, I requested a professor-level text that would analyze the concept of centralized and distributed knowledge according to Hayek in the context of artificial intelligence, also taking other types of knowledge into account.

When asked by ChatGPT for additional guidelines, I replied that the text should be approximately 4,000 words long, have a philosophical and economic character, but be written in a popular science style. As a result, I received 6,200 words of a well-structured text that was not only enjoyable to read but also included interesting examples that stimulated the imagination. One of them was AlphaGo’s famous move 37, often interpreted as an intuitive AI decision.

Will OpenAI’s Deep Search Replace Scientists?

Deep Search is a tremendous improvement in producing research texts, yet it does not even come close to replacing researchers. This tool represents a much greater qualitative leap than the introduction of language models themselves, which, when dealing with more complex topics, often functioned merely as a better internet search engine.

Previously, AI models mainly helped with faster information retrieval and text correction, but their usefulness in scientific analyses was limited. Some researchers developed methods for extracting more value from language models, for example through automated data analysis systems, but this required their own expertise. It should also be remembered that many scientists used AI for research even before the launch of ChatGPT.

Although the text generated by Deep Search looks like a professional academic paper—not only at first glance, but also upon closer reading—it does not cover the topic comprehensively and omits key aspects of the issue.

I noticed some of the omissions by comparing the results with answers generated by other Deep Search tools.

Analysis of the Generated Text

At the beginning of the text, ChatGPT presents the general context of the topic, and then lists various types of knowledge, as I requested. It connects them with the concepts of centralized and distributed knowledge, but does so in a way that is not entirely clear. The structure would be better if the models grouped the types of knowledge according to their belonging to a centralized or distributed system and the degree of their intermediacy between these two approaches.

You can see the full, unedited text here:

Next, ChatGPT analyzes the centralized processing of large amounts of data by AI, citing examples of companies such as Facebook, Amazon, Google, and Uber.

An interesting aspect is the analysis of applications such as Uber or Glovo, where a central algorithm may be unfair to many delivery drivers because it does not take their local knowledge into account.

There is also a reference to Jack Ma and his concept of a digital twin of the economy (Digital Twin).

The model indicates that in China, information about every transaction is immediately transmitted to a central system, which in theory allows the government to make real-time decisions. A quote from Daron Acemoğlu then appears, provocatively asking whether Hayek’s argument for the superiority of the market would still hold if central planners had today’s AI computing power.

This is one of the most important themes in the text, yet ChatGPT does not develop it sufficiently, which is a significant error.

Issues Concerning Applications and Analysis of Decentralization

Next, the model provides examples of how artificial intelligence supports decentralized knowledge gathering. It refers to the concept of the wisdom of the crowd and how AI can leverage it, as well as to systems such as Spotify and Netflix, where user preferences shape recommendation algorithms.

The problem is that ChatGPT’s conclusions are inconsistent. The model presents these examples as opposing centralization, whereas in reality they undermine Hayek’s theses more than they confirm them. This passage would fit better in an analysis of the digital twin and AI centralization.

The model also mentions the geographical decentralization of knowledge, giving the example of farmers in Africa who can take pictures of plants and obtain information about their diseases. In this case, AI is advancing trends initiated by the development of the internet, rather than creating them from scratch.

An interesting thread concerns the collaboration of multiple intelligent AI agents. In most discussions about the future of AI, the vision of a single centralized intelligence dominates, but systems based on multiple cooperating agents can eliminate the problems of central management. ChatGPT gives smart grids as an example, in which each household has its own AI system communicating with others. The issue of people’s motivation is also important, and smart cities are given as an example.

It is also worth paying attention to the problem of the homogenization of knowledge—artificial intelligence may promote a single, useful, but incomplete point of view.

What was missing?

  1. No reference to Mises and the problem of economic calculation, which significantly expanded on the issue addressed by Hayek.

  2. The lack of criticism of Hayek’s two types of knowledge, which emerged even before the development of AI. I found this information by chance in Google Gemini, although the model did not provide it itself—it did, however, point to a graphic from an article critical of Hayek.

  3. Insufficient discussion of hybrid models and how to avoid excessive centralization of decision-making.

  4. Lack of an in-depth analysis of the limits of predictability in epistemology, which Perplexity carried out without difficulty.

  5. No detailed description of the Chinese Digital Twin system and the AI models used for stock market analysis.

  6. No reference to motivation according to Kirzner, apart from a brief mention in the context of cities.

  7. Lack of analysis of predictive sales models and whether AI can accurately predict changes in consumer needs.

  8. No mention of the DIKW model (Data, Information, Knowledge, Wisdom), which could help analyze the structure of knowledge.

This list is quite subjective—omitting one of these elements would not be a problem, but omitting Mises and lacking a deeper analysis of central planning using AI are significant gaps.

Conclusions

Deep Search in ChatGPT produced high-quality text that could pass for a professional academic article. However, compared with Perplexity, it lacked more comprehensive context, and the conclusions in some places were too superficial. Ultimately, although this tool can significantly speed up research, it cannot replace in-depth analysis conducted by experts.