171 | The Ultimate Battle of Deep Research: (ChatGPT vs Gemini vs Perplexity vs Grok vs YOU)

0:00Hello, and welcome to another live episode of the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage your eye to improve efficiency, grow your business, and advance your career. This is Isar Maitis, your host, and I am really excited to be here today. This is an episode I was planning on doing for a while, meaning since deep research on ChatUPT came out after it came out on Google and a few other tools. And I really wanted to do this, but we had guests every single week. And so I didn't have the bandwidth to actually do another episode in a week. I think two is enough.

0:33And so finally this week, we have an opening and I can do an episode just myself. So the topic of today's podcast is going to be deep research. So multiple deep research tools became available in the last few months. And it's a game changer across multiple aspects of people's day-to-day work and aspects of the business. And with that, I wanted to really dive in and see what the differences are. And as I was working on it, more tool came out. So it we're going to learn today is, first of all, what is deep research or deep search in some of the tools and why is it different than the traditional models that we've used so far, how to use them in the most efficient way, what are the differences between the deep research tools across the different platforms, and what are my recommendations on how to use those in the most effective way in the business.

1:25So we have a lot to talk about. So with that in mind, let's get started and let's explain, first of all, where did this all come from and why does it matter? And then we'll dive into the practical, tactical aspect of this. So in the beginning, when ChatGPT came out, which was the first model we had, there was no way to connect to the internet, meaning it was a standalone system. And all that system knew was what it knew from its training data. And that's it. Basically, you couldn't know what's happening or research anything that wasn't in the original training data, that was obviously a huge limitation.

2:02That went away when the first connection between ChatGPT came initially through the connection with Microsoft browser Bing. And so that was step one. And then later on, it became more of an integrated solution as part of ChatGPT. Some of the other models followed suit with allowing you to actually browse the internet and get access to solutions in real time. And that evolved to all the other tools as well. Right now, as far as I know, the only tool that's still not connected to the internet, and that is crazy to me, is Claude. So I actually like Claude very, very much, but it's still offline. It's not connected in real time to any platform.

2:37But then online research tools started showing up. The first tool that I know that had serious quote unquote deep research is actually none of the very known tools, and it's u.com, Y-O-U.com. U.com had a deep research functionality very early on. I've been using it for about a year. And so way before the recent deep research tools came to be known. And shortly after that, we got Perplexity. So Perplexity was the first tool that really combined, if you want, the search engine capability together with a large language model where you can have a conversation with it, but it also knows how to research the web.

3:14But then came the era of deep research. So what is deep research and why is it different than everything we talked about so far in the first three minutes? Deep research is more of an agentic approach to research. What it actually does is it goes through a significantly larger volume of data and tries to distill and analyze the information in it and create a report. It works like a researcher would. A researcher that either works in a company or at an academia or wherever else, it actually does a very thorough work. And it approaches it in different ways, which we're going to dive into. And so the way it works is it actually looks at the problem that you present to it.

4:00It dissects it to smaller, researchable problems, elements, pieces of information that it thinks it needs to get. in order to do the research properly. And then it goes ahead and researches a large amount of websites. The smallest I've seen is probably, I don't know, 30, 40 websites. The most that I've seen was 500 plus websites that it researched in order to get the answer. So it's very different than even the early days of perplexity, or if you use perplexity or chat UPD just with the regular search, where we'll go and research six websites, eight websites. It's a much different level of research.

4:35So that's kind of like the general difference is that it's more of an agent that will decide on its own more things and more topics it needs to research in order to get a more deeper research and more thorough response. And also the output is usually longer and more detailed. It comes out as a report with a lot of segments in it and headers and subheaders and bullet points, just like a human would write a research report about a specific topic. So that's on a high level introduction to what is it. Now, the different tools are taking different approaches to this across multiple aspects of how these tools work.

