Comparing Leading AI Deep Research Tools: ChatGPT, Google, Perplexity, Kompas AI, and Elicit

Introduction

Artificial intelligence has quickly moved from simply answering trivia questions to tackling complex research tasks. In professional and academic settings, AI deep research tools have emerged to help users sift through vast information and produce meaningful insights in minutes. These tools act like virtual research assistants — searching the web or databases, analyzing content, and even compiling reports. This evolution is crucial for individuals and businesses that need to make sense of large amounts of data or stay informed on detailed topics without spending days on manual research.

In this article, we compare five leading AI deep research tools: ChatGPT Deep Research, Google Deep Research, Perplexity Deep Research, Kompas AI, and Elicit. Each of these has a unique approach to assisting with research. We will examine how they perform in terms of depth of research (how thoroughly they find and analyze information), usability (ease of use for both casual and professional users), report readiness (whether they give you a structured, ready-to-use report or just raw answers), continuous research capabilities (how well they allow ongoing, iterative investigation), and each tool’s unique strengths and limitations. The tone here is neutral and journalistic — our goal is an objective look that’s useful whether you’re a general user curious about these tools or a tech industry professional considering them for serious research work.

Tool-by-Tool Breakdown

ChatGPT Deep Research

ChatGPT Deep Research is OpenAI’s newest feature that transforms the popular ChatGPT into a research powerhouse. Unlike the standard ChatGPT, which is a live chat assistant, ChatGPT Deep Research operates as an autonomous research agent. When you provide a complex query, this mode will spend 5 to 30 minutes browsing the web, reading sources, and synthesizing information before returning with an in-depth answer. The output isn’t just a quick paragraph; it compiles a structured report with citations, almost like a briefing a human analyst might write. This makes the findings much more traceable and credible, as you can check the linked sources directly.

Depth of research: ChatGPT Deep Research is optimized for thoroughness and reasoning. It runs on a specialized new model (referred to as OpenAI’s upcoming “o3” model) that’s tuned for complex analysis. In tests, it scored exceptionally well on a challenging expert-level reasoning benchmark, far outpacing previous models. For example, it achieved an unprecedented 26.6% on the rigorous Humanity’s Last Exam test, whereas an earlier GPT-4-based version scored only 3.3%. This suggests that ChatGPT Deep Research can handle very complex, multi-faceted questions with a high level of accuracy and depth. It doesn’t just find facts; it tries to reason through them and connect dots across subjects.

Usability: Using ChatGPT Deep Research feels a bit different from a normal chatbot conversation. You enter your research question or task, then essentially let the AI “go off” and work for several minutes. During that time, it autonomously performs searches and analysis steps that would normally require many manual queries. This is convenient — you don’t need to babysit it — but it also means there’s a waiting period. The interface is still ChatGPT, so it’s familiar to existing users, but the feature is currently only available to certain subscribers (OpenAI Pro at about $200/month, with 100 queries limit) and initially only in the U.S. This hefty price and limited rollout mean it’s geared toward professional users who need heavy-duty research regularly. Once the agent finishes its work, you receive a well-organized response. You can then ask follow-up questions in the same chat if needed, treating the report as a starting point for further discussion. That continuous chat capability is useful for refining results, though the heavy lifting is done in the initial autonomous run.

Report readiness: One of ChatGPT Deep Research’s highlights is that it delivers information in a ready-to-use format. The answer usually comes formatted with an introduction, structured sections or bullet points, and references to sources. It can even accept various input types — OpenAI notes it can take in spreadsheets, PDFs, or images as input to incorporate into its analysis, with the ability to produce tables or charts in its report (features that are being rolled out in updates). In short, it’s aiming to give a comprehensive research report, not just a chat reply. For a user, this means less time editing or reformatting the output for a presentation or document. It’s almost report-ready by design.

Continuous research capabilities: ChatGPT is inherently conversational, so after getting the deep research report, you can continue the session. For instance, you might ask it to clarify a certain section, dive deeper into a sub-topic, or update the report with new data later. However, the Deep Research agent itself does the bulk of its work in one go. It plans and executes a multi-step research process autonomously, rather than requiring you to prompt it step-by-step. This is great for efficiency, but it also means you have less direct interaction during the research process compared to a normal chat. You do have the option to guide it initially (for example, by specifying particular aspects you want investigated), and afterwards you can refine the results through follow-up questions. In practice, it behaves like a very diligent assistant: you give it a task, wait, then get the report, and then you can iterate if needed. It currently does not “track” a research project over multiple days on its own (each deep research task is separate), but you can save the conversation or report and come back to it as context in a new query.

Strengths & limitations: ChatGPT Deep Research’s strength lies in its analytical power and depth. It acts like a top-tier analyst who can comb through the internet and produce a nuanced report, complete with source citations, in one session. It’s particularly targeted at professionals in fields like finance, policy, science, and engineering who need comprehensive insights on complex topics. It’s also versatile enough to handle consumer questions (e.g., comparing products or making a detailed pros-and-cons list) with a high level of detail. Another strength is its ability to integrate varied data types — something not all competitors do.

However, there are notable limitations. OpenAI has acknowledged that, despite the advanced reasoning, the tool can still make mistakes or include misinformation if not carefully validated. Like any large language model, it can “hallucinate” — meaning it might confidently cite a source that doesn’t exactly support the claim or mix up facts. The company advises that user verification remains critical. In other words, you shouldn’t blindly trust the report without checking the key facts or sources, especially for high-stakes decisions. Another limitation is accessibility and cost: at $200/month for limited queries, it’s an expensive option largely restricted to well-funded users or organizations. It’s also initially only accessible via the web interface (mobile and desktop apps support is reportedly coming soon). So while it’s a powerful tool, it’s not yet democratized for everyone. Finally, because the agent takes several minutes to run, it’s not ideal for very time-sensitive quick lookups — it shines when you can afford to let it methodically research a topic in depth.

Google Deep Research

Google Deep Research is Google’s answer to AI-driven research assistance. Integrated into Google’s new Gemini AI assistant (the successor to Google’s Bard and part of Google’s AI ecosystem), Deep Research is an “agentic” feature that conducts multi-step web research on your behalf. This tool became available to Gemini Advanced subscribers around December 2024, marking Google’s entry into this deep research arena alongside OpenAI. Just like ChatGPT’s approach, Google Deep Research aims to save users hours of digging through search results by automating the process and delivering a synthesized report.

Depth of research: Google Deep Research leverages what Google is best known for — search. When you pose a complex question, the Gemini AI will formulate a multi-step research plan and then execute it by iteratively searching the web. It finds information, then uses what it finds to launch new searches, refining its understanding continuously. This mimics how a human researcher might start broadly and then narrow down to specific sub-questions. The depth is quite high in terms of coverage because Google can tap into a vast index of webpages and data. Over a few minutes, it scours numerous sources. The end result is a comprehensive report organized into sections with key findings, and it also lists the sources and related content it used. Google explicitly notes that Gemini’s agent will surface not just well-known sources but also relevant websites or organizations you might not have found on your own, which could enrich the research with diverse perspectives.

In terms of accuracy or reasoning benchmarks, early indications suggest Google’s tool was a bit behind OpenAI’s in pure performance. On the Humanity’s Last Exam benchmark, Google’s deep research (with an earlier Gemini model) reportedly scored around 6.2%, which is much lower than OpenAI’s agent scored. This implies that at launch, Google’s synthesis and reasoning might not have been as strong as ChatGPT’s specialized model on that test. However, it’s worth noting that Google has since launched Gemini 2.0 with improved capabilities, and continuous updates are likely. Also, a reasoning benchmark is just one metric — for many practical questions, Google’s real-world search prowess may still retrieve more up-to-date or varied information. Depth-wise, Google Deep Research is excellent at gathering a broad base of facts (thanks to the search engine backbone), but it may still be honing its complex analytical reasoning compared to OpenAI’s offering.

Usability: For anyone familiar with Google’s interface, Deep Research is designed to be intuitive. It lives within the Gemini AI assistant (accessible via a web app and, in early 2025, also the Gemini mobile app). The process typically starts by the user entering a research query. The AI then proposes a research plan — essentially an outline of how it will tackle the question — which you can review and tweak before it proceeds. This is a nice usability feature: you get some control to make sure the AI is focusing on what you want. Once you approve, you hit “Start research,” and then you can sit back while the agent works through its steps. The interface may show progress or you might just wait a few minutes for the result. The final output is presented in a clean, readable format with headings, bullet points, and source links, similar to a report. One-click export to Google Docs is available, which is extremely handy for users who want to edit the results or integrate them into a larger report or presentation. Because it’s Google, the tool is also integrated with the user’s Google account, meaning you can seamlessly save the output to your Drive, share it, or continue working with it in Docs.

In terms of accessibility, Google Deep Research is part of the Gemini Advanced subscription (approximately $20/month) and was quickly rolled out to users in many countries (over 150 countries, supporting 45+ languages as of the end of 2024). This wide rollout and lower price point (compared to ChatGPT’s Pro tier) make it relatively more accessible to both general users and professionals globally. Using it does not require technical expertise; if you can use Google search or Google Docs, you can use this tool. The guidance in the interface (like the ability to modify the research plan) can be helpful for non-experts to ensure the AI is on the right track. Overall, usability is a strong point — it feels like an extension of the familiar Google search, but with the heavy lifting done for you.

Report readiness: Google’s output is explicitly described as a “comprehensive report … neatly organized with links to the original sources”. Users will see the information divided into understandable sections or questions, making it easy to digest. It also includes a list of sources at the end or throughout, under a “Sources and related content” section. The style is intended to be neutral and informative, offering “helpful, easy-to-read insights” and even a concluding summary of findings. Because the results can be exported directly to Google Docs, the transition from research to report write-up is almost seamless. You might just add a title page or some commentary, and you have a deliverable report. This is a key difference from a typical Google Search results page; instead of ten blue links and snippets that you have to assemble yourself, Google Deep Research gives you something closer to a finished analysis. Of course, a cautious user might still polish the language or double-check certain facts, but the structure and content are largely there. In short, report readiness is high — it’s designed for “read it or share it” immediacy.

