NotebookLM can help you question, compare, and organize a collection of research papers. What it cannot do is take responsibility for your literature review.
That distinction matters. A literature review involves more than producing summaries. You need to select appropriate studies, understand how they were conducted, assess the strength of their evidence, compare their findings, and decide what conclusions the literature supports.
NotebookLM can assist with parts of this process because its responses are grounded in the sources you upload. It can retrieve information from papers, compare studies, and attach citations that take you back to relevant passages. These features make it useful for building an evidence matrix.
The workflow below keeps the original papers, your extraction decisions, and the final synthesis connected. It also gives you several points at which to check the AI’s work before relying on it.
Google’s current help centre uses the name Gemini Notebook for the product previously known as NotebookLM. I will continue using NotebookLM in this guide because that is still the name many researchers recognize and search for.
What Is NotebookLM?
NotebookLM is a source based AI research tool developed by Google. You create a notebook, add documents, and ask questions about those materials.
Unlike a general chatbot that may draw on broad model knowledge, NotebookLM is designed to answer from the sources contained in the notebook. Its responses include citations that connect statements with passages in those sources.
Depending on the current version and account, NotebookLM can also generate reports, notes, mind maps, audio overviews, video overviews, and other source based outputs. These features can help with orientation, but the chat and citation tools are more important for the evidence matrix workflow.
For literature reviews, I use NotebookLM as a reading and comparison assistant. I do not use it to decide which studies belong in the review or to write conclusions that I have not independently checked.
What Is an Evidence Matrix?
An evidence matrix is a table in which every row represents a study and every column represents a category of information you want to extract.
The categories might include the research question, participants, setting, methods, findings, limitations, and relevance to your review.
| Study | Purpose | Sample | Method | Main findings | Limitations |
|---|---|---|---|---|---|
| Study 1 | |||||
| Study 2 | |||||
| Study 3 |
This structure forces you to examine the same questions across all included studies. It also makes similarities, disagreements, and gaps easier to see.
The concept is well established in research writing and evidence synthesis. Florida International University describes a synthesis matrix as a way to organize the arguments presented across sources. Johns Hopkins University similarly recommends using a matrix to record the main points of sources and identify relationships among them.
In systematic and scoping reviews, a more formal version of this process is usually called data extraction or data charting. The University of North Carolina Libraries describes it as collecting predefined information from included studies in a table.
Before You Open NotebookLM
Do not begin by uploading a folder of papers and asking AI to write the literature review. Begin with the review itself.
Write down the question you are trying to answer, the boundaries of the review, and the information you need from each study.
| Decision | Questions to answer |
|---|---|
| Review question | What exactly am I investigating? |
| Population | Who or what is being studied? |
| Context | Which settings or educational levels are relevant? |
| Phenomenon | What intervention, practice, technology, or experience matters? |
| Outcomes | What kinds of results am I looking for? |
| Study types | Which research designs will I include? |
| Date range | Which publication years are relevant? |
| Exclusions | What will I leave out, and why? |
If you are conducting a formal systematic review, you will need a documented search strategy, inclusion criteria, screening process, and reporting method. PRISMA 2020 provides reporting guidance for systematic reviews, including how studies were identified, selected, appraised, and synthesized.
NotebookLM does not replace these procedures. It enters the workflow after you have found, screened, and selected the papers.
Step 1: Search for Studies Outside NotebookLM
Search scholarly databases that fit your discipline. In education, these might include ERIC, Scopus, Web of Science, PsycINFO, Education Source, or Google Scholar.
Keep a record of:
| Search information | Example |
|---|---|
| Database | ERIC |
| Date searched | September 14, 2026 |
| Search terms | “generative AI” AND “teacher education” |
| Filters | Peer reviewed, 2023 to 2026 |
| Results returned | 142 |
| Results retained | 18 |
NotebookLM should not be your sole literature search tool. Its source discovery feature may help locate related materials, but a reproducible review requires searches conducted through appropriate academic databases.
For a formal review, retain your search strings and screening decisions. PRISMA requires transparent reporting of how records were identified and selected.
Step 2: Finalize the Papers Before Uploading Them
Remove duplicates, screen the titles and abstracts, and apply your inclusion and exclusion criteria.
Upload only the papers that have reached the relevant stage of your review. Mixing included studies, background readings, opinion pieces, and excluded papers in the same notebook can make later comparisons difficult.
Rename each file before uploading it. A consistent naming system makes citations easier to follow.
A simple format is:
Author_Year_ShortTitle.pdf
For example:
Kasneci_2023_ChatGPT_Education.pdf
Also check that the PDF contains searchable text. If the document is a scanned image, perform optical character recognition before uploading it. Otherwise, NotebookLM may have difficulty reading or citing its contents accurately.