5:17So let's start with how you get to these tools. The way you get to these tools, and I'm going to share my screen, and those of you listening at home, don't worry about this. I'm going to share with you exactly what's on the screen. But let's start with, I'll just start with whatever order I'm in right now. So I will share my screen. So the very first one we're looking at is perplexity. Perplexity, as I said, was the first tool that kind of went in that direction. But right now, they have a deep research mode. The way you get to it is if you click on, in a regular conversation, there's a drop done menu right in the middle of the search box that allows you to select the different options.

5:54There's auto, raw search, deep research, reasoning with R1 and reasoning with O3 mini. And if you pick deep research, that's how you get to it on perplexity. So it's a drop done menu in perplexity. The interesting thing in perplexity, if you'll see, if I'll scroll down to a search of already done, what you can see is the next time you get the prompt line, basically after you've done deep research, you can select again, meaning in your new prompt line, you can select again, there's the drop down menu shows up basically every time the prompt line shows up. So you can switch during the conversation from doing deep research to doing other stuff, which is actually very powerful because sometimes your follow-up questions, you do not want it to do deep research.

6:36You just want to follow up questions on the search it's already done. And if you have deep research turned on, then it's going to try to do deep research again, which is not exactly what you mean. And so most of the tools enable you to do that. If we go to DeepSeek, you'll see the same thing. So if I'm in the DeepSeek solution, you can see that deep research. So they call it search instead of deep research. And you have like a toggle button within your search, sorry, within your prompt line, there's a toggle button for the search option. It's a similar thing in new.com. There's just buttons and you can see it says research, GPT4O, smart, genius, and so on.

7:12So In ChatGPT, it's the same thing. Inside the actual prompt line, there's deep research button that you can turn on and off for different steps of your work. And the only place you cannot do it is in Gemini. In Gemini, to pick deep research is when you pick the model in the beginning. So now your entire universe that you're working in for that particular conversation is Gemini 1.5 Pro with deep research. And you're stuck with that. And that I said is somewhat of a disadvantage. On Grok, it's the same like all the other tools where it's a toggle button that you can turn on and off, and they call it deep search instead of deep research, but it's the same exact concept.

7:49So this is how you get into the deep research mode in all of these tools. Now, let's talk a little bit about some other differences on how it works once you start doing it. So you prompt it, just telling it what you're trying to research. So you want to tell it what is the topic. You can go deep. You can go wide, depending on exactly what you're looking for. But the more details you're going to provide on exactly what you want to get, you will get an answer. Now, most of these tools will just start the research. And I'll explain in a minute what are the different processes that they take. But the first, the only two that are a little different are actually Gemini and ChatGPT.

8:27So in Gemini, what it will do, it will think about your problem. It will think about the research it needs to do. it will come up with a research plan and it will show you the research plan and it will ask you, do you want to add or change anything? So you can see what it's planning as far as I'm going to research this, I'm going to do that, I'm going to summarize this, I'm going to look at this. I think these are the right directions to go. And once you go, it goes and you don't really know what it's doing behind the scenes. On ChatGPT, there's a slightly different approach that I actually like even better.

8:56So in ChatGPT deep research, if you give it any task, it will ask you two to five questions. It will ask you two to five questions as a follow-up to clarify what you want to focus on, or if there's specific things you want to know that you didn't mention, or are there specific sources that are more relevant, or whatever it thinks will help it do the job better, which is actually pretty cool. So I actually really did that approach. And so there's a question in the task, which one does it? So Gemini shows you the plan, and then you can go in and change the plan and ChatGPT will ask you some follow-up questions.

9:31I've seen it as little as one and as many as like six or seven, but that's about what it's going to ask you. The other tool is just go. Once you hit the go button, they will just do the research. Now, even in the research itself, there's varying levels of how it actually approaches it. So let's start with the one that I think is different than the other ones, and that's Gemini by Google. Gemini by Google doesn't show you the process as it's doing the process. Now, Now, I don't know if it doesn't actually do a process that is similar to the other ones or not, because they're not exposing it. So from a transparency perspective, it's very hard to know what it's doing.

