TL;DR
- Feed 3-7 source documents plus your decision into a long-context model and ask for a fixed one-page structure: TL;DR, TAM/SAM/SOM, top trends, top risks, recommendation, open questions.
- As of June 2026, ChatGPT Plus ($20), Claude Pro ($20), and Google AI Pro ($19.99) all run 1M-token models, so a few hundred pages of reports fit in one pass. For verifiable citations, NotebookLM (free) is the safest because it refuses to answer outside your uploaded sources.
- The output is a draft, not the brief. Trace every number back to a source page before it touches a board memo or fundraising deck — models still hallucinate plausible market sizes.
The task
You are heading into a strategy meeting with five market reports, three analyst decks, and a folder of competitor blog posts. The decision-maker wants a one-pager: market size, top trends, top risks, and a recommendation. Reading everything cold is a full day. A focused AI workflow gets you to a defensible first draft in 30-45 minutes.
This is a recurring task for founders pre-fundraising, corporate strategy teams, product marketers entering a new segment, and investment associates running first-pass diligence.
Which tool, and why
Pick by how much you trust the numbers, not by brand. The split that matters: a chat model with a big context window is faster and more flexible for synthesis; NotebookLM is slower but every sentence carries a clickable citation back to your source.
| Tool (June 2026) | Price | Context / capacity | Best for | Watch-out |
|---|---|---|---|---|
| NotebookLM | Free (Plus tiers via Google AI Pro/Ultra) | 50 sources free, up to 500,000 words/source | Source-grounded synthesis with inline page citations | No academic-style references; refuses to extrapolate beyond sources |
| ChatGPT Plus (GPT-5.5) | $20/mo | ~320 pages in-app per chat (full 1M only on $200 Pro) | Fast synthesis, sharp prose, charts | Will confidently fill gaps unless you forbid it |
| Claude Pro (Sonnet 4.6 / Opus 4.7) | $20/mo | 1M tokens; PDFs read best under ~100 pages; 30MB/file | Careful multi-doc reasoning, nuance on disagreements | Projects only pull relevant chunks, not the whole library at once |
| Google AI Pro (Gemini 3.1 Pro) | $19.99/mo | 1M tokens (~1,500 A4 pages) | Largest single-pass corpus, native PDF/image reading | Verify any number not visibly quoted from a source |
For a one-pager built on a manageable corpus (under ~200 pages of text), a single 1M-token chat model handles the whole stack. When the brief feeds a high-stakes decision and every figure must be auditable, run the synthesis in NotebookLM, which ships responses backed by inline citations and refuses to answer when the answer is not in your sources. If you want a model-specific, source-anchored walkthrough first, the Gemini research tutorial covers grounded summaries end to end.
When not to rely on AI alone
Never put an AI number into a board memo or fundraising deck without tracing it to a source page. Models hallucinate plausible-looking market sizes, and a wrong TAM in a deck is a credibility hole you cannot patch live.
Skip pure AI for proprietary or specialized research — insider channel checks, expert-network calls, primary survey data — where the model has no training signal and will simply guess.
What to feed the AI
- The 3-7 source documents (text or text-extracted PDFs; keep each PDF under ~100 pages for reliable full-text reading on Claude)
- The decision you have to make, in one sentence
- Your time horizon (12 months, 3 years, a decade)
- Your geography and segment focus
- One or two priors you suspect are true and want challenged
The “priors to challenge” line is what sharpens the output. It turns the model into a devil’s advocate instead of an agreeable summarizer.
Copy-ready prompt
Paste this into ChatGPT, Claude, or Gemini, or use it as the chat prompt over a NotebookLM notebook. Replace each [bracketed] placeholder with your own value.
You are a strategy associate preparing a one-page brief.
Decision needed: [one_sentence_decision]
Time horizon: [months_or_years]
Geography + segment: [scope]
Priors to challenge: [priors_list]
Sources (each separated by ---):
[paste_sources]
Output:
1. TL;DR: 3 bullets, each starting with a verb the decision-maker can act on.
2. Market size: TAM / SAM / SOM with the exact number, year, and source
citation for each. If sources disagree, show the range, say which you
trust more, and why.
3. Top 5 trends: each with one supporting data point and its source.
4. Top 5 risks: each with severity (high/med/low) and a leading indicator
to monitor.
5. Recommendation: 1 paragraph that names the trade-offs.
6. Questions the sources do not answer: 3 bullets.
7. Confidence: high/medium/low per major claim.
Quote numbers verbatim with source page. Mark anything you infer with
"[inference]". If a number is not in the sources, write "not in sources" —
do not estimate.The last line is the load-bearing one. Without an explicit “do not estimate” instruction, GPT-5.5 in particular will fill a missing TAM with a confident guess.
The one-page layout
A literal single page: TL;DR at the top, market size and trends in the middle, risks and recommendation at the bottom, and “what’s missing” as a footer. The footer is where you scope the next round of research.
Every number carries its source citation inline. Make that a hard rule in the prompt, not a hope.
How to check the output
- Pick the three numbers that most affect the recommendation. Open the source PDFs and verify each one’s value, year, and methodology footnote. If a single one is wrong, audit the rest — one fabrication means the model was guessing, not reading.
- For each trend claim, search the source for the exact phrase the model quoted. Generic trends with no supporting data point get downgraded or cut.
- Sanity-check the math separately. Use AI to extract numbers; use a spreadsheet to validate ratios, units, and growth-rate periods before anything gets quoted. A “40% CAGR” with no stated window is not a number.
In NotebookLM this is faster: click any inline citation to jump straight to the source passage, which collapses verification into one click per claim.
Common mistakes
- Numbers without sources — looks authoritative, fails any diligence call
- A recommendation that quietly ignores the trade-offs
- Confusing TAM with SAM (and the model will happily go along)
- Quoting a growth rate with no time period
- No “what’s missing” section — over-confidence is the real failure mode here
Keep the brief alive
Save the brief, the sources, and the prompt together as one artifact. After the meeting, note where the brief drove the right call and where it misled you. Re-run the prompt each quarter with fresh sources — that turns a one-off summary into a moving picture of the market.
FAQ
- How many sources is too many? Past ~200 pages of text, pre-summarize each document by section first, then synthesize the summaries. A single 1M-token model can technically hold more (Gemini 3.1 Pro fits roughly 1,500 A4 pages), but recall on a buried figure degrades long before the window fills.
- Which tool gives the most trustworthy citations? NotebookLM, because it is source-grounded: it cites a specific passage for each claim and declines to answer when the source does not cover it, which is why lawyers and academics adopted it.
- Can AI rank sources by credibility? It will try, and newer, peer-reviewed, or primary-data sources usually win — but the model often does not know publication dates, so verify the ranking yourself.
- Plus or the free tier — does it matter for this? For a few hundred pages, ChatGPT Plus or Claude Pro at $20/mo is plenty; the $100-$200 tiers only matter when you need the full 1M-token in-app window or much higher daily limits.
Related
Pair the brief with deeper research summary prompts, benchmark against rivals using competitor analysis with AI, and organize the underlying notes through the market research organize AI workflow.
Tags: #Data analysis #Workflow #Research