Continuous research capabilities: After Google’s agent presents the report, the interaction doesn’t have to end. You can ask follow-up questions or tell Gemini to refine certain parts of the report just by continuing the conversation. Because it is part of the Gemini AI chat, it retains context. For example, if the report has a section about “autonomous vehicle sensor trends” and you want more detail, you could ask, “Can you give more details on LiDAR vs radar from that section?” and it would understand you’re referring to the prior research. This makes the research process feel continuous and interactive, even though the heavy research steps are done in batches. Moreover, since the output can be saved to a Google Doc, you essentially have a living document of your research. You can come back later and ask the AI (in a new session) to update the findings as new information becomes available, or extend the research to related questions. One thing to note is that Google’s agent, like OpenAI’s, performs its thorough web search in a contained session. It is not continuously monitoring the web for you after that. If you want updated info a week later, you’d run it again. But the ease of doing so (and the ability to edit the research plan each time) supports an ongoing research workflow. Also, being a Google product, one could imagine future integration with alerts or new data, but as of now it’s user-initiated continuous research.

Strengths & limitations: Google Deep Research’s major strength is the integration of Google’s search expertise with AI summarization. It can tap into the most up-to-date information on the open web (something OpenAI’s tool might miss if its browsing is limited or if it’s constrained by training data). The reports come with plenty of direct links, which encourages transparency — you can verify or read further by clicking through. It’s also user-friendly for those already in the Google ecosystem, and the price and availability make it a bit more approachable for a wider audience. Another advantage is the support for many languages and regions from early on, reflecting Google’s global reach; non-English content or multilingual queries may be handled better by Google’s tool given their expertise in search localization.

On the flip side, one limitation is that the analytical reasoning quality appears to be still maturing. The comparatively low score on a rigorous reasoning benchmark suggests it might summarize well but could miss some deeper insights or connections that a tool like ChatGPT Deep Research would catch. In practical terms, some users have found that Google’s reports are informative but sometimes lack the nuanced conclusions or creative synthesis that comes from a top-tier language model’s reasoning. It may list facts from various sources without always analyzing conflicts or implications deeply. This is likely to improve as Google refines Gemini (especially with Gemini 2.0 and beyond). Another limitation is that it’s a new service bundled with Gemini Advanced — some users might not even be aware it exists, as it’s not as high-profile as ChatGPT. And while the interface is generally good, the initial need to subscribe and go into a specific AI app (rather than the regular Google search page) could be a barrier for casual users. Lastly, like all AI tools, it’s not immune to errors. It might misinterpret a source or provide an outdated link. Google’s credibility might lead users to trust it blindly, but experts still advise reviewing the sources, since AI summarization can sometimes misquote or cherry-pick information inadvertently.

Perplexity Deep Research

Perplexity AI is an AI-powered answer engine that has gained popularity for its concise answers with cited sources. With Perplexity Deep Research, the platform extends into deeper, report-style outputs. This mode is essentially Perplexity’s version of an autonomous research assistant. When activated, it takes your query and performs a flurry of searches, reading through a wide array of sources, and then synthesizes the information into a detailed report. Impressively, it manages to do this quite fast — typically completing the whole process in under three minutes. The aim is to mirror the thoroughness of a human researcher who might spend hours, but compress that into a couple of minutes of AI work.

Depth of research: Perplexity Deep Research is designed to cover a lot of ground quickly. It executes multiple iterative searches and examines hundreds of results if necessary. As it finds information, it uses that to refine its subsequent searches — an iterative approach similar to what Google and ChatGPT agents do. The difference is that Perplexity has always prioritized direct source citations and factual accuracy, so the Deep Research mode continues that tradition. It pulls in facts from many sources and makes sure to reference them in the final output. One interesting aspect is that Perplexity’s tool is not just limited to plain web search; it also has some coding and data analysis capabilities in other modes. For example, the company has mentioned integrating a Python mode in the past. In the context of Deep Research, this could mean it’s able to run simple analyses or parse data if needed (though the primary function is web research). In terms of accuracy, Perplexity’s team shared that their Deep Research achieved about 20–21% on the extensive Humanity’s Last Exam benchmark, which puts it in a high tier of performance (just a few points shy of OpenAI’s result on that test). This indicates it’s quite capable at reasoning through complex questions. It even surpassed some competitors like an earlier version of Google’s and a tool called DeepSeek on that benchmark. Essentially, Perplexity Deep Research offers a strong balance of breadth and accuracy: it may not have the absolute cutting-edge model that OpenAI’s proprietary system does, but it leverages available AI models very effectively to get solid results.

Usability: A big selling point of Perplexity Deep Research is its accessibility. The feature is available on the Perplexity web interface, and using it is as simple as selecting the “Deep Research” mode from a menu and entering your question. The interface remains clean and straightforward — much like a search engine or Q&A chat. For non-technical users, this is friendly: there’s no need to script a research plan or decide which tools to use; it’s one click and go. It is also available for free (with some limits), aligning with Perplexity’s mission to democratize AI research tools. Specifically, anyone (even without a subscription) can run up to five Deep Research queries per day at no cost. This is great for students, independent researchers, or curious individuals who might not want to pay for a subscription. For power users, Perplexity offers a Pro subscription (around $20/month) which raises the limit to 500 Deep Research queries per day — effectively removing concern about usage caps for most people. Compared to some competitors that are paywalled or limited by region, Perplexity’s approach is very user-friendly and inclusive.

Another aspect of usability is speed. Because it often completes tasks in a couple of minutes, users get near-instant gratification. You don’t have to wait 20–30 minutes as with some heavier tools, unless the question is extremely complex. The results page you get is easy to read, with the answer synthesized in paragraphs and often broken down by subtopic. There’s an option to export the report as a PDF or share it as a link (a “Perplexity Page”). This means if you want to share your research findings with someone else, you can do so with a clean read-only webpage or a PDF document that Perplexity generates for you. That’s a convenient feature for collaboration or presentation. On the whole, Perplexity Deep Research is perhaps the easiest to access for the general public: no special setup, works in your browser, and you likely won’t hit a paywall unless you exceed the free query limit on a heavy day.

Another aspect of usability is speed. Because it often completes tasks in a couple of minutes, users get near-instant gratification. You don’t have to wait 20–30 minutes as with some heavier tools, unless the question is extremely complex. The results page you get is easy to read, with the answer synthesized in paragraphs and often broken down by subtopic. There’s an option to export the report as a PDF or share it as a link. This means if you want to share your research findings with someone else, you can do so with a clean read-only webpage or a PDF document that Perplexity generates for you. That’s a convenient feature for collaboration or presentation. On the whole, Perplexity Deep Research is perhaps the easiest to access for the general public: no special setup, works in your browser, and you likely won’t hit a paywall unless you exceed the free query limit on a heavy day.

Report readiness: The output from Perplexity Deep Research is a well-structured answer that often reads like a mini-report. It will typically have an introduction or overview, followed by key points or findings, sometimes arranged with headings or bullet points for clarity. The tool emphasizes clarity and detail — one source calls the reports “coherent and detailed.” After analyzing sources, Perplexity compiles the information into a clear report that you can directly export as a PDF. Each major statement in the report is usually accompanied by a citation (linked to the source), which upholds transparency. Compared to a normal Perplexity answer (which might be a single paragraph with a few citations), the Deep Research output is more expansive — covering multiple facets of a question. For example, if you asked about “the impact of climate change on global agriculture,” a Deep Research report might come back with sections on crop yield data, economic impacts, regional case studies, and future outlook, each citing different studies or articles. It’s not just a summary; it’s compiled insight. While the report is quite ready-to-use, you might still want to add your own conclusion or tailor it for your specific use-case, but the heavy lifting of gathering and summarizing info is done. The fact that you can get it in PDF form means it’s already in a presentable format. If you need to edit or add to it, you could copy the text into a document editor or, as some users do, ask follow-up questions in Perplexity to get more details on a certain point and then combine those. Overall, the outputs are highly usable and almost publication-ready for internal purposes.

Continuous research capabilities: Perplexity’s platform does support some level of continuous interaction, though it’s not as dialog-focused as ChatGPT or Google’s assistants. After you get a Deep Research report, you can indeed follow up with additional questions. Perplexity has a conversational mode where context from previous questions is retained for a while. However, the Deep Research feature might be seen more as a one-shot process that you invoke for a given query. If you have a slightly different or deeper question, you might trigger another Deep Research task. The interface doesn’t string together multiple deep research tasks into one long session automatically; you would run them one after the other if needed. That said, the iterative nature is built into each Deep Research run — the tool itself iterates internally, so the user doesn’t have to guide it step-by-step. And because it’s quick, doing multiple runs with adjusted queries is not burdensome. Users can also refine their queries or click on sources to explore more manually if they want to dig further. There isn’t a project management interface (as in, you can’t save a “project” with multiple related queries in a folder within Perplexity, at least as of now), so continuous research is more about the user guiding successive queries. The shareable “Perplexity Page” feature does allow continuity in the sense that you have a persistent link to your results which you or others can revisit, but it’s static. In summary, Perplexity supports continuous exploration in a conversational manner, but it may not have the same structured multi-turn planning feature as, say, Google’s research plan step or Kompas’s iterative deepening. It’s straightforward: ask, get report, ask another question if needed. Given its speed and free availability, many users might do iterative research by simply asking multiple specific questions and aggregating the answers.

