Researching with AI
Exploring the New “Agentic” Research Capabilities in LLMs
Over the past few months, a powerful new capability has quietly rolled out across multiple large language models (LLMs), capturing a significant amount of attention: deep AI-driven research tools. It’s been an evolving landscape but now that things have settled out a bit, it’s time to explore them. These tools, which multiple AI companies have named "Deep Research" (because AI marketing is hard, I guess?) will be referred to here simply as AI Research.
AI Research tools allow users to perform extensive search and analysis tasks with minimal effort. With a single prompt, they search public information sources, compile summaries, and deliver insightful analysis — all without requiring constant user input.
Why manually search Google when you can generate a comprehensive, cited research report with a single prompt? Let’s dive in.
AI Research as an early form of AI agents
The broader AI community sees agents – AI applications that work autonomously with minimal human intervention – as the next major phase in AI development. This is distinct from tools like Microsoft’s Copilot that focus on utilizing AI to enable and empower individual work. AI agents can do narrow and useful work independently in an ongoing fashion.
This trend aligns with OpenAI’s five-stage AI innovation framework (according to Bloomberg and summarized by Inc.):
Stage 1: “Chatbots, AI with conversational language”
Stage 2: “Reasoners, human-level problem solving”
Stage 3: “Agents, systems that can take actions”
Stage 4: “Innovators, AI that can aid in invention”
Stage 5: “Organizations, AI that can do the work of an organization”
AI Research represents a narrow, early-stage agent designed specifically for conducting comprehensive research. It bridges the gap between standard chat-based LLMs and fully autonomous AI agents.
When to use AI Research vs AI Chat
AI Research tools function differently than traditional AI chat — and often require more time and processing power. AI Research queries take time as they comb through the web and generate analysis (simple searches could be a few minutes but more complex could take 20-30min!). In addition, some services limit their usage - for example, OpenAI’s Plus plan currently includes only 10 Deep Research searches per month, while Google’s Gemini and Grok offer more liberal access.
Standard AI Chat Uses
In general, stick to standard AI Chat when you’re looking for a quick answer or fact-check (that you should double check), summarization and simplification, creative work, casual research, brainstorming, comparisons, and so forth. Efforts that are real-time, only need surface level understanding, that are interactive, and don’t need heavy source material.
AI Research Uses
Consider a more in-depth research search if you’re looking for something more substantial, such as comprehensive industry/market reports, comparative analysis across different sources, source-cited briefings, trend and historical analysis, competitor intelligence or benchmarking, risk assessments, academic literature reviews, policy tracking, etc. Efforts that require a more thoughtful, deliberate, and cited approach.
What are some more general uses that we might want to consider AI Research for?
Explore personal interests or hobbies through curated learning paths.
Prepare for job interviews with in-depth company profiles.
Research health conditions to better understand a medical diagnosis.
Evaluate relocation options based on custom criteria.
Analyze market investments or real estate opportunities.
Plan complex home improvement projects with comprehensive background information.
What tools are available?
AI Research is a new AI LLM capability that is increasingly becoming part of the baseline expectations with AI models. As of early 2025, several platforms offer this functionality:
Google released a “Deep Research” capability for Advanced ($20/mo) Gemini users in December 11, 2024.
Genspark launched a similar but free “Deep Research” agent in January 2025.
OpenAI announced its own “Deep Research” for Pro ($200/mo) users on Feb 2, 2025, but just recently made it available to all Plus ($20/mo) users with limited use (currently 10 searches/month).
Perplexity announced a free Deep Research capability on Feb 14, 2025.
xAI released a “DeepSearch” capability (via live broadcast) integrated into their new Grok 3 model on Feb 19, 2025 that users can use for free.
There are others as well. The speed of adoption is remarkable — in less than three months, a single feature evolved from a Google experiment into a baseline capability across multiple leading platforms.
