Project: Research Companion
A general-purpose chatbot is a poor researcher. It has a knowledge cutoff, cannot access live sources, blends information from across its training data without attributing anything, and has no memory of what you researched last week. This project builds a personal research companion — a stack of three specialist AI tools working together that covers discovery, deep reading, and organisation. You will end up with a reliable, repeatable research workflow for any topic. No coding required.
Why Generic Chat Fails at Research
| Research need | Generic chat (ChatGPT / Claude) | Specialist tool |
|---|---|---|
| Live, up-to-date information | Training cutoff — may be months out of date | Perplexity — real-time web search with citations |
| Source attribution | Blends sources — hard to verify any claim | Perplexity — numbered inline citations, every claim sourced |
| Reading your own documents | May mix uploaded content with training data | NotebookLM — grounded only in what you upload |
| Academic literature | Cannot reliably find real papers or citations | Elicit — 125 million academic papers, structured extraction |
| Continuity across sessions | No memory of past research threads | NotebookLM notebooks — persistent, topic-specific workspaces |
| Comprehensive report on a topic | Possible but slow and requires careful prompting | Deep Research (Perplexity / ChatGPT / Gemini) |
Your Three-Tool Research Stack
Each tool does one job extremely well. Together they cover the full research cycle.
Tool 1 — Perplexity AI (Discovery)
Your starting point for any research question. Perplexity searches the live web and academic sources simultaneously, then writes a structured answer with numbered inline citations you can click through to verify. Every claim links to a real source.
Tool 2 — Google NotebookLM (Organisation & Deep Reading)
Your research workspace. Upload PDFs, articles, notes, web pages, and YouTube links. NotebookLM answers only from what you uploaded — it cannot confabulate from training data. Use one notebook per research topic and it becomes a persistent, searchable knowledge base for that subject.
Tool 3 — Elicit (Academic Literature)
For any question that needs peer-reviewed evidence. Elicit searches 125 million academic papers and returns structured data — not just titles, but key findings, methodologies, sample sizes, and limitations — in a sortable table. Designed for systematic literature reviews without needing a university library subscription.
When to Use Each Tool
| Research situation | Best tool | Why |
|---|---|---|
| Quick background on an unfamiliar topic | Perplexity standard search | Fast, cited, current — gets you oriented in 60 seconds |
| Comprehensive report you can actually use | Perplexity Deep Research | 50+ sources, structured multi-page output, exportable |
| Understanding a set of specific documents | NotebookLM | Source-grounded — answers come only from your files |
| Turning dense material into audio for commuting | NotebookLM Audio Overview | Podcast-style discussion of your uploaded sources |
| "What does the research actually show about X?" | Elicit | Real papers with real data — not AI summarisation |
| Finding papers related to one you already have | ResearchRabbit / Connected Papers | Citation network maps — visually explores what a paper cites and is cited by |
| Deep report with more narrative depth than Perplexity | ChatGPT Deep Research or Gemini Deep Research | Better synthesis and argumentation — slower but more analytical |
Build Your Research Companion — Step by Step
Step 1: Set Up Your Tools (15 minutes, one time)
Step 2: The Standard Research Loop
For any new research question, follow this sequence:
| Step | Action | Tool |
|---|---|---|
| 1 | Ask your question in plain language. Review the cited sources, not just the answer. | Perplexity standard search |
| 2 | Follow up with 2–3 clarifying questions in the same Perplexity thread to go deeper. | Perplexity (follow-up mode) |
| 3 | Save key articles, PDFs, or web pages. Upload them to a topic-specific notebook. | NotebookLM notebook |
| 4 | Chat with the notebook to extract specific information, compare sources, and spot contradictions. | NotebookLM chat |
| 5 | If empirical evidence is needed, run the question through Elicit to find real studies. | Elicit |
| 6 | Add the most relevant papers from Elicit to your NotebookLM notebook. | NotebookLM |
| 7 | Generate a Mind Map or Audio Overview to consolidate your understanding. | NotebookLM Studio |
Step 3: Perplexity Deep Research (for big questions)
For research questions that need a comprehensive report — a market overview, a technology landscape, a policy comparison — use Deep Research. It spends 2–4 minutes analysing 50+ sources before writing a structured, multi-page response.
Step 4: Build a Persistent Research Library
The real power of this setup is that research accumulates. Each NotebookLM notebook becomes a searchable, queryable library for that topic — and it never forgets.
Copy-Paste Prompt Templates
Ready-to-use prompts for common research situations.
Initial topic orientation (Perplexity)
Give me a structured overview of [topic]. Cover: (1) what it is in plain language, (2) the main sub-topics or components, (3) the key debates or open questions, (4) the most important recent developments in 2024–2025. Cite your sources for each section.
Deep Research briefing (Perplexity Deep Research)
Write a comprehensive briefing on [topic] as of 2025. Structure it as: (1) Executive summary (3 sentences), (2) Background and context, (3) Current state of the field, (4) Key players and perspectives, (5) Major trends and forecasts, (6) Open questions and controversies, (7) Recommended next sources for deeper reading. Use at least 30 sources.