Strengths & limitations: Accessibility and speed are top strengths for Perplexity Deep Research. It lowers the barrier to entry for advanced research assistance by being free (to try) and fast. This makes it an attractive tool for students, journalists, or independent researchers on a budget. It consistently provides sources for every claim, aligning with its reputation as a reliable answer engine. Another strength is that it’s already multi-platform (accessible via web, and the company mentions plans for iOS, Android, Mac apps soon), which means you can integrate it into your workflow easily. Its performance on complex benchmarks shows it’s not far behind the best, so you’re getting a lot of capability for little cost.

On the downside, since Perplexity uses existing models (likely OpenAI’s or similar, given it cites OpenAI and Anthropic partnerships elsewhere), it may not have a proprietary edge in reasoning that OpenAI’s own agent has. It cleverly orchestrates these models but might occasionally stumble on very complex logical tasks slightly more often than ChatGPT’s dedicated system. Also, the brevity that is a hallmark of Perplexity’s normal answers is expanded in Deep Research, but some users might still find the reports not as extremely exhaustive as if a human spent a day on the task. They are detailed, but there might be cases where you want even more depth or a longer narrative, in which case a tool like Kompas AI or ChatGPT might delve a bit further (albeit at more cost/time). Another limitation is that Perplexity’s knowledge and access are bounded by what it can search on the web. If information is behind paywalls or not online, it won’t magically produce it. And while it’s generally good at not hallucinating (because it sticks to sources), any AI summary can occasionally misinterpret a source. The user should be ready to click the citations and verify important facts, especially if the information is critical. Lastly, as a relatively smaller player compared to OpenAI or Google, Perplexity might not have the same level of user community or awareness, so finding user guides or discussions might be less abundant (though it has an active subreddit). In practice, however, it’s so easy to use that extensive documentation isn’t really needed.

Kompas AI

Kompas AI is a newer entrant specifically focused on deep research and report generation. Unlike the others, Kompas is built from the ground up to feel like a dedicated research platform rather than a chat or Q&A tool. Think of it as an AI-powered research analyst that works with you in a structured way. Its creators describe it as a “multi-step research platform” that coordinates multiple AI agents to gather and synthesize information. The end product is a well-organized long-form report, and the system even provides an interface for refining and editing that report with AI assistance.

Depth of research: Kompas AI emphasizes going beyond surface-level information. When you start a research task, Kompas automatically scans hundreds of sources across the web to pull out deeper insights. It doesn’t stop at the first page of search results; it attempts to cast a wide net and then filter out the noise. In fact, the platform is described as reviewing “hundreds of web pages” while “filtering out irrelevant data”, so that the insights you get are substantive. Under the hood, Kompas deploys multiple specialized AI agents in parallel to handle different aspects of the research. This could mean one agent focuses on statistics, another on news articles, another on academic papers, etc., which are then consolidated. This multi-agent approach is akin to having a research team working together, and it allows Kompas to cover more ground in a given time. The service also reportedly evaluates the reliability of sources as it synthesizes information, aiming to ensure that each point in the final report is backed by credible evidence. This reliability-check mechanism is important for depth, because it’s not just accumulating facts — it’s also weighing them for trustworthiness.

Another aspect of depth is context and analysis. Kompas doesn’t just copy snippets; it performs in-depth analysis to find trends, correlations, and draw conclusions. For example, if tasked with analyzing a market trend, Kompas might identify patterns over time or differences between regions by piecing together data from multiple sources. The iterative nature of Kompas is interactive: after the initial pass, users can instruct the system to “Research Further” on certain subtopics or questions that emerged. This means you can drill down into a particular aspect in multiple rounds, each round adding more depth. Because you can do this multiple times, the depth is really only limited by how much time you want to invest and how much information exists. Kompas is designed to help manage and structure that iterative deepening without losing track of the overall project.

Usability: Kompas AI’s user experience is one of its distinguishing features. Unlike a blank chat box, Kompas presents a structured workflow. When you enter a topic or a question, the system automatically generates a research outline — essentially a plan broken into sections or key questions that it will investigate. This outline is visible to you, giving a clear picture of where the research is headed. It’s a bit like the AI saying, “Here’s how I plan to tackle this — let me know if this looks good.” This can save users from having to brainstorm all subtopics themselves. After you approve or adjust the outline, Kompas begins retrieving information. The results populate into a draft report format within the app, sorted into the relevant sections.

Drilling deeper is straightforward: each section or even specific points may have an option for “Research Further,” where with one click you instruct the AI to gather more detail or additional sources on that point. This is Kompas’s way of enabling continuous, guided research. It feels very intuitive — much like using an outline in a document and then fleshing it out. Because the interface is built around the report, you see the document evolving in front of you.

Once the initial content is in place, Kompas offers AI-assisted editing tools. You can select text and ask the AI to adjust the tone (maybe make it more formal or more simplified), reorganize the content (if you want to restructure the argument or flow), or even translate sections if needed. This built-in “AI Edit” feature means you don’t have to copy-paste into another editor for refinement; you can polish the report right in Kompas. Of course, you can also type your own edits manually at any time — the platform allows full manual control over the text.

For collaboration or later reference, you can save the project and export or share the final report easily. Kompas supports exporting the report in common formats (like Word or PDF). The presence of an internal saving mechanism indicates you can return to your research project later, which is great for long-term projects.

For general and professional users, this approach is very user-friendly once you get the hang of it, because it breaks down the process into clear steps. General users who might be intimidated by an open-ended AI chat are guided by the outline and editing tools. Professionals get fine-grained control over the content and structure. There is a slight learning curve in that Kompas is a new interface — unlike ChatGPT or Google which people might have used analogs of (chat or search), Kompas is its own style of tool. But the interface’s focus on clarity (with sections, numbered steps, etc.) helps flatten this curve. Also worth noting: Kompas AI offers a free tier or trial, which means users can likely test it out without commitment. Being a startup product, Kompas is evolving, but initial reviews highlight that it successfully delivers a more structured, report-ready user experience from the outset, as opposed to the more free-form chat of some competitors.

Report readiness: This is where Kompas really shines. The output is essentially a long-form report, neatly formatted with headings, subheadings, and coherent narrative flow. Because the tool itself is oriented around producing a report, you don’t have to assemble anything — it’s building the document as it gathers info. Each section of the outline becomes a section in the report, with the relevant findings filled in. The writing style aims to be clear and professional, suitable for reading by decision-makers or for inclusion in white papers, etc. Kompas also ensures that each point is backed by references or evidence, often in the form of citations or footnotes within the report. This is crucial for report credibility. The end result is something you could plausibly hand off as a draft of an analysis or incorporate into your work with minimal editing.

An added benefit is the AI editing features mentioned before. If the first draft of the report is good but not exactly how you want it, you can tweak phrasing and tone using AI suggestions. For example, you might say “simplify this paragraph” or “make this section more concise,” and the system will adjust it for you, right in the report. This is more efficient than having to do those edits manually or regenerate content via trial-and-error prompts. By the time you hit export, the content can be very close to final form.

Another notable aspect is that Kompas supports collaborative and diverse use cases of report output. They highlight use cases like academic theses, market research reports, business strategy documents, and even policy briefs. Each of these requires a certain formal structure, and Kompas’s flexible outline approach can cater to that. For example, an academic literature review might have sections for methods, findings, and gaps; a market research report might be structured by market segments or by SWOT analysis sections. Kompas, by letting you refine the outline, can adapt to the structure you need.

Overall, Kompas likely delivers the most “ready-to-publish” style of output among the tools discussed here. You may still need to give it a read-over for accuracy and add any personal insights or confidential info that the AI wouldn’t know, but the heavy drafting work is done in a very organized way.

Continuous research capabilities: Kompas AI was built with continuous, iterative research in mind. The presence of the “Research Further” feature is a testament to that. You can use it to continuously deepen the content. For instance, if your initial question was broad, Kompas might create a section for each subtopic. You can then click “research further” on one subtopic to get more details, which might add sub-sections or expand that part of the report. You can do this multiple times, essentially conducting a multi-round exploration without losing the original context or structure. The tool keeps track of what’s been done and where you are adding more information.

Additionally, because you can save projects, Kompas supports ongoing research over time. You might start a report today, then next week new information comes out or you have new angles to explore — you can return to the saved project, possibly update some sections or add new “research further” prompts, and update the report. It’s like having a living document that an AI can continue to work on whenever you need.

This continuous workflow is somewhat unique. Other tools have conversational continuity, but Kompas has document-centric continuity. It’s very useful for long-term projects (like a thesis or a market analysis that evolves over a quarter). The structured UX means you can always see the big picture of your research, no matter how many iterative deep dives you’ve done.

One limitation to mention is that currently the AI operates when prompted by the user at each step; it’s not autonomously monitoring changes or new data unless asked. So while you can come back and refresh sections, Kompas isn’t, for example, pushing updates to you proactively. But given the ease of use, this is a minor issue — a quick prompt and the AI can update a section with the latest info in minutes.

Strengths & limitations: Kompas AI’s major strengths are its structured approach, continuous refinement, and report-focused output. It stands out by providing an all-in-one workspace for research: you search, outline, write, and edit all in one place. This structured UX is ideal for users who want a clear beginning-middle-end process. It’s particularly powerful for producing long-form content like detailed reports, white papers, or extensive comparative analyses. For example, a business strategist can use Kompas to generate a competitor analysis report that’s sectioned by each competitor, with data and insights under each — and they can refine each section as needed. The integrated editing tools (AI Edit) give it a polished edge; none of the other tools offer built-in post-processing of the text to that extent. Kompas also allows manual edits freely, which means the user and AI truly collaborate on the document.

