DeepResearch: The AI Superpower That’s Making Google Searches Look Prehistoric

Everyone seems to be adding DeepResearch to their LLM (Large Language Model) chatbot’s which are seemingly in everyone’s pocket or desktop already.

While your regular AI assistants like ChatGPT and Claude are great at quick responses based on their training data, DeepResearch takes things to a whole new level.

Remember when DeepSeek showed us how AI could “show its work” through chain-of-thought reasoning? DeepResearch is the next leap forward leveraging reasoning.

DeepResearch promises to transform how professionals conduct research, analyse data, and synthesise complex information in a novel way.

How does it work?

Think of DeepResearch as your personal research assistant on steroids. Instead of just pulling from a static knowledge base, these systems actively search the web, read through documents, and piece together comprehensive reports. They can handle text, images, and PDFs, while being smart enough to:

• Judge which sources are trustworthy and resolve conflicting information, much like a human researcher would when faced with contradictory data
• Adapt their search strategy as they learn more about the topic, following leads and connections just like a detective
• Pull everything together into a coherent narrative that actually makes sense

How can I try it?

• Google’s got their version in Gemini Advanced (paid)
• OpenAI’s added their DeepResearch button for ChatGPT (paid)
• Perplexity AI uses their own DeepSeek implementation and they give you 5 free searches daily
• Abacus.ai has all of the popular models in one place (my referral code if you fancy a spin — only $9 a month with the first month free)

Clearly the goal here is to start replacing (or at least augmenting) those research analysts which are usually trawling through Google, Scientific Research papers and Company information to build ‘McKinsey’ style presentations.

You could really see this type of research query being used for deeper analysis of Legal, Finance and Scientific tasks.

Here is an example of a prompt you can modify but its a lot of trail and error to find out what works. Enjoy.

Precision Prompts: “Compare NVIDIA H100 vs AMD MI300X for LLM training, focusing on power efficiency and Tensor core utilization”

Source Guardrails: “Prioritize peer-reviewed studies from IEEE Xplore”

Output Formatting: “Structure as IMRAD report with summary table