Step 3: Work in Small Batches
NotebookLM can accommodate larger source collections, but capacity is not the same as clarity.
From my own use of the tool, I prefer to work with no more than ten papers in a notebook when conducting detailed extraction. This is not the platform’s source limit. It is a practical choice that makes the responses and citations easier to inspect.
You can organize notebooks by:
| Option | Example |
|---|---|
| Review theme | AI feedback studies |
| Study design | Qualitative studies |
| Population | Preservice teachers |
| Review stage | Extraction batch 1 |
| Publication period | Studies published from 2023 to 2024 |
If your review contains 40 studies, you might create four extraction notebooks with ten papers each. Keep the final evidence matrix outside NotebookLM so all verified entries eventually come together in one place.
Step 4: Design the Matrix Before Asking Questions
Decide what you need to extract before prompting NotebookLM. Your fields should follow directly from your review question and planned analysis.
Here is a general evidence matrix for educational research:
| Bibliographic information | Study information |
|---|---|
| Full APA reference | Research purpose |
| DOI or stable URL | Research question |
| Country | Theoretical framework |
| Publication year | Educational setting |
| Methods | Evidence |
|---|---|
| Research design | Main findings |
| Participants and sample size | Reported effect sizes |
| Data collection | Authors’ interpretations |
| Data analysis | Limitations |
| Intervention or exposure | Relevance to the review |
Add fields that matter to your topic. A review of AI tutoring might include the AI model, comparison condition, duration, subject area, and type of learning outcome. A review of teacher experiences might include teaching level, professional background, and analytic approach.
Avoid asking the AI to decide what is important after extraction has started. Changing the columns halfway through can produce inconsistent entries.
Step 5: Extract One Study at a Time
Select one source in NotebookLM and ask the same structured questions for every paper.
A useful extraction prompt is:
Using only the selected source, complete the fields below. If the paper does not report an item, write “Not reported.” Do not infer missing details. For every answer, provide a citation to the exact supporting passage.
- Full reference
- Research purpose
- Research question
- Country and setting
- Study design
- Participants and sample size
- Data collection methods
- Data analysis methods
- Main findings
- Limitations reported by the authors
- Relevance to this review question: [insert question]
The phrase “using only the selected source” is important when a notebook contains several papers. It reduces the risk that information from one study will be blended with another.
The instruction not to infer is equally important. There is a meaningful difference between “not reported” and an AI generated guess.
Step 6: Verify Every Extracted Entry
NotebookLM’s citations make verification easier, but a citation is not proof that the interpretation is correct.
Open each citation and compare the response with the original passage. Ask:
| Verification question | What to check |
|---|---|
| Does the passage support the claim? | Direct agreement between answer and source |
| Is the information about this study? | No confusion with cited background research |
| Is the number correct? | Sample size, percentage, effect size, or duration |
| Is the result qualified? | Conditions and uncertainty have been preserved |
| Is it a finding or interpretation? | Results are separated from the authors’ discussion |
| Is the limitation reported or inferred? | Your judgment is labelled separately |
Pay close attention to research design and causal language. A correlational study does not demonstrate that one variable caused another. A small qualitative study may provide detailed insight without supporting claims about prevalence.
Research on AI assisted data extraction supports this caution. A feasibility study by Schmidt and colleagues found useful performance in some extraction tasks, but accuracy varied by domain and causal inference methods were among the areas where errors occurred. The authors concluded that human verification remains necessary. Read the study.
Step 7: Transfer Verified Information to an External Matrix
Do not leave the only copy of your evidence inside NotebookLM.
Transfer each verified entry to a spreadsheet or reference management system. Include a location marker so you can return to the evidence.
| Study ID | Finding | Source location | Verification |
|---|---|---|---|
| S01 | Students reported improved confidence | p. 12 | Verified |
| S02 | No significant difference in test scores | p. 8, Table 3 | Verified |
| S03 | Teachers raised privacy concerns | p. 17 | Verified |
You can use Google Sheets, Excel, Airtable, Notion, or specialist review software. The tool matters less than having one stable, auditable record.
For systematic reviews, stronger procedures are required. Cochrane recommends independent duplicate extraction of outcome data because this reduces errors and limits the influence of one reviewer’s decisions. Cochrane data collection guidance.
NotebookLM should not silently replace a second reviewer.
Step 8: Compare Studies Across the Matrix
Once the individual entries are checked, NotebookLM can help you look across studies.
A comparison prompt might read:
Compare the selected studies in relation to the review question. Organize the answer under areas of agreement, disagreement, differences in study design, differences in population or setting, and unanswered questions. Cite every statement. Do not treat the number of studies supporting a claim as evidence of its quality.