Another strength is the emphasis on source reliability and evidence-backed content, which is crucial for any serious research use. And despite being powerful, it offers free usage options, lowering the barrier for trial. On a technology level, Kompas leverages both OpenAI and Anthropic models, meaning it can combine strengths (Anthropic’s Claude is known for large context and stable output; OpenAI’s GPT for cutting-edge reasoning). This multi-model backbone likely contributes to its robust performance.

When it comes to limitations, as a newer tool, Kompas AI doesn’t have the same track record or user base as something like ChatGPT or Google, which means it might not have been tested on as wide a variety of queries by the public. Users might encounter occasional quirks or less polish in certain niche areas simply because the tool is younger. Also, the structured approach, while a boon for many tasks, could feel a bit constrained for someone who just wants a quick ad-hoc Q&A. For very simple questions, firing up Kompas and going through an outline might be overkill — a quick chatGPT question might suffice. Thus Kompas is really optimized for when you have a more involved research query.

Another factor to consider is that Kompas’s value is in its structure; if a user doesn’t want to engage with an outline or editing and just hoped for one-shot answers, they might not be using the tool to its full advantage. So it caters to users who want that structure. In terms of data limitations, Kompas, like others, depends on publicly available sources and the quality of the models it uses. If all sources on a topic are thin or biased, the report will reflect that. It tries to mitigate this by pulling from many sources and filtering, but it’s not infallible. Finally, being a startup product, pricing and limits might change as it evolves. Currently having a free option is great, but heavy users might at some point have to consider a paid plan if one is introduced beyond a certain usage (this is speculative, as we only know “free options” exist as per the launch info).

Elicit

Elicit takes a different approach to AI-assisted research. Developed by the nonprofit research organization Ought, Elicit is an AI research assistant geared towards academic literature. Its primary function is to help users find, summarize, and extract insights from scientific papers and studies. In essence, Elicit is like an AI-powered academic search engine combined with a research paper analyzer. If the other tools are like having an all-purpose research assistant, Elicit is like having a very knowledgeable librarian or research intern who specializes in academic papers.

Depth of research: Elicit’s depth is specialized. It shines when your question is something that scholarly articles, studies, or reports have addressed. For example, if you ask “What are the effects of sleep deprivation on memory?”, Elicit will search through a database of published research (it uses sources like Semantic Scholar, which indexes over 100 million papers) and find relevant studies that contain answers. It doesn’t just keyword-match; it uses language models to read through abstracts and sometimes full texts of papers to identify those that likely answer your question. It can then present the answers in a synthesized form along with the references to the papers.

One powerful feature is that Elicit will often give you a table of results, where each row is a paper and columns include things like the paper’s title, a one-sentence summary (abstract), the year, and a direct answer extracted from that paper if available. This allows you to see multiple perspectives or findings from different studies at a glance. It’s a great way to get a quick literature review on a topic. Elicit can also pull specific data from papers if you ask a more focused question (like “What was the sample size in study X?” or “Which paper provides statistics on Y?”). Additionally, it has tools for discovering connections — for instance, you can select some relevant papers and ask Elicit to find other papers similar to those (helpful for broadening a literature search).

However, Elicit’s depth is limited to what exists in academic literature. It doesn’t browse the general web for you. If your research query is more about current events, market stats, or something that experts haven’t published papers on, Elicit might not find much. Also, Elicit often relies on abstracts and summaries of papers rather than full texts for its answers. It has access to many papers, including non-open-access ones, but frequently the tool is pulling from the abstract (since full texts might not be freely available). This means the depth of its answer might be constrained by what is stated in an abstract, which is usually a high-level summary. If deeper details are buried in the paper, Elicit might not capture them unless the paper is open-access or it specifically has been able to parse it. It is continuously improving in this regard, but it’s a known limitation: “Elicit often only has access to the abstract or summary of many papers” which can limit the detail of its answers.

Overall, for academic questions, Elicit provides an impressive depth by aggregating knowledge from multiple studies. It essentially does a mini systematic literature review in seconds. For non-academic questions, its depth is not better than a regular search engine, because it’s not designed for that.

Usability: Elicit’s interface is somewhat like a search engine with a research twist. You enter a question, and it returns a list (or table) of answers/papers. There are features to refine the search: you can add filters (like only show papers after 2015, or only certain fields of study), or you can choose what information you want to extract. Elicit allows you to add columns to the results table for specific things. For example, for each paper result, you can add a column for “sample size” or “outcome measured,” and if Elicit can find that info in the text, it will fill the table accordingly. This is a powerful feature for researchers doing evidence synthesis or data gathering from papers.

For a general user, Elicit might feel a bit more complex than a simple Q&A chat. It’s really designed with researchers in mind — people who don’t mind reading paper titles and skimming summaries. The interface is still user-friendly (it’s web-based and has clean design), but it’s not as conversational. It’s more of a tool or app. If you’re not used to academic literature, the results might feel dense (imagine getting results like “Smith et al. (2019) — Effects of Sleep Deprivation on Working Memory — Summary: Found that 24h sleep deprivation reduced recall by 20% in adults.” It’s useful info, but not a plain-speech answer).

That said, Elicit provides helpful features for usability: you can click on a result to get more details, often including a longer summary of the paper or direct quotes that seem relevant. You can also ask follow-up questions, which essentially runs a new query possibly using context from the previous one. And because Elicit is focused, you’re unlikely to get off-track information — it won’t suddenly talk about something unrelated; it sticks to your research question.

Importantly, Elicit is free to use. It’s been offered as a public tool by Ought and does not have a paywall for its main features. You may need to sign up for an account to use some advanced features or to save your work, but there’s no subscription fee. This makes it accessible for students and researchers anywhere. There is also no restriction by country since it’s just a web app.

Elicit doesn’t require any setup, but to use it effectively, one should have a clear question in mind. The better your research question, the better the results. In library terms, it helps to phrase things like you would for a literature search (though natural language works too).

Report readiness: Unlike the other tools, Elicit does not directly generate a single narrative report or essay in response to your query. Instead, it provides the building blocks for you to create your own report. It’s more of an information-gathering and summarizing tool. If you need a ready-to-use literature review summary, Elicit will give you the key points and references, but you (or another writing tool) would have to stitch them into a cohesive narrative if that’s required.

For example, using Elicit you might end up with a table of 5 papers that each answer your question in slightly different ways. It’s then up to you to write something like, “Multiple studies have examined the effects of sleep deprivation on memory. Smith et al. (2019) found a 20% reduction in recall after 24 hours without sleep, while Jones & Lee (2020) observed impairment in working memory tasks after 36 hours…” and so on, citing those studies. Elicit helps you get those facts quickly and ensures you know which paper said what.

Where Elicit does provide some ready text is in summarizing individual papers (it can generate a summary of a single study) or listing out the direct answers from each paper in a concise form. It even has a feature where you can ask a question and it will try to synthesize an overall answer by combining information from top papers. That synthesized answer is somewhat like what a narrative might be, but it’s not guaranteed to read as smoothly as something like ChatGPT would generate, and you’d still likely want to check each part against the source papers.

So in terms of report readiness, Elicit is more of a research assistant than a report writer. It gets you the facts and citations quickly, reducing the grunt work of literature review. But it doesn’t output a ready report or presentation. People often use Elicit in combination with other tools: for instance, get the data from Elicit and then use a writing assistant (even ChatGPT itself) to help draft the report, or write it manually. Elicit thus accelerates the research phase, and the writing phase remains the user’s responsibility.

Continuous research capabilities: Elicit can be part of a continuous research workflow, but it doesn’t have a memory of a “project” in a narrative sense. You can certainly refine your queries: if the initial question is too broad, you can narrow it down, or you can click on one of the returned papers to ask a more specific question about it. There’s even a feature where you can have a sort of conversation, but it usually manifests as doing another search with context. It’s a tool that assumes the user is inquisitive and somewhat analytical themselves. There’s no hand-holding in writing a conclusion; it just hands you evidence.

One great usability feature: no login is required for basic use. You can go to elicit.org and just start asking. That reduces friction. Another is that results come quite fast (a few seconds usually) because it’s mostly searching a pre-indexed database. It might take a little longer if you ask it to summarize a specific paper’s PDF, but generally it’s speedy. If you ask a new kind of question, sometimes Elicit says “I’m not sure how to help with that” which is at least an honest response and better than giving a wrong answer — but it also means the user has to realize maybe they should phrase differently or that Elicit isn’t the right tool for that query.

Strengths & limitations: Elicit’s key strength is handling academic literature with ease and rigor. It is unparalleled when you need to quickly find scholarly references or get a sense of what the research community has to say about a very specific question. It literally can save days of work in a literature review by automating search and initial summarization. Elicit ensures that the information you get is grounded in published research — it only shows you what actually exists in scientific literature. This means there’s a lower risk of complete hallucination; it won’t make up a source or claim out of thin air because it’s drawing from real papers. If it gives an answer, that answer is traceable to a paper in the results list.

It also has some neat features like being able to extract numeric info or other details into tables, which other tools don’t offer. For example, if you’re comparing statistics from multiple studies, Elicit can help compile those.

The limitations mainly stem from its narrow focus. It doesn’t directly pull in news articles, websites, or other non-academic sources (unless those have been indexed as such in Semantic Scholar, etc.). So it might completely miss information that is important but not in a journal. For instance, if you ask a question about a very new technology that hasn’t been studied academically, Elicit will give you little or nothing (or tangential results). In those cases, a general web research tool would do better. Elicit also might give answers that are too academic or not synthesized into plain language, expecting the user to interpret the results.

Another limitation is the abstract-vs-full-text issue. Because Elicit sometimes only has the abstract, it might not answer detailed sub-questions like “What was the methodology in that study?” if the abstract doesn’t cover it. The user might then have to actually read the paper (which Elicit does facilitate by linking out to PDFs or sources when possible). Elicit doesn’t rank sources by quality beyond some basic signals (it doesn’t yet fully understand which paper is more credible, other than maybe citation counts or publication venues as hints). So, a critical eye is still needed to assess if a study is high-quality or relevant.

