You have 12 PDFs, three competitor white papers, and an analyst report. The boss wants “the across-document picture by Friday.” Per-document summaries — the cheap version of this work — answer “what does Document 5 say?” but not “where do these 16 sources agree, disagree, or stay silent?” This tutorial walks researchers, strategy analysts, and graduate students through a four-question synthesis workflow that produces a credible cross-source brief in a working day, with citations that trace back to each source.
What this covers
A four-question synthesis pattern (consensus, disagreement, gap, recency) plus the tooling choice (NotebookLM versus Projects-style chats), source labeling discipline that keeps citations traceable, and the spot-check loop that catches AI’s tendency to invent agreement that does not exist. The output is a synthesis doc with per-claim citations you can defend.
Who this is for
Researchers, analysts, consultants, journalists, and students with five or more documents that need to be synthesized. Strategy teams comparing competitor positioning. Investors triangulating between management decks and analyst reports. Policy writers cross-checking academic literature against government statements. If your work involves a phrase like “across the literature,” this is the workflow.
When to reach for it
When per-document summaries are not enough — you need to compare across sources. (If the per-document summary itself is the goal — say, walking into tomorrow’s journal club having only read one paper — use the 10-minute research-summary workflow instead.) Also useful when you suspect sources disagree but cannot pinpoint the disagreement, or when you need to know what is missing across an entire literature.
When this is NOT the right tool
Two documents — just read them side by side. Hundreds of documents where the answer is statistical not interpretive — use a proper bibliometric tool. Highly sensitive sources (legal, classified, medical records) where upload is forbidden — synthesize by hand. Same-language sources only is not a requirement; multi-language works with one extra step.
Before you start
- Decide your “source tier” rules up front. Peer-reviewed paper, gray-literature report, blog post, internal memo — these are not equal evidence. Tag the tier and tell the model to weight accordingly.
- Pick the tool. NotebookLM for 5-50 docs with built-in citation surfacing; Claude Projects or ChatGPT Projects for fewer docs with a more conversational style.
- Standardize source labels. “Smith 2024” or “Internal Q3 Deck” — readable, unique, and short enough to inline in citations.
- Have your synthesis question written down before opening the tool. Vague questions produce vague syntheses; the structure of your question is the structure of your output.
Step by step
- Pick the right tool. NotebookLM for 5 or more docs (citations are first-class, retrieval is grounded). ChatGPT Projects or Claude Projects for 5 or fewer docs (conversation is smoother, but you must enforce citation discipline by prompt).
- Upload all sources. Give each source a stable label, ideally the same one you will use in your final doc. NotebookLM uses filenames; rename before uploading if necessary.
- Ask for consensus. “Where do all sources agree on [topic]? List each agreement as a bullet with a citation to every source that supports it. Do not invent agreement — only points all sources actually make.” The “do not invent” line is essential; models love to round disagreement into “broad consensus.”
- Ask for disagreement. “Where do sources disagree? For each disagreement, name which sources take which side, and quote one line per source.” Quotations let you spot-check; bare claims do not.
- Ask for gaps. “What does no source address? What questions are raised but not answered? What is implied but never argued?” The gap question is where synthesis becomes original — it tells you what to research next.
- Ask for recency. “Which claims have been superseded by later sources? Where does the older view differ from the newer?” Critical for fast-moving fields.
- Compile the output into a synthesis doc. Carry over source labels verbatim so citations remain traceable. Spot-check three citations by opening the source and finding the supporting text.
First-run exercise
- Pick five documents on a topic you already know something about — your own domain, or a recent project. You will catch the model’s invented agreements immediately.
- Run the four-question sequence. Save each response separately. Total time should be under 30 minutes.
- Spot-check three citations per response. Count: how many citations are accurate? How many invented? How many directionally right but wrong page?
- For the second run, change only one variable: a stricter quotation requirement, a different model, or an additional source tier.
Quality check
- Every claim in the synthesis has at least one citation. Unsupported claims either come from the model’s training data (not your sources) or are inventions.
- Spot-check ratio: for every 10 citations, verify 2. Below 80% accuracy, the synthesis is unsafe to ship without a full audit.
- The disagreements have specific quotes, not paraphrases. Paraphrased disagreement is where AI smooths real conflict into “different framings.”
- Source tiers are reflected. If a peer-reviewed paper and a blog post disagree, the synthesis should say so — not present them as equal.
- The gaps section has at least three items. Fewer means the model padded the consensus and disagreement sections instead.
How to reuse this workflow
- Save the four-question prompts as a template. New project, new sources, same questions.
- For recurring research (quarterly competitor landscape, weekly literature tracking), maintain a synthesis doc and re-run the workflow each cycle. Diffing syntheses reveals shifts in the field.
- Build a citation hygiene habit. Keep source labels stable across projects so old syntheses remain readable years later.
Recommended workflow
Choose tool by source count → upload labeled sources → consensus question → disagreement question → gap question → recency question → compile synthesis with verbatim citations → spot-check 20% of citations → final edit for narrative.
Common mistakes
- Mixing source quality without flagging (peer-reviewed and blog post treated as equal evidence).
- Asking “summarize” instead of “where do they agree, disagree, leave gaps.” Summary collapses the cross-document structure that synthesis needs.
- Losing citation labels in the final synthesis. Once labels drop, the synthesis is unverifiable.
- Trusting “all sources agree” claims without spot-check. The model overstates agreement to sound confident.
- Uploading documents but not telling the model their tier. Without tier, the model treats a Substack post like a Nature paper.
- Stopping at consensus. The original signal is in disagreement and gaps — that is where your contribution lives.
FAQ
- NotebookLM or Claude/ChatGPT Projects?: NotebookLM for 5+ docs with strong citation needs. Projects for fewer docs and more conversational synthesis.
- What if sources are in different languages?: Translate or summarize each first into a common language, then synthesize. Or use NotebookLM, which handles multilingual decently.
- How many sources is the upper limit?: NotebookLM handles 50 sources fluently; quality degrades above 100. For larger corpora, cluster first and synthesize per cluster.
- Can I trust the AI’s “no source addresses this” claims?: Only after spot-check. Models sometimes miss a source that does address the topic.
- What about contradicting sources I disagree with?: Include them. Synthesis is most useful when it surfaces views you disagree with, with reasoning.
- How do I handle paywalled sources?: If the model cannot read them, treat them as out of scope. Do not let it speculate about content it cannot see.
Related
- NotebookLM getting started
- Literature review prompts
- AI industry research workflow
- AI paper reading workflow
Tags: #Tutorial #Research #Long document