Comparing AI Research Tools
I recommend trying out more than one and checking out the differences in output for yourself. Each has different approaches, strengths and weaknesses. The general consensus as I understand it (at least as of March 1, 2025) seems to be something like the following:
OpenAI’s Deep Research – Considered the gold standard for thorough analysis and insightful synthesis across multiple sources.
Google Gemini Deep Research – Strong at searching (it’s Google, after all) but weaker at analysis, making the results useful but less refined.
Perplexity & Grok – Promising alternatives showing potential to rival Google’s capabilities.
Ethan Mollick stated the following when comparing OpenAI and Google: “Both outputs represent work that would typically consume hours of human effort - near PhD-level analysis from OpenAI's system, solid undergraduate work from Google's.” I think that’s a useful benchmark.
How to Use AI Research Agents
On a practical note: using these tools is easy — typically, you enable the option within a normal AI chat interface. Each platform handles it slightly differently, but they all aim for the same outcome: hands-free, autonomous research reports.
A few examples by major service below:
Let’s Test It: The AI Research Showdown
I decided to test out the AI Services that I use with deep search capabilities with the exact same prompt. Let’s see how the outputs differ. I’m personally curious how AI is being integrated into the US Federal government and came up with the following prompt (with a little help from GPT):
📝 Example Prompt: Conduct a deep research study on the adoption of Generative AI tools in U.S. federal agencies from 2020 to 2025.
· Objective: Understand trends, key use cases, obstacles, regulatory environment, and future outlook.
· Audience: Executive leadership team unfamiliar with government tech adoption processes.
· Scope: Focus on agency-level initiatives, funding programs, pilot projects, and regulatory frameworks.
· Sources: Prioritize GAO reports, Congressional Research Service papers, federal agency publications, and credible news outlets like FedScoop or NextGov.
· Deliverable: 5-page written report with sectioned headers, clear summaries, and direct citations (APA style). Include a comparative table showing adoption rates by agency type.
· Exclusions: Ignore content focused solely on state or local governments.
· Tone: Analytical and neutral.
I fed the exact same prompt into OpenAI, Gemini, and Grok and you can see the results for yourself. Personally, I think the consensus is largely correct - OpenAI’s output was exhaustive and analytical, Google’s was well-sourced but less synthesized, and Grok produced results closer to Google’s.
With all that said, AI Research tool is going to provide interesting and usable results. (Just remember, always verify anything important.)
Crafting Effective Deep Research Prompts
As stated above, prompting AI Research agents are going to be a little different than a typical chat. It’s going to be much more important to get your prompt right up front, because once you lock in your prompt, the AI Research agent starts spinning, so clear direction upfront is essential.
Below is guidance from GPT itself, which I followed loosely in my prompt above.
Key Components
Scope & Objective – Purpose and level of detail.
Context & Audience – Who it’s for and what they already know.
Sources & Timeframe – Preferred data sources and date range.
Format & Deliverable – Narrative, tables, visual aids, citations.
Depth & Exclusions – How detailed should it be, and what to omit.
Example Prompt Template
Conduct deep research on [TOPIC].
Objective: [Briefly state purpose].
Scope: [Timeframe, geography, industries].
Sources: [Preferred sources].
Depth: [High-level, moderate, exhaustive].
Audience: [Target readers].
Format: [Narrative, tables, visual aids].
Exclusions: [Omit specific content].
Citations: [Yes/no, preferred style].
Tone: [Neutral, advisory, etc.].
My recommendation as you get comfortable with AI Research tools is to first use AI chat to help you craft your prompt, then start a new chat and copy and paste it in for the real deal.
Final Thoughts: Research for Everyone
AI Research tools aren’t perfect, but they’re improving rapidly. There are also numerous options available for free. I encourage you to try a few, compare their outputs, and see what works best for your needs.
Mastering these tools today will prepare you for even more advanced AI capabilities that on the horizon. Expect more agents like this to come soon.
