How Steven Johnson researches and structures books with NotebookLM
Steven Johnson
Bestselling author Steven Johnson keeps about 7,000 research quotes in one NotebookLM notebook and asks it plain questions while he writes, so he can find the exact fact he needs without breaking his flow. He also used NotebookLM to compare Gold Rush sources side by side, which surfaced a person who became a possible opening for his next book.
Story written by: the tool's company
The problem
Johnson writes long nonfiction built on huge amounts of research, and digging out one quote or fact could pull him out of his writing. He also wanted a way to test early book ideas against his sources before committing to them. He needed a tool that worked from his own collected material, not general internet answers.
How it was built
- 1Load your research into one notebookNotebookLM
- 2Set up the draft for fast accessGoogle Docs
- 3Ask plain questions to find exact factsNotebookLM
- 4Tell the tool who you areNotebookLM
- 5Ask what one set of sources adds over anotherNotebookLM
- 6Follow up on what catches your eyeNotebookLM
- 7Stress-test a structure before committingNotebookLM
Tool step Person step
1. Load your research into one notebook
Johnson keeps about 7,000 quotes from more than 20 years of research as sources in one NotebookLM notebook. For his Gold Rush idea he added excerpts from H.W. Brands' The Age of Gold (highlighted passages imported with ReadWise), Lafayette Bunnell's Discovery of the Yosemite, and two public-domain accounts of the Indigenous people of the valley.
ToolNotebookLM
Check: He had already read the Brands and Bunnell books in full before importing, so he could judge the tool's answers against what he knew.
2. Set up the draft for fast access
He drafted The Infernal Machine entirely in Google Docs, with one tab per chapter, so he could jump between chapters and drop research quotes into future chapters as they came to him.
ToolGoogle Docs
3. Ask plain questions to find exact facts
While writing, he asked NotebookLM natural-language questions like "What was the name of that restaurant where Berkman and Goldman met?" to pull the exact quote or fact from his sources without leaving the writing flow.
ToolNotebookLM
Check: Answers come with citations that link back to the original passage, so he can confirm the quote in the source.
4. Tell the tool who you are
For the Gold Rush brainstorm, Johnson started by telling NotebookLM he was Steven Johnson, author of multidisciplinary history books, so it understood the kind of book he wanted to write.
ToolNotebookLM
5. Ask what one set of sources adds over another
He asked NotebookLM what the two Indigenous-perspective sources contained that the other two sources lacked, and the answer pointed out that one source gave short biographies that humanized individuals the other books treated as a mass.
ToolNotebookLM
Check: He judged the comparison himself against books he had already read cover to cover.
6. Follow up on what catches your eye
A name in the NotebookLM answer, Maria Lebrado, caught his eye, so he asked for more and got nearly 600 words of biography. Her late-life return to the valley suggested a possible opening scene and structure for the book.
ToolNotebookLM
Check: He decided for himself whether Lebrado would work as a major character, and says he is still not sure.
7. Stress-test a structure before committing
He asked NotebookLM to flesh out what events from his sources would fill each chapter of a "long zoom" idea that starts a million years before the Gold Rush and closes in step by step. The tool drafted possibilities from his source timeline rather than a final outline.
ToolNotebookLM
Check: He treats the output as possibilities and evaluates what works on his own; he notes the model leans too supportive, so he asks for more skeptical criticism when he wants it.
What made it work
- He curated the source set himself, so answers come from material he chose, not the model's average response
- He told the tool who he is and what kind of book he wants before asking questions
- He had already read the core books in full; the tool helps him probe, not replace reading
- He uses it to generate possibilities, then judges them himself
- Fast fact lookup let him stay in the flow of writing
Follow along
Practice steps suggested by Ai Siftly, based on the account above. We have not run them ourselves.
Pick two short sources and one question
Choose two short documents you have permission to upload and label them Source A and Source B, with title, author and date. Read both first, then pick one question both discuss.
Example to copy: How do Source A and Source B explain [same event]?
You should get: Two sources and one question.
Check: You have permission to upload both documents.
Add the sources and confirm they loaded
Import both documents into the notebook tool, then open each imported source and compare the beginning, middle and end with the original. A file name alone does not prove the text loaded.
You should get: A notebook with readable sources.
Check: Open each loaded source and confirm the full text came through.
Test with a question you can already answer
Ask a factual question you can confirm from your own reading, then open the passages the tool points to.
Example to copy: What reason does Source A give for [event]? Point to the passage; if it is not there, say Not found.
You should get: An answer linked to a source, or a not-found reply.
Check: The cited passage actually supports the answer.
Compare the two sources in a table
Ask for a table of claims, the passages behind them, and where the sources agree or differ. Open the original text and confirm at least one agreement and one difference; if the sources agree, do not force a difference.
Example to copy: Table: claim | A passage | B passage | agreement or difference | unanswered question. Do not fill gaps from outside knowledge.
You should get: A comparison table you can trace back to the sources.
Check: Confirm any claimed agreement or difference against the originals; if there is none, record that.
Build your own outline
Pick the important checked findings yourself, then ask the tool for a possible structure where each point has a supporting source.
Example to copy: Propose an outline from these checked findings. Each point needs a supporting source. Add no facts. [findings]
You should get: An outline with an evidence table.
Check: Read the sources again before publishing; the structural decisions stay yours.
Try it yourself
Find a matching workflow in the workflow finder. Search words: A I n o t e b o o k c o m p a r e r e s e a r c h s o u r c e s.
Sources and notes
- Interview: How bestselling author Steven Johnson wrote his latest book with Google Docs and an AI-powered notebook (Google Workspace Blog, 2024-05-29)
- Machine Readable (Adjacent Possible (Steven Johnson), 2025-06-25)
- All results here are self-reported: the 2024 interview is published on Google's own blog, and Johnson is Editorial Director at Google Labs, which helped build NotebookLM.
- The 7,000-quote notebook ran on an early 2024 version of NotebookLM, and the interview and the Substack post describe two separate projects.
- No time or cost figures are given; the outcomes are his own account of how the work felt.
- The follow-along guide is a suggested practice from Ai Siftly, not Johnson's exact process. Have ready: two to four short sources you may upload, a narrow question you can answer by reading them, and a notebook tool that imports documents.
Interview also describes a hypothetical (timeline from scanned articles 'I would be able to'), excluded here Qualitative only Employee of the vendor. Date note: 2024-05-29 interview; 2025-06-25 Substack post. Record dated 2026-10-27.