In summary, Elicit is superb for scholarly research and ensures a high level of trust in the origin of information, but it’s not a one-stop solution for all research needs, particularly outside academia or for producing a polished write-up.

Head-to-Head Comparison

Now that we’ve looked at each tool individually, let’s compare them directly across the key dimensions of research accuracy/depth, usability, report readiness, continuous research, and the type of user who benefits most from each. This side-by-side comparison will highlight where each tool excels and where they might fall short relative to one another.

Research Depth and Accuracy

All these tools aim to provide deep research, but they do so in different ways and with varying levels of success.

• ChatGPT Deep Research: arguably pushes the envelope on depth and reasoning. With its autonomous browsing and analysis cycle that can run for up to half an hour, it tends to dig very deep into a topic. It uses a cutting-edge model optimized for reasoning, which showed top-tier performance on a broad expert test. In real terms, this means ChatGPT’s agent is good at not only gathering facts but also interpreting them — finding connections, drawing inferences, and even spotting inconsistencies across sources. For example, if two sources disagree, ChatGPT might note that and discuss why. Its depth is such that it can incorporate a variety of data types (text, data tables, images) into its analysis, giving it a multifaceted understanding of a question. The accuracy is high in many cases, but it is also very much dependent on the AI’s judgment, which isn’t perfect. It might sometimes give a very confident answer that later turns out to have a minor factual error or a mis-cited detail. Overall, for breadth and reasoning, ChatGPT Deep Research is leading, but it must be used carefully to ensure that depth doesn’t come with a side of hallucination.
• Google Deep Research: offers depth through breadth of sources. Leveraging Google’s mighty search index, it can usually find information that others might miss, especially up-to-the-minute info or things hidden in corners of the web. It is excellent at covering what is out there on a topic. If there’s a statistic on a government website or a quote in a news article, Google’s agent will likely find it and include it. However, when it comes to accuracy and synthesis, Google’s AI still seems to be catching up to OpenAI’s in pure analytical prowess. It may not reason as deeply about the implications of facts; instead, it might list the facts and let the user draw conclusions. Think of it this way: Google Deep Research might give you 10 well-chosen facts from 10 sources (ensuring no stone is unturned), whereas ChatGPT might give you 7 facts and 3 insights or hypotheses drawn from those facts. Depending on your needs, one approach could be more useful. Accuracy for Google’s tool is generally solid regarding the facts it presents (since it quotes sources directly), but it could be less adept at filtering out subtle misinformation. It might include a point from a less-authoritative source because it can’t reason fully like an expert about which sources are gold and which might be speculative. That said, Google’s continuous refinement during the search does mean it’s checking itself — finding multiple confirmations for info — which helps reliability.
• Perplexity Deep Research: strikes a nice middle ground. It uses a respectable large language model combined with its own search and retrieval pipeline. In terms of depth, it’s quite capable: it will peruse many sources and tends to focus on those most relevant. It might not have the ultra-deep 30-minute dive of ChatGPT, but in practice its under-3-minute research often captures the key facets of a topic. Its benchmark performance around ~20% on an expert test shows it’s not too far behind ChatGPT’s agent in reasoning. One could say Perplexity is highly efficient — it finds a lot of info quickly and usually the accuracy is bolstered by the fact it cites everything clearly. If Perplexity says “Study X found Y,” it will give you the link to Study X right there. This makes it easy for users to verify each piece of information, which in turn keeps the overall accuracy of what you take away high (assuming you cross-check the major points). In head-to-head accuracy, Perplexity might only be a notch below ChatGPT’s exhaustive analysis and perhaps above Google’s more list-oriented approach, but all three are in a relatively high tier. None of them are immune to mistakes though: Perplexity could still mis-summarize a source if the model interpreting it slips up, but the transparency of sources helps mitigate that.
• Kompas AI: is designed for depth through iteration and structure. It may not have public benchmark figures, but qualitatively, Kompas’s approach of multiple AI agents and user-guided deepening suggests it can achieve very deep coverage on complex questions. In a way, Kompas’s depth is potentially unlimited because you can always decide to research further on any subtopic. It systematically organizes information, which means it’s less likely to overlook an angle once that angle is identified in the outline. Accuracy is a focus area for Kompas; it tries to verify reliability before adding content to the report. By filtering out “extraneous material” and focusing on relevant data, it reduces the noise that can sometimes lead to inaccuracies. If Kompas includes a piece of information, it’s probably something that appeared in multiple trustworthy sources or was deemed significant by its analysis. Also, since the user sees the outline early, there’s a chance to spot if something might be missing and prompt the AI to cover it — that collaboration enhances depth and accuracy. In practice, Kompas should be able to handle very complex multi-part questions (e.g., a research report on “The socio-economic impacts of renewable energy adoption in Southeast Asia over the past decade”) by breaking them down and tackling each systematically. The limitation might be raw reasoning or creative inference — Kompas will give you what the sources say and a solid synthesis, but it might not “think outside the box” as much as an OpenAI model could. However, for factual depth, it’s likely on par with the best, and the fact that it backs points with evidence helps maintain accuracy.
• Elicit has a different kind of depth. For academic questions, its depth is remarkable: it essentially reads through potentially hundreds of papers (via their abstracts mainly) and surfaces the core findings. You get the benefit of scholarly peer-reviewed knowledge, which is often the deepest and most verified form of information. If the information you need exists in academic form, Elicit will almost certainly find it and present multiple viewpoints or data points. That is a qualitatively different kind of depth — it’s the difference between “what does the internet say” and “what does the scientific community know”. Accuracy is generally high because you are literally getting the answers from published papers, which have been vetted to some degree. Elicit itself isn’t inventing anything; it’s quoting and summarizing actual studies. The caveat is that sometimes the summaries of papers can be incomplete. Also, if there is conflicting evidence in different studies, Elicit might show you both but won’t reconcile the conflict for you beyond listing them. It’s up to the researcher to interpret that (which is normal in research work). In areas not well-covered by academia, Elicit’s depth is shallow — it might give you irrelevant or no results. For example, a question about a very current event or a niche industry trend will stump Elicit or yield tangential papers. So, Elicit’s accuracy is superb within its domain (because it doesn’t hallucinate beyond showing real papers), but outside that domain it’s not applicable. It never lies, but it might say “no good information found” or give a generic answer if literature is lacking.

In summary, all tools bring strong research capabilities: ChatGPT and Kompas aim for comprehensive multi-angle analysis, Google ensures wide coverage and up-to-date info, Perplexity balances speed and quality with transparency, and Elicit provides scholarly rigor. If we think of “depth” as thorough understanding, ChatGPT and Kompas might take the crown. If we think of “accuracy” as factual correctness with evidence, Elicit and Perplexity have an edge due to their direct sourcing approach. Google is the broad net that may catch the latest or most diverse info, though it might need the user’s interpretation to reach the deepest insights. Each serves a different notion of what deep research means.

Usability and User Experience

How easy are these tools to use, and what is the experience like for a user?