You can also ask narrower questions:
| Analytical purpose | Prompt |
|---|---|
| Compare methods | How do the studies differ in design, sample, and data collection? |
| Find agreement | Which findings appear across more than one study? |
| Find disagreement | Which studies reach conflicting conclusions? |
| Examine context | Do findings differ by country, grade level, or subject? |
| Assess evidence | Which claims rely on experimental evidence? |
| Locate gaps | Which populations or outcomes remain understudied? |
Do not ask only, “What are the main themes?” That prompt often produces broad categories that sound plausible but hide important differences among studies.
Step 9: Build a Synthesis Matrix
An evidence matrix organizes information by study. A synthesis matrix reorganizes that evidence by theme, concept, or claim.
| Theme | Study 1 | Study 2 | Study 3 | Emerging interpretation |
|---|---|---|---|---|
| Student engagement | Increased | Increased | Mixed | Engagement often improved, but not uniformly |
| Learning outcomes | Not measured | Small improvement | No difference | Evidence remains limited |
| Teacher workload | Reduced | Not examined | Reduced initially | Possible benefit, but context matters |
| Privacy concerns | Reported | Reported | Not examined | Recurring concern across settings |
This shift is what moves you from summarizing one paper after another to writing an actual literature review.
A paragraph based on synthesis brings studies into conversation. It explains where they agree, where they differ, and how their methods or contexts may account for those differences.
Step 10: Draft From the Matrix
Write from the verified matrix and return to the papers as needed. Do not draft only from an AI generated overview.
A useful paragraph structure is:
- State the pattern or claim.
- Identify the studies that support it.
- Explain meaningful differences among those studies.
- Discuss the strength or limitations of the evidence.
- Connect the paragraph to the review question.
NotebookLM can help you test your draft against the source collection:
Review this paragraph against the selected sources. Identify any claim that is unsupported, overstated, missing an important qualification, or inconsistent with the cited study. Do not rewrite the paragraph.
This keeps the researcher in control of the writing while using AI for a focused audit.
Step 11: Conduct a Final Claim Audit
Before submitting the review, examine every important factual or interpretive claim.
| Claim type | Final check |
|---|---|
| Descriptive claim | Does the study report this information? |
| Numerical claim | Does the number match the paper? |
| Causal claim | Does the design support causal inference? |
| Comparative claim | Were the studies genuinely comparable? |
| Generalization | Does the sample support the stated scope? |
| Gap claim | Was the literature search broad enough to support it? |
| Citation | Does the cited paper support this exact sentence? |
Also verify every bibliographic reference using the original publication record, Crossref, the journal website, or your library database. Do not rely on an AI generated reference list without checking it.
What NotebookLM Can and Cannot Do
| NotebookLM can help with | NotebookLM should not control |
|---|---|
| Locating passages in uploaded papers | Defining the review question |
| Producing initial study summaries | Selecting inclusion criteria |
| Extracting reported study information | Deciding which studies to include |
| Comparing methods and findings | Assessing evidence without verification |
| Generating questions about the literature | Making final interpretive judgments |
| Linking responses to source passages | Writing claims you cannot defend |
| Auditing a draft against sources | Taking responsibility for the conclusions |
Privacy and Copyright
Research documents can contain unpublished findings, identifiable participant information, confidential institutional materials, or copyrighted content.
Before uploading a source, check whether you have permission to place it in an external AI system. Do not upload confidential interview transcripts, identifiable participant data, embargoed manuscripts, or materials restricted by an agreement unless your institution has approved that use.
Google’s current Gemini Notebook help information states that data handling depends partly on the account type. For Google Workspace and Workspace for Education users, uploads, queries, and model responses are not used to train AI models or reviewed by human reviewers. Institutional settings and agreements can still differ, so consult your university’s data governance policy.
Copyright also matters. Uploading a paper for personal scholarly analysis is different from sharing a public notebook containing copyrighted publications. Keep your notebooks private unless you have the right to distribute every included source.
Final Thoughts
NotebookLM is most useful for a literature review when it supports a method you have already defined. It can help retrieve information, compare studies, and connect claims with passages in the original papers. The evidence matrix gives that work a consistent structure.
The essential sequence is straightforward:
Search externally → select studies → design the matrix → extract one paper at a time → verify every entry → compare across studies → write from the checked evidence
The verification steps are what make the workflow useful. Without them, NotebookLM may simply help you produce a polished summary more quickly. With them, it becomes a practical research assistant that helps you stay close to the evidence while keeping the interpretation and final judgment in your hands.
