Deep Research Prompting

This data was originally featured in the March 5th, 2025 newsletter found here: INBOX INSIGHTS. March 5, 2025: Finding Joy in AI, Deep Research Prompting

In this week’s Data Diaries, we have an embarrassment of riches when it comes to AI-enabled deep research:

• For manual deep research, there’s Google’s NotebookLM
• For automated deep research there’s:
  • OpenAI o1 Deep Research (paid)
  • Google Gemini 1.5 Deep Research (paid)
  • Perplexity Deep Research (paid, but free users get 5 uses per day)
  • xAI Deep Search (paid, free users get an unknown but small number of uses per day)
  • Deepseek R1 with Search (free, though the web search part is unreliable)

Here’s the catch: all deep research tools use generative AI. Which means all deep research tools still must obey the basic tenets of prompting generative AI tools. What does that mean?

A prompt like:

Write a report on how to optimize a landing page

Is going to net you… well, let’s just say the results will be far from stellar. So how do you write a great prompt for a deep research tool?

Here’s the challenge. Each of the tools operates differently, but using a framework like the Trust Insights 5P framework will generally yield good results.

• Purpose: What are you commissioning the deep research for? What’s the point?
• People: Who is the deep research for and what are their requirements?
• Process: What are the essential components of your deep research? What areas do you want to explore? What time ranges are the data? Almost all deep research tools use chain of thought, so you don’t need to tell it what to do step by step, but you should absolutely focus on what should be allowed or disallowed.
• Platform: Which data sources should be allowed or not?
• Performance: What’s the final outcome expected to be? How should it be structured?

Here’s an example deep research prompt, the same as before, but structured with the 5Ps. For clarity, I’ve kept the 5P structure exactly; this is exactly what I’d paste into a deep research tool:

Let’s set up our deep research project. You’re an expert researcher in the field of CRO, or conversion rate optimization. Here are the project details:

• Purpose: Write a paper on the best practices for building high-converting landing pages in 2025.
• People: The audience of the paper is a CRO expert, so do not dumb down, simplify, or reduce extremely technical concepts. Technical language is encouraged.
• Process: We want the paper to focus on UI, design, copy, persuasion, writing style, structure, and expert tips and tricks. Use your knowledge to identify all the major topic areas of CRO and landing page optimization.
• Platform: Use publications from 2024 and 2025 only; do not use any content earlier than 2024. Focus on marketing and marketing technology publications.
• Performance: Be sure the paper is well structured and logical, ordering the categories in the order that a CRO expert would build a landing page.

Using the 5P framework does two things. First, it forces us to think through exactly what we’re asking for. It makes us consider the ways that a deep research tool might go off the rails, from pulling in old knowledge to extracting sources that aren’t credible.

Second, it helps the LLM understand the structure of what we’re asking as well, calling out the discrete parts of the inquiry. When we structure it with the 5Ps, including the 5P headings, the LLM has less ambiguity. It knows who the audience is, what is expected, what success looks like.

As Katie often says, prompting AI is like delegating work to a human. The more clear you can be in your instructions, the better the human or machine will perform. Use the 5Ps to help structure your deep research prompts!

[postsignature}