• ChatGPT Deep Research: The user experience here is essentially an extension of using ChatGPT. If you know how to prompt ChatGPT with a question, you can use Deep Research. The difference is you’ll wait longer for an answer. There’s minimal effort in terms of clicking or specifying details (no need to pick sources or outline — the AI does it). This simplicity is great, but it’s also somewhat opaque: the AI goes off and does stuff and comes back later with results. There isn’t much interactivity during the process. For some users, that’s fine (“let it work and get back to me”), for others it might feel odd not knowing what it’s doing in those 10–20 minutes. In terms of interface, since it’s built into ChatGPT, it’s very clean — just a chat box and the eventual answer. One potential usability drawback is managing expectations: a user might type a huge question and not realize they have to wait, or might wonder if it’s still working. Presumably, OpenAI provides some indicator that it’s researching. Another consideration is the cost and access; since it’s limited to certain subscribers, the average person can’t use it at all yet, which is a usability issue in the broad sense. For those who can, it’s straightforward. Professional users might appreciate not having to micro-manage the tool, while some power users might wish they could intervene (like, “no, skip that subtopic, focus on this”) mid-way — which currently isn’t a feature. But overall, if you describe usability as “how hard is it to get a good result?”, ChatGPT Deep Research is easy mode: one prompt, one answer (albeit after some time).
• Google Deep Research: Being part of the Google ecosystem, it benefits from familiarity. The interface has a bit more guidance than ChatGPT’s — you see a proposed plan and have buttons like “approve plan” and “start research”. This adds a tiny bit of complexity but also clarity. Users can actually see the breakdown of tasks, which demystifies what the AI will do. For those who want control, it’s reassuring. For novices, it might be slightly more steps than expected, but it’s explained in simple terms by Google (“Gemini will create a plan… you can add more aspects… then click start”). During the research process, Google’s interface likely shows some loading or a message that it’s gathering info. The final output is shown in a nicely formatted way within the same interface, and with the export button, which is great for usability. One strong point is integration: because you can export to Google Docs with one click, continuing work with the info (editing, adding your own commentary, sharing with a colleague) is frictionless if you already use Google’s productivity tools. Also, multi-language support and widespread availability mean users in non-English contexts can use it in their native language — that’s a huge UX win for inclusivity. The requirement of a Gemini subscription and needing to go to a specific Labs or app might hinder some casual users who just stick to normal Google search. But as these agent features possibly merge into main search in the future, that will become even more seamless. Overall, Google offers a user-friendly, mildly interactive experience that’s quite polished.
• Perplexity Deep Research: Usability is one of Perplexity’s bragging rights. There’s basically no learning curve — if you can use a search engine or type a question, you can use it. The interface is minimalistic. One thing to mention is that switching to Deep Research mode is an extra step (you have to toggle that mode or select it, as opposed to the default quick answer mode). But that’s just a single click in the UI. The speed at which results come is also a user experience factor: waiting just a couple of minutes is relatively painless, and you can actually watch it populate intermediate findings if you stare at the screen (sometimes Perplexity shows a spinner and partial text updating). For a user, getting the answer quickly means they can iterate or move on faster — it feels snappy. The readability of the answers is good; Perplexity usually writes in clear, digestible language, often breaking down answers into bullet points or short paragraphs which are easy to skim (especially since sources are in brackets right next to statements). Since it’s web-based and shareable, if you want someone else’s input, you can just send them the Perplexity page link. On the downside, because it’s so straightforward, it may not guide users through complex queries like Kompas does. If a user asks a very broad question, Perplexity will attempt it in one go, and if the question wasn’t specific enough, the result might be a bit shallow or generic. It’s up to the user to then refine the question and try again. So users still need some skill in asking the right research question. It’s not a hand-holding workflow; it’s more trial-and-error if the first attempt isn’t spot on. But thanks to being free and fast, trying again is no big deal — you won’t feel you “wasted” a costly query.
• Kompas AI: Kompas introduces a slightly different paradigm, so the usability depends on how quickly one adapts to it. Initially, a new user might find it unusual that after they type a topic, they see an outline instead of an immediate answer. But once it’s understood that the outline is the AI’s plan (which you can edit if you want to focus the research), many users will appreciate that clarity. It actually helps in cases where the user’s question is fuzzy — the outline might prompt them to clarify what they want. The multi-step process (outline -> populate -> refine -> edit) is logically organized. It might feel like more work than just asking a question on ChatGPT, but the trade-off is you get a more tailored result. For serious research tasks, users often prefer a bit of upfront structure, so Kompas caters to that mindset. The interface with integrated editing tools is a plus for advanced users; you don’t have to switch contexts to do things like tweak wording. For a casual user, those extra buttons might be overwhelming or unnecessary, but they don’t have to use them. They can just focus on the content. One potential usability challenge is that Kompas likely works best on larger screens (since you might be looking at an outline and a report draft simultaneously). It’s a very productivity-oriented interface. On a mobile phone, for instance, that might not be as easy to navigate, though presumably Kompas can be used on mobile browsers, it might not be as smooth as on a desktop. Another consideration: because Kompas is relatively new, it might have the occasional glitch or slow moment as it’s still being optimized. However, nothing in our info suggests it’s particularly buggy — just something to keep in mind whenever adopting a new tool. Learning to trust the tool’s process is part of usability, and Kompas gives transparency (outline, sources) that foster trust. In summary, Kompas might require a tiny bit more user involvement, but it’s designed to be intuitive in a logical, stepwise way. For users who like structure, that involvement actually enhances the experience. For those who just want a quick answer with no steps, Kompas might feel like overkill for simple queries.
• Elicit: Elicit’s usability is tailored to researchers. If you’re used to Google Scholar or academic databases, Elicit feels like a dream come true because it simplifies many steps (search, gather, summarize all in one). The interface with tables and filters is efficient, but if someone comes expecting a chat or a simple list, they might be taken aback by seeing a spreadsheet-like result. It’s not colorful or full of prose; it’s very information-dense (which is good for the right user). Elicit does attempt to answer in plain language at times (like it might give a direct answer sentence at the top if it can), but often the meat of it is in the list of sources. So for a general user asking, say, “How to improve sleep quality?” Elicit might throw up 10 paper references about sleep studies, which is probably not what they expected or want. That kind of user would prefer a narrative answer from ChatGPT or similar. On the other hand, a grad student asking “What are the known effects of XYZ compound on plant growth?” will get exactly what they need — a set of papers to read or at least cite. Elicit’s UI has some learning curve: you should know that you can scroll right to see more columns, or that you can click to expand details, and know how to interpret the output. It’s a tool that assumes the user is inquisitive and somewhat analytical themselves. There’s no hand-holding in writing a conclusion; it just hands you evidence.

One great usability feature: no login is required for basic use. You can go to elicit.org and just start asking. That reduces friction. Another is that results come quite fast (a few seconds usually) because it’s mostly searching a pre-indexed database. It might take a little longer if you ask it to summarize a specific paper’s PDF, but generally it’s speedy. If you ask a new kind of question, sometimes Elicit says “I’m not sure how to help with that” which is at least an honest response and better than giving a wrong answer — but it also means the user has to realize maybe they should phrase differently or that Elicit isn’t the right tool for that query.

For someone doing a literature review, Elicit’s ability to save the query or export the results (you can copy the table or download it) is useful. It doesn’t automatically integrate with writing tools, but you can take those references and put them in a citation manager easily enough.

Output and Report Readiness

What do you actually get from these tools, and how ready is that output for consumption or use?

• ChatGPT Deep Research: The output is a cohesive report-style response in the chat. It typically includes an introduction, detailed sections, and a conclusion or summary of key points. It also includes citations to sources in-line or as footnotes, following the user’s requested format. For someone who posed a question, the answer is immediately useful — you can read it and get a comprehensive understanding. If you needed to present it to someone, you might want to copy it into a document and do some formatting (since in the chat it’s just text with markdown, presumably). But substance-wise, it’s largely ready. The tone is usually neutral and formal, which works for a report. If needed, you could ask ChatGPT to adjust tone or format within the same conversation after it presents the first result (e.g., “Provide the above as a bullet-point executive summary” or “Turn this into a slideshow outline”). Because it’s ultimately a chat AI, it’s flexible to reformat or augment the output if you ask. So, in terms of report readiness, ChatGPT is high — you get a well-structured answer — and it’s adaptable since you can further converse with it to change the output style. You don’t get a nicely designed PDF or a doc automatically, but that’s a minor inconvenience. The content quality is usually sufficient that some have even described it as potentially “equal to the output of a research analyst in tens of minutes”. So it’s fair to say ChatGPT’s deep research mode produces outputs that could be directly used in many cases, especially internal reports or as the draft of something client-facing.
• Google Deep Research: The output is very much a formatted report. It has sections and headings, sources listed, and an overall narrative structure. It’s probably quite similar in content style to what ChatGPT produces — thorough and factual — but with Google’s flavor of often listing out key findings succinctly. Because Google’s result lives in the Gemini app, one might review it there. But since it has a one-click export to Google Docs, the expectation is that users will export and then have a fully editable Google Doc. In that Doc, everything is already organized; the user might just do some light editing or add any specific context. So in terms of readiness, it’s very high — practically designed to be immediately taken into a report or an email. The inclusion of live links to sources makes it easy for someone reading the report to verify info, which is great for transparency if you’re sharing it. You might need to remove those source URLs or convert to footnotes depending on your audience (some might find bracketed URLs in text a bit distracting in a formal report), but that’s trivial with Docs. Google’s output likely also incorporates any specific instructions you gave in the plan; for instance, if you specifically asked it to include a table of data, it might actually create one in the Doc. Right now, at least text and simple tables are expected. The key is that Google’s result is structured and polished enough that you could basically forward it as-is to a colleague to read.
• Perplexity Deep Research: The output here is a detailed answer that reads somewhat like a well-researched article or a very long-form Q&A response. It isn’t delivered as a separate document, but the Perplexity page you get is shareable and nicely formatted for reading (with section headings if applicable, and each statement followed by superscript numbers linking to sources). In terms of content, it’s report-like. However, it might not automatically structure things into as many explicit sections as ChatGPT/Google do unless the question had multiple parts. Perplexity’s style often focuses on enumerating key findings and providing a narrative that connects them. It also can provide an answer in a conversational tone if the query was conversational, but for research mode presumably it defaults to an objective tone. The PDF export feature is a plus: when you export, you get a PDF with the answer text and citations. That PDF can be included in an email or printed, etc. It won’t have fancy visuals (unless the answer contained links to images, but I think it’s mostly text), but it’s quite functional as a report. If one needed to integrate it into a larger report, they’d have to copy text out of the PDF (since it’s not a .docx export), but one could also share the perplexity link which remains accessible. The output is definitely meant to be consumed directly by the user; it’s not raw data or fragments — it’s a coherent write-up. So, you could say Perplexity’s outputs are nearly ready-to-use, maybe just needing a bit of reformat if you want them in a different layout or combined with other content.
• Kompas AI: Kompas outputs a fully structured report within its interface, which can then be exported easily. Because the user can intervene in formatting and content during the process (with editing tools), by the time you finish with Kompas, the report is truly your report. That means it’s as ready as you need it to be. If you used all the features, you might have even tailored the tone and ensured consistency throughout. The output will have a clear structure given the outline-driven approach. It likely includes an introduction, sections aligned with each main point of your outline, and possibly a conclusion or recommendations section if that was part of the outline. Since Kompas is aimed at long-form professional documents, it probably pays attention to things like an executive summary or key recommendations (especially if the user indicates they want those). The text will have citations or at least reference links to sources for evidence-based points, maintaining credibility. When you export, you can choose a format like Word or PDF, which means you immediately have a file that can be shared or further edited in standard software.

In terms of polish, because Kompas allows manual edits, a user might correct any awkward phrasing or add specific data points before finalizing. So the end result could be on par with something a human analyst wrote with AI assistance. Essentially, Kompas’s output is as close to presentation-ready as you can get in an automated way. You may still need to proofread it — as one should with any AI-generated text — but the structure and content should already meet the requirements of a formal report. This is a standout feature: Kompas doesn’t just answer a question, it helps you produce a deliverable.