10:06But I think they're also taking a different approach. And I'll explain why. I think what Google does is they take the brute force of Google. Think about it. They indexed the entire internet. They know what's in every single website on the planet. And so their approach is when they build the plan, basically that phase where they're asked for your approval, they're reviewing lots and lots of websites and they're building the plan based on that. And then they just go and execute the plan versus all the other tools, literally all of them, using more of a reasoning model to do that where they go step by step, evaluating and redirecting and making slightly different changes as the process goes on.

10:49So the reason I think Google doesn't do that, in addition to the fact it's not showing it, is when you use Google to do deep research, you're picking, as I said, from a drop-down menu, which model you're using, and you're using 1.5 Pro, which is not a reasoning model. So I think this is what's happening. And so this is Google different than the other ones. The other ones have varying levels of how, I would say, active the reasoning model is when it comes to making decisions and changing and manipulating the process as it's happening. So there's varying levels of activity in these models as far as how active they are in the reasoning.

11:31The most active and actually the coolest to watch because of that is Grok. And let's look at a simple example in Grok and let's just look at it and I'll show you a history of one of them. So what I did, I needed a benchmark. I wanted to compare all of them on a specific topic and I needed a topic that I can verify the facts, right? So the whole point is to see if they're making information up, how detailed they are, how much information they can actually find. So I needed a topic that I knew everything about. And I started thinking about different ideas on what topic can I research that I know everything about, that I can verify the information it comes back with.

12:07And I started this, like I said, as soon as Chachupiti came out. So that's, I don't know, like a month ago when it came out on Pro. And then it hit me that the one thing I know about everything is me. So I decided to do the research on myself. And so the prompt that I use, and I'm going to read it to you, and then we're going to look at what each and every one of the tools did and so on. So the prompt I wrote is, what can you tell me about Isar Maitis from multiply.ai? Find everything you can about him from the time he was born until today. I want to know everything, hobbies, education, professional background, skills, online communities, etc.

12:43So that was the prompt. I gave the same exact prompt to Perplexity, DeepSeek, U.com, ChatUPT, Gemini, and Grok. So six different tools. And I gave them the task, the same exact task to see what they were doing. And I want to start with Grok just to show you the reasoning aspect of it and see how cool that is. And I'm going to share my screen again. But again, don't worry, I'm going to read everything that is in there. So here you see the prompt in Grok and you see what it's doing. and it's starting by thinking and it's thinking and then it says, okay, researching Isar Matis, the request to gather comprehensive detailed about Isar Matis from Multiply AI covering his life, hobbies, education, and careers.

13:20I'm considering starting with a web search to find general info about him using his name and company. So that's what it's starting with. And then it's looking for 10 results found. It found my LinkedIn profile. It found the about us page of my website, Multiply.ai. It found my podcast, the Leveraging AI podcast. It found a generative AI business transformation course, which is the course I've been teaching for two years. And it found Venvelo, which is a venture capital that I'm in, plus five more sources that I can see that provide general information about me. Then it's browsing the results. And then it's starting to find information.

13:54And then it continues to think what to do next. So then under multiply, it found my bio. So then it researched the bio. And then it searched Venvelo for that. And then it found additional websites. And then from that, it browses more results and it gets more and more ideas on what to look for. It found a different website about me called the eTribe, which is my old website. And it found my biography, but it found that that website is not live anymore, which is true. So it didn't know what to do with that. And then it found Crunchbase and many more websites. So it's going through the process as if it's a human person, like it's going step by step.

14:26And as it found another lead, it follows that lead and follows those steps. And then it keeps on going in all these different directions before it actually writes the report. So that's what I'm saying. It's very active when it's doing what it's doing. If you look at ChatGPT as a different example to see what ChatGPT is doing, you will be able to see, first of all, in this case, it only asked me one question. So the question was, let me know which areas are most important to you. I would like to gather all, I'll gather all available information. And so in this particular case, I didn't tell it exactly what I want, but it researches the process for seven minutes.

