TL;DR
One tool will not carry a serious research task end to end. This workflow splits the job across the tool that each stage is actually good at: Perplexity for the landscape scan, NotebookLM for grounded synthesis of sources you control, a Deep Research agent (ChatGPT, Gemini, or Claude) for the draft, and you for the citation audit and the final brief. Six stages, one folder you can clone next time. A cold run takes 2-4 hours; with a saved template, 60-90 minutes.
Who this is for
Analysts, PMs, consultants, founders and grad students who run research weekly and want a system, not a one-off prompt. If you only need a quick one-tool answer, this is overkill — open Perplexity and stop. This guide is for the brief that someone will push back on.
Why one tool isn’t enough
Each tool has a different failure mode, and chaining them cancels the failures out:
| Stage tool | Strength | Failure mode it has alone | What the next stage fixes |
|---|---|---|---|
| Perplexity | Fast, citation-first web scan | Shallow synthesis; cites whatever ranks | NotebookLM re-grounds on sources you vetted |
| NotebookLM | Answers only from your uploads, no web invention | Blind to anything you didn’t upload | Deep Research adds fresh web material |
| Deep Research (ChatGPT / Gemini / Claude) | Long, structured, multi-source draft | Confident but sometimes wrong or stale citations | Your Stage 4 spot-check catches the fakes |
| You | Judgment, “what’s actually news for the reader” | Slow; can’t read 40 sources fast | The first three stages did the reading |
The tools handle breadth and recall. You handle the part a model cannot fake: deciding what matters.
Tool tiers and limits (as of June 2026)
Pin these before you start so you don’t hit a wall mid-task. Limits move, so re-check the official pages linked below.
| Tool | Free tier | Paid tier | What unlocks |
|---|---|---|---|
| Perplexity | ~3 Pro Searches + 5 Deep Research per day | Pro $20/mo ($200/yr ≈ $16.67/mo) | Unlimited Pro Search, 20 Deep Research/day, model picker |
| NotebookLM | 50 sources/notebook, up to 100 notebooks, 500K words/source | Plus, bundled with Google AI Pro $19.99/mo | 300 sources/notebook, higher daily caps |
| ChatGPT Deep Research | 5 lightweight runs/month | Plus $20/mo (~25/mo) · Pro $200/mo (~250/mo) | More full-fat runs before the o4-mini fallback kicks in |
| Gemini Deep Research | Limited daily runs | Google AI Pro $19.99/mo | Deep Research + Deep Research Max (extended compute) |
| Claude Research | Not on Free | Pro $20/mo ($17/mo annual) | Multi-source Research mode, fewer hallucinated cites |
Sources: Perplexity pricing, ChatGPT Deep Research help, Google AI Pro plans.
Notice the overlap: Google AI Pro ($19.99/mo) gives you both NotebookLM Plus and Gemini Deep Research in one subscription, which is why it’s the most efficient single paid pick if you run this workflow often. Free tiers of Perplexity + NotebookLM + ChatGPT still cover most one-off projects.
Step by step
Every stage ships a copy-paste prompt. Replace the [...] placeholders with your real topic, then send to the named tool.
Stage 0 — Write the research question on one line
If you cannot state it in one sentence, split it into 1 primary + 2-3 sub-questions. Use this template:
Primary question: [verb + object + scope + time]
Example: "How do B2B SaaS content sites convert SEO traffic to newsletter signups in 2026"
Sub-questions (max 3):
1. [sub-question 1]
2. [sub-question 2]
3. [sub-question 3]
Deliverable: [e.g. "1500-word brief + 3 charts + 5 recommendations"]
Deadline: [e.g. "by Thursday afternoon"]
Audience: [e.g. "my co-founder + 2 investors, all SEO-literate"]If the primary question cannot be stated in one sentence, you cannot open AI yet — keep splitting. This single line is what you will paste into every later stage, so it earns its time.
Stage 1 — Landscape scan (Perplexity)
Open perplexity.ai and start a new thread. On Pro, turn on Pro Search and pick a strong reasoning model from the picker (Sonar for speed, or a frontier model like GPT-5.5 / Claude Sonnet 4.6 / Gemini 3.1 Pro for harder framing). Send:
I am researching: [primary question].
Scan how experts, leading practitioners, and academia map this:
1. List 5-7 main points of disagreement (not consensus points)
2. For each: position A (max 1 sentence) + position B (max 1 sentence) + 1-2 representative sources each (with links)
3. What has materially changed in the last 12 months (new data, events, tools)
4. Three "pseudo-consensus" points that are repeated often but rest on weak evidence
Only use sources from 2024 or later. End with a Markdown list of all cited links.Copy every citation link into a local file research_[topic]/sources.md with a date and a one-line annotation. Free tier caps you at roughly 3 Pro Searches per day, so make this prompt count rather than firing off five vague ones.