• Elicit: The “output” of Elicit is more of an information set than a composed answer. You get bits and pieces: a list of papers, short summaries, maybe a synthesized statement if possible, and any specific data you asked for in columns. If we compare it to a report, it’s like having the references and notes section of a report, but not the narrative. So Elicit’s results are not ready-to-present in the sense of a written report or an article. They are ready-to-use for research purposes: you can quickly take those and incorporate them into a report you’re writing. But if someone just looked at an Elicit output expecting a report, they’d see an outline of sources or a table, which isn’t what most would call a finished product.

For academic writing, Elicit gives you exactly what you need to start writing the related work or introduction section — key findings with citations. But you must do the writing and interpretation. Where Elicit does provide some ready text is in summarizing individual papers (it can generate a summary of a single study) or listing out the direct answers from each paper in a concise form. It even has a feature where you can ask a question and it will try to synthesize an overall answer by combining information from top papers. That synthesized answer is somewhat like what a narrative might be, but it’s not guaranteed to read as smoothly as something like ChatGPT would generate, and you’d still likely want to check each part against the source papers.

So in terms of report readiness, Elicit is more of a research assistant than a report writer. It gets you the facts and citations quickly, reducing the grunt work of literature review. But it doesn’t output a ready report or presentation. People often use Elicit in combination with other tools: for instance, get the data from Elicit and then use a writing assistant to help draft the report, or write it manually. Elicit thus accelerates the research phase, and the writing phase remains the user’s responsibility.

Continuous Research and Refinement

This aspect looks at how well each tool supports an iterative research process — can you refine your question, track progress, or handle long-term research needs?

• ChatGPT Deep Research: In terms of iteration, ChatGPT allows follow-up questions in the same chat thread. You can absolutely say, “Thanks for the report. Now, can you dig deeper into point X?” and it will do more work (possibly another deep research run or at least use what it already found plus a quick search). This conversational continuity is useful. However, ChatGPT’s agent doesn’t automatically update itself unless prompted. There’s no notion of a persistent project space beyond the chat history. If you close the chat and come back a week later, the context is saved in the conversation, but the information might be outdated (e.g., it won’t have noticed new events in the interim unless you explicitly ask again and it browses fresh). For ongoing research (like over months), ChatGPT isn’t tracking a topic over time; you have to manually re-ask to get updates. Another thing: since the deep research mode uses up a “query” allowance and takes a while, you might not run it for every small tweak. You might use it once for the big groundwork, then do lighter follow-ups in normal GPT mode to clarify or expand, which is a viable strategy but means the heavy-lifting agent is not continuously active — it’s more on-demand. If you plan to refine your question significantly, you might even start a new deep research query. ChatGPT doesn’t have an explicit feature to merge findings from multiple separate queries (except you can copy results from one chat to another, etc.).

For continuous research within one session, ChatGPT is good — the conversation can evolve fluidly. For long-term or large-scope research management, it lacks a bit of structure that tools like Kompas provide. You might end up with multiple chat threads on subtopics and have to organize that yourself. Also, because ChatGPT will eventually limit the length of conversation (there’s a context length limit, albeit huge in the new models, but still), extremely long sessions might hit some boundaries. In essence, ChatGPT is continuous in the conversational sense but not in a project-tracking sense.

• Google Deep Research: Google’s agent is built into a chat as well (Gemini assistant), so you have conversational continuity just like ChatGPT. You can refine the query by telling it to focus on something else, or ask a new question in the same thread which might reference the previous work. One advantage with Google is the integration with Docs — if you exported the initial report, you can manually note what else you want or even edit that doc and then perhaps ask Gemini to update based on the new points (though it’s not clear if Gemini can take a Doc as input for further research yet, but you could certainly copy text from it back into Gemini as context if needed). For ongoing tracking, Google doesn’t have a standalone “research project dashboard” either. However, Google being Google, one could use other products in tandem: for example, set Google Alerts for new info, or use Google Drive to keep versions of the research. That’s not the AI agent doing it automatically, but an ecosystem workaround.

Google’s agent does have the explicit notion of a plan and multi-step search in each session, but if the user’s needs change, they’d generate a new plan accordingly. The limited rollout means presumably one uses it deliberately for specific tasks, rather than having it running continuously. In a continuous research scenario like “monitor this topic and update me weekly,” Google’s agent doesn’t yet do that (though Google could potentially integrate it with something like Google Alerts in the future — not currently).

One unique thing: Google’s deep research can handle follow-ups in multiple languages or different angles easily because of the chat — you could ask it to translate parts of the report or find related info in another language site, leveraging Google Translate under the hood perhaps. That could be seen as continuous refinement in a cross-lingual way.

Overall, Google’s continuous capability is similar to ChatGPT’s: conversation-based refinement is supported and even encouraged, but long-term tracking is manual.

• Perplexity Deep Research: Perplexity operates both in a single-turn deep dive and conversationally. You can ask a follow-up question after the report, but in my experience, the follow-up might not automatically consider the entire context of the deep research result unless you specifically refer to it. The system might drop some context between queries because the Deep Research might be treated as a somewhat standalone process. However, Perplexity does have a “thread” mode (in normal Q&A, you can click continue and it keeps context to some degree). If that extends to Deep Research mode, it means you could do a Deep Research, then click “ask a related question” and it would keep some of the findings in mind or at least the conversation history.

That said, the UI of Perplexity is not project-based — it’s Q&A based. Continuous research is more a matter of asking successive questions and maybe keeping the browser tab open. If you close it, you’d have to rely on memory or saved links. Perplexity Pro might have history features or allow longer context with GPT-4. There was mention of “Perplexity Copilot,” a feature where the AI can proactively ask you clarifying questions and help refine — that indicates a form of interactive refinement, though it’s still in the domain of a single session, not a multi-day agent.

Because Perplexity is fast and free, doing an iterative approach (many quick deep research queries adjusting the focus each time) is practical. It’s not burdensome to run things multiple times. However, you as the user have to integrate those results. The tool itself doesn’t merge them for you aside from what you ask in one thread. The output is more of a collection of individually complete answers that you might need to tie together manually.

• Kompas AI: This tool is explicitly designed for ongoing and iterative research. The “Research Further” function allows you to continually build on what you have, which is a big plus for refinement. For example, after the initial report is drafted, you might realize you want more detail on one section; you click a button, and Kompas goes deeper on that, appending or integrating the new findings. This can be done for multiple sections, essentially layering depth incrementally. Because it keeps everything organized in one document, you don’t lose track of context — it’s all right there in your report outline. You can also revise the outline midway if your focus shifts, and then have the AI fill in new sections. This is a very powerful way to handle evolving research questions. If you consider long-term research (like spanning weeks or months), Kompas’s ability to save projects means you have continuity. You can come back later, see what you did (outline, content, sources), and then press “update” or “research further” to refresh sections with new info that might have emerged. It’s like having a living document that an AI can continue to work on whenever you need.

This persistent, structured approach is something the others don’t explicitly offer. They are session-centric, while Kompas is project-centric. For a user, that means less chance of duplicating work or forgetting what’s been done — Kompas sort of serves as the memory of your research effort.

• Elicit: Elicit doesn’t have a concept of a single ongoing project, but you can certainly conduct systematic research using it. For example, you might start wide, then narrow. It won’t automatically know you’re doing a series of related searches unless you make use of features like selecting papers and refining (which is a form of continuous drill-down within one session). If you sign in, Elicit does let you save query results as a list of papers, which is handy — you could have a collection of relevant literature saved in the tool and then ask different questions of that collection. That’s continuous in the sense of focusing on a set of sources over multiple queries.

However, outside of that, Elicit doesn’t update you if new papers are published unless you search again later. There’s no agent running in the background updating your results. You as the researcher would check again maybe after a few weeks to see if new studies have come out.

Where Elicit aids continuity is in how it lets you pivot: from one paper you can jump to related works, or from one query you can quickly tweak it and run another. It’s all manual but supported by the UI with quick actions. Also, since it’s free and accessible, you can come back to it anytime.

Who Benefits Most from Each Tool?

Each of these tools has an ideal user or scenario where it shines:

