Study a Scientific Replication Attempt with Its Original Paper in NotebookLM

A source-grounded workflow for comparing an original scientific paper with a replication attempt without reducing the result to a simple success-or-failure label. The method below treats NotebookLM as a source-navigation and comparison aid, not as a substitute for domain judgment. Google’s current help documentation, checked on October 8, 2026, explains that imported material becomes a source used for answers, that users can select specific sources, and that inline citations can be opened in context. Those controls are useful only when the source set is deliberate and every important claim is reviewed.

This guide produces replication comparison. It assumes you have permission to use the documents and have removed personal, confidential, or restricted material. It also keeps the task distinct from general notebook setup: the objective is a reviewable evidence product for a scientific replication attempt and its original paper.

Define the claim before comparing results

Start by writing one sentence that identifies the exact effect, relationship, or procedure the replication tried to test. Do not begin with a broad question such as whether the original paper was right. A replication may target one experiment, one outcome, or one statistical path inside a much larger paper. Record the original sample, setting, measures, analysis, and reported result beside the replication target. The public OSF record for a replication of Stinson and colleagues, for example, identifies the original citation, the proposed mediating effect, recruitment frame, model, and planned fit criteria. That level of specificity creates a fair comparison. It also prevents a difference in one outcome from being misreported as a judgment on every claim in the original paper.

Build a source pack that preserves chronology

Create a small, deliberate source set: the original paper, its supplementary material, the replication protocol or registration, the replication report, and any correction or author response. Add a short source note containing publication date, version, and role. Put preregistration before results in the reading order so that planned decisions are not confused with explanations written after seeing data. If a source is a scanned PDF, check whether tables and formulas are readable after import. Keep publisher pages or repository metadata as separate sources when they establish dates or versions. The goal is not to upload every paper that cites the study. It is to preserve the evidence chain from original claim to planned replication and reported result.

Create a method-and-result comparison matrix

Ask NotebookLM for a table with rows for population, sample size, recruitment, exclusions, intervention or exposure, measures, timing, analysis, outcome definition, effect estimate, uncertainty interval, and missing-data treatment. Require a citation in both the original and replication columns. Add a third column labelled “comparison judgment” and allow only controlled values: matched, materially different, unclear, or not applicable. Inspect every citation. A shared variable name does not prove that the measures were administered in the same way. Likewise, a larger sample does not automatically make a design equivalent. The matrix should expose which differences could plausibly affect comparability without pretending to calculate a new meta-analysis.

Interpret disagreement without a binary verdict

Separate direction, magnitude, precision, and statistical threshold. A replication can estimate an effect in the same direction with a wider interval, a smaller magnitude, or a result that crosses a conventional threshold. Those are different patterns. Record whether the replication authors defined success in advance and whether the reported analysis matches that rule. Then distinguish direct replication from conceptual replication. If the population, task, or measurement changed, state what kind of generalization the study tested. Use language such as “the replication did not meet its preregistered criterion” instead of “the original was disproved” unless the evidence truly supports that stronger conclusion.

Use prompts that force source separation

Prompt in stages. First ask for the original claim with citations only from the original paper. Next select only the protocol and ask for planned deviations and success criteria. Then select the replication report and ask for observed results. Finally select all sources and request the comparison matrix. This sequence reduces source blending. Useful follow-ups include: “Which differences are documented before data collection?”, “Which explanations appear only after the results?”, and “Show the exact passage supporting each judgment.” Save verified answers as notes with names that encode their role, such as Original claim, Planned test, Observed result, and Open question.

Check failure modes before writing the conclusion

Watch for version mismatch, outcome switching, incompatible effect measures, inaccessible supplements, and citations that point to discussion text rather than methods or results. NotebookLM can help locate passages, but it does not decide whether two statistical models are equivalent. Do not infer missing analytic details from a summary. If the replication report depends on code or data that are not in the notebook, mark the comparison incomplete. Also avoid treating author prestige, journal name, or citation count as evidence about the replicated effect. The final comparison should include an uncertainty section naming every missing source or unresolved methodological difference.

Write a reviewable replication comparison

Use four outputs: a one-paragraph claim definition, the method matrix, a results comparison, and a bounded conclusion. Every important sentence should link back to a checked citation. End with a short list of what would change the conclusion, such as obtaining a missing supplement or confirming an exclusion rule. This structure makes the review useful to editors and students because it shows the path from evidence to language. It also provides a correction trail: if a source version changes, the affected row can be updated without rewriting the entire article.

A practical review sequence

Before drafting, open the source panel and verify that every file is the intended version. Google documents that Drive sources can update and that imported web pages may include only page text, while comments and footnotes from Google files may not be imported. Those limits matter whenever a conclusion depends on annotations, layout, or nested pages. Use a naming convention that includes date and version, select only the relevant sources for each question, and open the cited passage rather than trusting a summary.

Run three passes. The inventory pass asks what sources and fields exist. The comparison pass requests a table that keeps each source’s evidence separate. The challenge pass asks for contradictions, missing fields, boundary cases, and unsupported inferences. Save only checked findings. For a broader introduction to source selection and answer verification, read the NotebookLM research sources guide.

Frequently asked questions

Sources

• Google: Add or discover new sources for your notebook
• Google: Use chat in Gemini Notebook
• OSF replication record

Primary sources checked October 8, 2026. Product interfaces and limits can change; confirm the current help pages before relying on a specific capability.