Stage 2 — Source synthesis (NotebookLM)
Pick the 8-15 most relevant sources from Stage 1, download each as PDF or save a web archive. Open notebooklm.google.com, create a New notebook, and upload them. Free notebooks hold up to 50 sources at 500K words each, which is more than enough here. In chat:
Based only on the uploaded 8-15 sources, answer:
1. Consensus: 3-5 facts or judgments every source agrees on
2. Disagreement: explicit conflicts between sources — for each, name which sources hold which side (use file names)
3. Gaps: information none of the sources covers but that is necessary to answer "[primary question]"
4. Data: every key number, growth rate, and market size that appears, each with a source citation
Do not use information outside the uploaded sources. Every claim must cite source filename + section.NotebookLM will not invent web facts — it answers only from your uploads, which is exactly why it belongs here. Use Add note to save the synthesis.
Stage 3 — Deep Research draft
Pick one agent: ChatGPT Deep Research (Plus gives ~25 runs/month before it falls back to a lighter o4-mini version), Gemini Deep Research (use Deep Research Max when you want maximum source coverage), or Claude Research (cleaner prose, fewer hallucinated citations). Paste the Stage 2 synthesis as context, then:
I have already done the landscape scan and source synthesis. Here are the findings:
[paste Stage 2 output]
Now produce a brief based on the above plus the latest material you can search for:
- Length: 1500 words
- Structure:
1. Executive Summary (200 words, 3 most important takeaways)
2. Current state (300 words)
3. Major disagreements and evidence (500 words, with sub-headings)
4. Data and trends (300 words + a key-numbers table)
5. 3-5 recommendations for [audience] (200 words)
- Every factual claim must carry an inline citation [n]
- End with a full citation list — link + publication date for each
- No pre-2024 sources as primary evidenceIf you run Stage 3 in ChatGPT specifically, pair this with the ChatGPT research workflow for Project setup and the question-challenge / weak-section re-research loop. If you’d rather stay in one tool with fresh web data and Workspace integration, the Gemini Deep Research workflow gives a ~45-minute path to a defensible brief.
Stage 4 — Citation spot-check
This is the stage people skip, and it’s the one that saves the brief. Open the Deep Research output, pick 3-5 inline citations at random, and for each:
- Click the link and confirm the page exists
- Cmd+F the quoted phrase in the original page and confirm it’s actually there (not invented)
- Flag it weak if: the original says the opposite, the source is a blog with no underlying data, or the source is more than 2 years stale
Even the better agents misattribute a fraction of citations, so treat every number as a claim to verify, not a fact. Re-run Deep Research for the section containing any weak citation using the same prompt plus “only use sources of type X” (e.g. primary filings, peer-reviewed, official stats).
Stage 5 — Human synthesis
Deep Research gives you a fact mosaic; “what counts as news for the reader” only you can write. Open a new doc, do not paste the AI executive summary, and write your own:
- “If the reader reads only 3 sentences, what should they take away?” Write those 3.
- Which one finding violates the reader’s current intuition? Put it first.
- 3-5 recommendations, each = verb + object + how to measure done.
Stage 6 — Archive
Create a local folder research_[topic]_[YYYYMMDD]/:
research_saas_newsletter_2026_05_21/
├── 00_question.md # primary + sub-questions (Stage 0 template)
├── 01_perplexity_scan.md # Stage 1 output + citations
├── 02_notebooklm_synth.md # Stage 2 synthesis
├── 03_deep_research.md # raw + redone version
├── 04_citation_audit.md # weak citations + reruns
├── 05_my_brief.md # your hand-written brief
└── prompts.md # every prompt used (next time, change placeholders)Next time you do a similar topic, duplicate the folder, change the name and 00_question.md, and you’re already 80% set up. The archive is what turns a 4-hour cold start into a 60-minute repeat.
Recommended flow at a glance
Question → landscape scan (Perplexity) → source synthesis (NotebookLM) → Deep Research draft → manual spot-check → human-written synthesis → archived bundle.
Common mistakes
- Skipping the one-line question and opening AI with a fuzzy goal
- Letting one tool do all 6 stages (you lose every cross-check above)
- Treating the Deep Research output as the final brief instead of a draft
- Skipping the Stage 4 spot-check and shipping a hallucinated citation
- Never archiving, so you redo this from scratch every time
FAQ
- How long does the full workflow take?: A first run takes 2-4 hours; once you have a saved template and folder, it drops to 60-90 minutes. Stage 3 (Deep Research) is the longest wait — most agents take 5-15 minutes per run.
- Do I need paid tools?: No. Free tiers of Perplexity (~3 Pro Searches + 5 Deep Research/day), NotebookLM (50 sources/notebook), and ChatGPT Deep Research (5 lightweight runs/month) cover most one-off projects. Pay only when research is weekly.
- Which single subscription gives the best coverage?: Google AI Pro at $19.99/mo bundles NotebookLM Plus (300 sources/notebook) and Gemini Deep Research in one plan, so it’s the most efficient paid pick for this exact workflow.
- Why not just use ChatGPT for all six stages?: ChatGPT can’t ground answers to a fixed source set the way NotebookLM does, and using one tool removes every cross-check that catches its mistakes. The whole point is that each stage audits the previous one.
- How accurate are the AI citations?: Treat them as unverified until Stage 4. Even strong agents misattribute or stale-link a share of citations, which is exactly why the spot-check is mandatory, not optional.