NotebookLM — source synthesis
Across all the sources in this notebook, what are the 5 most important claims or findings? For each one, tell me which source(s) support it, and whether any other source contradicts or qualifies it.
NotebookLM — gap finder
Based on the sources in this notebook, what important questions about [topic] are NOT answered by what I have uploaded? What additional source types or perspectives would fill those gaps?
Elicit — evidence search
[Paste this into Elicit's search bar] Does [intervention/behaviour/treatment] improve [outcome] in [population]? After Elicit returns papers, add these columns to the extraction table: sample size, study design (RCT/observational/meta-analysis), effect size, and limitations.
Research-to-output summary (NotebookLM or Claude)
I have researched [topic] and gathered [N] sources. Based on this material, write a 500-word summary I can share with [audience — e.g. colleagues / my manager / a general reader]. The summary should: cover the key findings, note any important uncertainties, and end with 3 actionable implications. Do not add information beyond what is in my sources.
What Works Well
Tool specialisation beats one-tool-does-all
Using the right tool at each stage — Perplexity for discovery, NotebookLM for depth, Elicit for evidence — produces far better results than trying to use a single chatbot for everything.
Source-grounding is the key safety feature
NotebookLM's refusal to answer outside your sources is a feature, not a limitation. If it says "I don't see that in your sources", that is useful information — you need to find and upload the relevant material rather than getting a confident hallucinated answer.
Audio Overviews dramatically improve retention
Generating a NotebookLM Audio Overview of your sources and listening to it as a podcast — on a commute, during exercise — significantly improves recall compared to reading alone. Available in 80+ languages from August 2025.
Research compounds across sessions
A notebook you built two months ago is immediately queryable when the topic comes up again. You are not starting from scratch — you have an annotated, searchable library of everything you already found.
Failure Modes
Trusting Perplexity citations without clicking through
Perplexity cites sources inline, but it can still misrepresent them. The citation format is not a guarantee — it means "this claim came from this URL", not "this URL fully supports this claim". Open the top 2–3 sources for any important claim.
Uploading low-quality sources to NotebookLM
NotebookLM is only as good as what you upload. If you fill a notebook with opinion pieces, company blog posts, and SEO content, the AI will synthesise those confidently. Garbage in, garbage out — even in a source-grounded system.
Using Elicit for non-empirical questions
Elicit is powerful for evidence-based questions with measurable outcomes. It is not useful for opinion, strategy, or emerging topics with little peer-reviewed literature. Use Perplexity for those instead.
Treating Deep Research output as finished work
Deep Research produces impressive-looking reports that can contain errors of synthesis or missed nuance. Treat the output as a high-quality first draft — review the cited sources, spot-check the key claims, and add your own judgement before using it professionally.
One giant notebook for everything
Mixing unrelated topics in a single notebook makes it harder to query. "What are the main findings?" becomes ambiguous when the notebook contains sources on three different subjects. Keep notebooks scoped to a single topic or project.
Extend Your Research Companion
2025–2026 Developments
Deep Research became a standard feature across all major platforms (Feb 2025)
Perplexity, OpenAI, and Google all launched "Deep Research" modes in early 2025. Each runs dozens to hundreds of searches automatically, synthesises across sources, and produces a structured report. What required a research analyst for hours can now be done in under five minutes by anyone. Perplexity's version launched February 14, 2025 and is free for 5 queries per day.
NotebookLM's Audio/Video Overviews went global (Aug–Sep 2025)
In August 2025, NotebookLM removed the English-only restriction, expanding Audio Overviews to 80+ languages. The September 2025 update added format choices — Deep Dive, Brief, Critique, Debate — making it genuinely useful for different research styles. Video Overviews followed in the same rollout. A Lecture Mode (single-host, 30-minute structured talk) was in testing as of late 2025.
Research tool convergence — every AI is adding search
ChatGPT, Claude, and Gemini all added or improved live web search in 2025. The distinction between "chat AI" and "search AI" is eroding. What makes dedicated tools like Perplexity and Elicit still worth using is specialisation: Perplexity's citation density and academic mode, Elicit's structured paper extraction. As general tools improve, the bar for specialist tools rises — but their structural advantages remain.
Checklist: Do You Understand This?
- Can you explain why using three specialist tools beats one general-purpose chatbot for research?
- Do you know when to use Perplexity versus NotebookLM versus Elicit?
- Can you describe the standard 7-step research loop from initial question to organised notebook?
- Do you know what Perplexity Deep Research does and how many free queries you get per day?
- Can you explain why NotebookLM's source-grounding is a feature, not a limitation?
- Do you know how to use Elicit for an evidence-based research question?
- Can you name three failure modes and how to avoid them?
- Do you know what the NotebookLM Audio Overview formats are as of September 2025?