15:02And you can see what it's doing here is I'm putting together info on StarMate. It's examining personal history, education, career, blah, blah, blah, blah, blah. And then it's searching and then it's collecting data. And then again, it's starting to find websites and it's tracing path to my being an F-16 pilot and so on. And then Rollins College, but it's not as active as making decisions on more and more steps. It's kind of doing the same thing. The third one that is doing a similar thing and showing the work is perplexity. So if we go to perplexity, we'll be able to kind of see the same stuff as far as how perplexity is doing its work.

15:35It's going to show you the sources, but it's showing you less of what it's doing. It's just showing you general high-level steps. It might be doing similar stuff in the background. We just don't know that. So for each and every one of the steps, you can see I found some initial information about Isar Maitis, including his professional role at Multiply.ai and a few details regarding educational background, blah, blah, blah, blah, blah, blah. So significantly less active as far as what the agent part of it is doing. So you can see different approaches to the problem. But the reality is we don't really care about the approach as long as it gives us the best output, right?

16:10So which one gives you the best output? So even in the output, they have very different approaches. The most unique one that again, I find really cool is actually Grok. So what Grok does is if you go down to the report, it always starts, all of them, by the way, tell you how many websites they researched. So if you look right now, we can see that Grok researched 92 web pages. Think about how long it will take you to research 92 web pages to reach a conclusion. Gemini research, it's going to be on the bottom on Gemini, so each and every one is a little different, only 21 websites. But by the the largest number of websites I've ever seen was Gemini.

16:46It was like I said, over 500. On ChatGPT, 20 resources. On you.com, I don't know if they say, I don't think they say in you.com how many websites. They do not. But you get the point. It's going to research many different websites and it's going to give us an answer. Which one actually gave the best results? So let's go on a few of them and see what they've done. Let's start with the worst one. And I've tried this across multiple things. The worst one consistently performed was DeepSeek, which I was actually surprised. So sometimes it works okay. Sometimes like in my case, it worked very poorly.

17:17It basically gave me one fact and then said something like due to technical issues, the search service is temporarily unavailable. And then it gave me a little more information and that's it. So I didn't found much. I tried this. Sometimes it does work, but in all the cases, it's not as good as the other deep research tools. Proplexity actually did a very good job. So it did a comprehensive report. A SAR made this from military pilot to innovator and entrepreneur. So it gives it a header as if it's providing this as something people would want to consume. Isar Matis, a seasoned entrepreneur and CEO of Multiply AI, has cultivated a multifaceted career spanning from military service, tech startups, and innovation, born in 1973 in Israel, and so on and so forth, go through my F-16 pilot and being at the Air Force Academy and then the different businesses and the exits that I have.

18:04And it gives this executive summary that is relatively short, but really great. Like if you just read that, you kind of know who I am. Grok does the same thing. It starts with key points and it does this for every single search. It does this quick executive summary in bullet point format, which I absolutely love. And so if you're looking for a quick executive summary at the beginning and then more details, the best play to go is going to be Grok. And then Grok gives you more of a breakdown. The weird thing about Grok is after it gives you the breakdown at the bottom, every single time, it gives survey notes, which is basically giving you more details.

18:36And it's almost one-to-one repeating what was in the more detailed research. So I don't understand why it's repeating it twice. And again, it's doing it every single time. But the fact that there's three different levels of depth you can go into is relatively cool. And I find this, for my usage, very helpful. Again, in 99 % of cases, I don't research myself. I actually research stuff that I want to find information about. So let's look at perplexity again. So it started with this opening, and then it goes to early life and education. It talks about my childhood and early education, talks about military training and leadership development, academic pursuits, and then goes to my professional career, early ventures and tech startups.

19:11Each and every one of them has between a paragraph to two paragraphs of information, e-commerce leadership and scaling success, multiply AI and consulting, and then he dives deep into what I'm doing right now, including outcome-centric thinking, evaluation shift, professionalized skills, and so different things that I teach, infinite scalability, different things that I refer to in my courses. So we go deeper into the content that I'm generating right now, mentorship and community building. So he talks about my Friday AI Hangouts and other communities that I'm a partner on. And then he goes to contribution to AI and thought leadership.