• ChatGPT Deep Research: This tool is well-suited for professionals and analysts who face complex, multidimensional questions and need a thorough, reasoned analysis, and who have access to the premium service. For instance, a policy analyst comparing various economic models or a finance professional analyzing a company across markets could benefit from ChatGPT’s depth. It’s also great for users who want minimal effort in guiding the research — you hand off a broad question and get back a detailed answer. Busy executives or consultants might use it to get up to speed on a topic quickly. General curious users could enjoy it for deep dives (“Explain the history of quantum computing in detail”), but the price barrier means it’s not for the casual hobbyist broadly. So, it benefits high-end users who need deep analysis: think along the lines of researchers, strategists, advanced students, or journalists doing investigative work.
• Google Deep Research: This is a natural fit for anyone already in the Google ecosystem — from business professionals to students — who wants comprehensive info with minimal fuss. Market researchers, marketing strategists, and small business owners could use it to quickly gather competitive intel or trend analysis, since it can pull from news, blogs, company sites, etc., and then they can export to Docs and share with their team. It’s also beneficial for international users or research across languages, given Google’s global advantage. If someone needs a quick but thorough briefing on something and doesn’t have specialized research tool access, Google’s relatively lower barrier is attractive. Educators or students might use it for preparing lessons or essays, getting a lot of info in an organized form. And because it’s integrated with familiar tools, even less tech-savvy users (who might find ChatGPT’s idea of an “AI agent” intimidating) could use it by essentially doing what they normally do (searching) but letting the agent compile the results. In short, power internet users and professionals who value breadth and convenience will benefit, as will those who need to collaborate using Google’s suite.
• Perplexity Deep Research: Given that it’s free for modest use, students and independent learners stand out as beneficiaries. A college student writing a paper can use Perplexity to gather facts and sources without paying for expensive tools. Journalists or bloggers could also use it to fact-check or quickly gather background info with sources for an article. It’s also great for casual curious minds who love to explore topics in depth but don’t have institutional access to pricey software. Because it’s quick, even professionals might use it as a first pass to scope a topic before using heavier tools. For example, a business analyst might run a Perplexity deep research to get a general lay of the land, then decide if they need more detailed follow-up via other means. Researchers in developing regions or small companies/startups with limited budgets could rely on it heavily, since it delivers a lot of value at low cost. Its ease of use means it’s also suitable for non-experts: someone who’s not a research specialist but needs to find info (say a content writer or a product manager gathering user research from public sources) can use it with little training. Essentially, Perplexity benefits the broad public and cost-conscious users as well as those who prioritize having sources at their fingertips for credibility.
• Kompas AI: This tool seems ideal for users who need to produce polished research reports or documents as an output, and who appreciate a structured approach. Business consultants, market research analysts, report writers, and policy researchers come to mind. If your job is to regularly create reports (market analyses, competitive reports, white papers, detailed reviews), Kompas is tailored for that workflow. It’s also a strong match for teams or individuals who have ongoing research projects — for instance, a think tank researching a topic over several months, or a company investigating a new market — because it allows saving and iterative building. Graduate students or thesis writers might find Kompas very helpful too: writing a literature review or a background chapter with Kompas could save time, given you can structure it and fill in the content iteratively. Another group is non-technical professionals who might find a chat interface too unstructured; Kompas’s guided process could be more approachable for, say, someone in a managerial role who isn’t an expert at prompting AI but knows what sections a report should have. By providing a template and fill approach, Kompas reduces the skill needed to steer the AI. It also might appeal to content creators and writers who are preparing long-form content (like an e-book or a comprehensive guide) — they can outline their chapters and let Kompas gather the raw material for each, then refine it. In essence, Kompas benefits those aiming for high-quality, shareable research documents and those who value a guided, step-by-step research process.
• Elicit: Elicit is the go-to for academic researchers, students, and anyone who needs scholarly evidence. If you’re writing an academic paper, doing a systematic literature review, or even just a research proposal, Elicit can save you enormous time in finding relevant literature. Scientists, social science researchers, evidence-based practitioners (like policy analysts, doctors looking up medical research, etc.) would find Elicit extremely useful to quickly gather what’s published on a question. It’s also great for PhD students or undergrads working on literature reviews for theses or dissertations. Librarians and information specialists might use it to assist patrons with research questions. Writers of research-heavy content (like a science journalist or a non-fiction author) could use Elicit to ensure they’re citing actual studies and to discover studies they might have missed via normal search. On the other hand, Elicit is not aimed at someone who just wants an easy answer without doing any reading — those users will not get as much from it. It specifically benefits those who are willing to read the source material and just want help finding and summarizing that material. Also, in fields where information reliability is paramount (e.g., healthcare, academic publishing), Elicit’s approach of giving the original sources is very valuable and in some cases might be the only acceptable way.

Each tool, therefore, has a somewhat distinct primary audience: ChatGPT DR for high-end analytical work, Google DR for broad and convenient info gathering, Perplexity for accessible research to the masses, Kompas for structured report creators, and Elicit for academic literature deep-divers. Of course, there’s overlap and many people might use a combination depending on the task.

Final Thoughts

AI deep research tools are rapidly transforming how we approach information gathering and analysis. As we’ve seen, the leading solutions each carve out their own niche on the spectrum from guided report generation to open-ended literature discovery. A few key insights emerge from this comparison:

• No one-size-fits-all: Each tool excels in different scenarios. If you need a polished report with minimal assembly required, Kompas AI’s structured approach and continuous research workflow can be a game-changer, guiding you from outline to final draft seamlessly. If your priority is raw analytical power and you have access to premium AI, ChatGPT Deep Research delivers top-notch reasoning and depth, acting like a versatile domain expert — albeit one you should double-check. For quick, fact-focused answers with sources on hand, Perplexity’s Deep Research mode is a speedy and convenient ally that democratizes research by being freely available. Meanwhile, Google’s Deep Research offers a very user-friendly on-ramp with global and cross-lingual reach, which is perfect for broad exploratory questions and integration with everyday workflows. And for scholarly research needs, Elicit stands out by directly tapping into scientific literature, ensuring your insights are grounded in peer-reviewed knowledge.
• Balancing depth and trust: One recurring theme is that these tools can gather a lot of information — but more information isn’t automatically better unless it’s accurate and relevant. They each have mechanisms to provide evidence (citations, links, etc.), which is crucial. In professional settings, it’s reassuring to have sources like those provided by Perplexity or Kompas right alongside AI-generated text. The best practice with any of these tools is to use that feature: verify the key facts, and use the AI’s output as a starting point, not the final unquestionable truth. Encouragingly, the tools are getting better at filtering out misinformation — perhaps future versions will cross-verify facts among multiple sources automatically before presenting them.
• Usability vs. control: There’s a bit of a trade-off between ease-of-use and the level of control or structure. ChatGPT and Perplexity are extremely easy: you ask, you get answers. Kompas and (to some extent) Google give you more control over the process (outline or plan), which can lead to more tailored results but requires a tad more input from the user. Neither approach is inherently better — it depends on the user’s needs. It’s heartening to note that AI tools are evolving to cater to different user styles. Those who want a hands-off experience can get it; those who want to steer the AI like a research assistant can do that too. Over time, we might even see hybrid interfaces (for instance, ChatGPT might introduce optional outline generation like Kompas has, or Google might allow an “auto” mode for simpler queries).
• Report readiness is the new benchmark: One clear trend is that these tools are moving beyond just chatty answers toward deliverables. Kompas explicitly positions itself as producing long-form reports. ChatGPT and Google’s agents also produce structured write-ups with sections and even tables. This reflects what many users actually need — something they can take into a meeting or include in a paper. It wouldn’t be surprising if in the near future all AI research assistants offer multiple output formats (brief memo, detailed report, slide deck outline, etc.) as options. The current leaders already hint at that: OpenAI’s plans to include charts and more visuals, Google’s seamless Docs export, Kompas’s editing for style — all point to AI not just fetching info, but helping polish it for the end audience.
• Continuous learning curve: Using these tools also changes how one learns and researches. It’s now feasible to regularly offload initial scouting of a topic to an AI, freeing humans to do higher-level thinking — like interpretation, strategy, and creative insight — on top of the AI’s findings. However, it also means developing new skills: knowing how to phrase queries, how to prompt for more detail, when to trust the AI and when to verify with your own eyes. For example, a good researcher might use Elicit to gather sources, then use ChatGPT to summarize those specific sources, then maybe use Kompas to organize the entire narrative — orchestrating multiple tools.

Looking to the future trends in AI deep research, we can anticipate even more integration. It’s plausible that eventually a single platform might incorporate all these strengths: imagine an AI that can search academic papers, web articles, and internal company data, then draft a fully formatted report, ask you for feedback on which sections to expand, continuously update it with new findings, and even cite everything in your preferred citation style — all in one workflow. Companies are racing toward that vision. We may also see more specialization before we get one-tool-to-do-it-all; for instance, domain-specific research AIs (for legal research, or clinical medical research) that know the nuances of those fields.

Another trend is collaboration features — currently research AI is often a solo experience, but in businesses, team members might want to collaborate on an AI-assisted research project. Google’s advantage with Docs might set the stage for collaborative AI research, where multiple people can chat with an agent in a shared document space. Kompas’s approach could similarly be extended to team workspaces where a team collectively builds a report with AI help.

From an objectivity standpoint, the competition between these tools is driving rapid improvement, which benefits users. It’s also making deep research capabilities more accessible: what used to require a trained research analyst and days of work can now be done (not perfectly, but decently) by an AI in minutes. That doesn’t eliminate the role of humans — rather, it shifts our role to being editors, strategists, and fact-checkers who work alongside the AI.

Closing recommendations: Given the current landscape, here’s a brief guide for different user needs:

• If you’re a student or casual researcher on a budget: Start with Perplexity Deep Research for quick, well-sourced overviews. If you need scholarly sources, add Elicit to your toolkit to find actual papers to cite. These two free tools in combination can cover a lot of ground.
• If you’re a professional needing a thorough report but don’t have time to curate it yourself: Consider Kompas AI. It will help structure the project and produce a near-finished report that you can easily tweak to perfection, which is ideal for business or policy contexts where presentation matters.
• If you’re an analyst or researcher dealing with very complex questions and you have access to premium AI services: ChatGPT Deep Research is worth the investment for its depth and reasoning. It can save you time by handling complexity, but be prepared to verify its outputs. You might use it to tackle the hardest parts of a problem.
• If you value integration and the latest information: Google’s Deep Research via Gemini is a strong choice, especially if you’re already using Google Workspace. It’s great for broad, up-to-date research that you can immediately incorporate into documents and share. For example, a team prepping for a marketing plan might use Google’s AI to gather all recent trend data and then collaboratively build on that in Docs.
• If you need academic credibility above all: Elicit should be your first stop. It ensures you’re building on established knowledge and can point you directly to the seminal papers on your topic. Once you have those, you could even feed them into another tool (like ChatGPT with browsing or Kompas) to summarize or compare findings.

In conclusion, AI deep research tools are proving to be invaluable assistants, each with their own personality and strengths. Whether you prefer an AI that behaves like a knowledgeable librarian, a tireless research intern, or a draft report writer, there’s an option available. As these tools continue to evolve, we can expect research — be it for business, science, or personal curiosity — to become an increasingly collaborative dance between human insight and AI-powered data gathering. The key for users is to stay objective, leverage the unique advantages of each tool, and remain aware of their limitations. With that balanced approach, these AI tools can significantly amplify our ability to learn and make informed decisions in an information-rich world.