NotebookLM can help you question and compare a collection of papers, but a useful literature review needs more than a summary. The workflow below keeps the original papers, your extraction decisions, and a checked evidence matrix connected. Google’s current help centre uses the name Gemini Notebook; this guide keeps NotebookLM in the title for readers searching under that name.
This is a suggested research workflow based on product documentation. The example question and blank matrix are tools to adapt, not findings from a completed review.
1. Define the review before opening a notebook
Write the question, the kind of review you are conducting, and the inclusion criteria. For example: “How do undergraduate students use feedback to revise argumentative writing?” Decide whether you need intervention studies, qualitative accounts, reviews, or a combination. Record your search strategy separately.
A notebook containing five selected papers is a reading collection. It should not be described as a comprehensive review of a field simply because the AI can summarise it. Keep the discovery and screening record outside the notebook.
2. Add a small, clearly labelled source set
Start with a few papers you have screened and are allowed to upload. Give each a stable label such as author, year, and short title. Google’s source documentation explains supported imports and their limitations. Check the imported text against the original, especially tables and material that may not transfer as expected.
Keep the original PDF and citation in your reference manager. Record which version you used, such as an accepted manuscript or version of record. If a source changes later, you need to know which version supported the entry in your matrix.
3. Build the evidence matrix
Copy the field list below into a document or spreadsheet, with one row per study. The blank structure is intentionally more important than a polished AI narrative: it shows what you checked and where the evidence came from.
- Study ID and citation: a stable label and full reference.
- Population and setting: who participated and where.
- Design: how the study was conducted.
- Intervention or phenomenon: what was examined.
- Outcome or theme: what was measured or interpreted.
- Reported finding: a careful paraphrase.
- Evidence location: page, section, table, or source passage.
- Limitations: author-reported limits and your separate appraisal.
- Verification: unchecked, checked, or needs clarification.
Blank row to copy: Study ID: ___ | Population: ___ | Design: ___ | Focus: ___ | Outcome: ___ | Finding: ___ | Evidence location: ___ | Limitation: ___ | Checked by/date: ___
4. Extract one paper before comparing many
Select one source and ask: “For this paper, identify the population, design, outcome measure, reported findings, and limitations. Cite the supporting passage for each field. Use ‘not reported’ when the information is absent. Keep the authors’ statements separate from interpretation.”
Open each citation and read around it. Google’s chat documentation explains source selection and navigating citations. The presence of a citation helps you investigate an answer; it does not remove the need to check the claim.
Correct the row yourself before using the same fields across more papers. If the AI confuses a sample characteristic with an outcome, make that distinction explicit in the next prompt.
5. Compare findings without erasing differences
After checking individual rows, ask: “Compare these studies by population, method, and outcome. Identify where findings appear similar and where the studies are not directly comparable. Cite the sources. Do not count agreement without considering what each study measured.”
Two studies can use the word “feedback” for very different activities. One may examine teacher comments over a semester; another may study a single automated response. Preserve those differences in the synthesis instead of combining them into a broad claim that feedback works.
6. Write from the checked record
Organise the review around a question or theme, then bring studies into that argument. Keep factual claims traceable to the checked matrix and original papers. Separate what the evidence indicates, what remains uncertain, and your interpretation.
A useful paragraph plan is: claim, supporting studies, an important difference or limitation, and the implication for your question. Draft the argument in your own words. If AI helps reorganise it, compare the revised text with the source record before accepting it.
Common problems and fixes
- A source is missing: check the import and selected-source list.
- A field is filled with a guess: replace it with “not reported” until you verify it.
- A citation supports only part of a claim: narrow the claim or find additional evidence.
- Different outcomes are merged: create separate columns or themes.
- The summary sounds too certain: restore the study’s design limits and qualifications.
Keep your search log, screening decisions, original papers, and final matrix as the durable research record. NotebookLM can assist the reading process; the review method and conclusions remain yours to justify.
Related guides and further reading
Explore more Research Tools and AI in Education.