19:42He talks about my podcast and educational initiatives, many other things. And then he talks about my personal life and stuff like that and online presence. So a very detailed, great report. I would say two things about perplexity, one positive, one negative. The positive, it's the fastest out of all these tools. So we got used from these tools being able to say something and then in five seconds get an answer. These tools will take anything between three minutes to 20 minutes to give you the final summary. The longest I always see is Gemini taking the longest amount of time to get you the results.

20:12So that's the one thing. The downside is that I see it hallucinating every now and then. And I ran this research on me with perplexity more than once. And at least once it made stuff up. The other one that made stuff up, which actually surprised me because it's been around the longest is you.com. So u.com, if you go to that, to just see its format, you can see similar format, early life and education with significantly less information than perplexity, educational background, professional background with bullet point breakdown, which is actually easier to follow. But they have me as the head of AI at Platica, which I'm not, or I wasn't, co-founder of CEO at People, which I'm not, founder of Prospero Technologies, which I'm not.

20:52And so out of some of the skills, the professional history that it has my current role as CEO at Multiply, and then it made up three other roles. Not good if you're doing research. Now, the reason I wanted to research myself is because I can verify this very quickly. But if this was me researching something else, then it will be very hard for me to go and verify. That being said, I always give you links to the sources. So you can go back and check where the information came from in each and every one of these tools, but it's still going to be a lot of work. My suggestion, always go and check the resources, but I will show you in a few minutes how you can do something even cooler and more impactful when it comes to finding what is actually accurate and what is not.

21:33So what am I referring to here? And I just see a lot of stuff happening in the chat. So I'm going to open to see if you guys have any questions. What model is used by you.com? They don't tell you. So I think what they're doing is they're writing on other models behind the scenes. I know that for a fact, but they've developed their own agentic approach to the deep research part of it. So a lot of it is their own proprietary secret sauce. One huge benefit of you.com, and by the way, you can do the same thing in perplexity if you have the enterprise account, is you can mix existing information that you have in your company, large documents in a high volume of them, and have them be a part of your deep research.

22:12So you can combine the research on proprietary data that you own together with information in the outside world. and you can do it in the paid version of you.com and you can do it. And then they also have a pro version of the search that may not hallucinate as much. I don't know. And you can do it in Perplexity if you have the Enterprise version as well. Enterprise said, Perplexity said in this past week or so that they're going to provide this functionality to what they call personal files. And they didn't exactly explain what that is. So maybe in the next few weeks or days or tomorrow or once we're done with this live, I don't know, they will actually release a version that allows you to do it for anyone, then you don't have to have the enterprise license.

22:54So there's another question from Denny on when answers are not cut and dry, where does it differ as far as opinions? I haven't tried that thoroughly enough. I've tried it, but I don't have a conclusive opinion on which one actually does better in thinking. But I will take that back to what I said in the beginning. Most of these tools allow you to go back and forth from the deep research model to their thinking models. And I will show you in a minute how that can be highly beneficial. But you can start the research to get the dry facts with a deep research kind of model, and then in the same chat, switch to a reasoning model and then ask it for opinions.

23:35And that would be my recommendation for this kind of use case. so let's continue let's look at chat gpt for a minute chat gpt as an overall i think that's the best job right now so it's not the fastest it doesn't have the largest number of sources it doesn't think as much as grok but i think as far as the output from all the searches that i've done and again right now every time i do deep research i go to all the tools just to learn which one is doing what better. The other thing that will help, and I'm opening a parenthesis for a second, is most of these tools show you how they think. That helps you know what you may want to use in your prompt to push them in a specific direction to think about specific things.

24:21So it might be, I will probably give more thought and more depth when I prompt Gemini that is not going to have that thinking process than if I'm going to give it to Grok or ChatGPT as an example that I know do a lot of thinking. So just food for thought on that. But let's look at the report on Chachipiti for a minute. You will see that it's the longest, the most detailed, lots of formatting options that help you see what it's doing. So there's bold sections inside the text on important sections. It's broken down into clear segments. Every segment is significantly longer than in any of the other reports.