How to Turn Research PDFs into a Comparison Table with NotebookLM

You have eight papers open and keep forgetting which one used interviews, which one followed participants for a year, and which one never measured the outcome you care about. This is exactly the kind of mess a comparison table can fix.

NotebookLM, now called Gemini Notebook, can build a data table from uploaded PDFs and export it to Google Sheets. The useful part is not the automatic table. It is choosing columns that force the papers to answer the same questions.

Start with the decision the table needs to support. If you are deciding which papers deserve a close read, you may only need population, method, outcome, and main finding. If you are preparing a literature review, you will probably want more detail.

For a set of studies on remote work and employee well-being, I would begin with these columns:

That last column matters. A table becomes much easier to trust when you can jump from a neat cell back to the paragraph or table it came from.

Avoid starting with vague columns such as “Methodology” or “Insights.” They invite equally vague answers. “How was well-being measured?” gives the tool a concrete extraction job.

Put the papers for one comparison task in the same notebook. Give the files recognizable names, then select the PDFs you want included. In the Studio panel, choose Data Tables and customize the table by describing the rows and columns. Google's Gemini Notebook help confirms that you can specify the table structure and export the result to Google Sheets.

A prompt for the remote-work example could be:

“Not reported” is one of the most useful instructions in the prompt. Without it, a clean table can hide the difference between information found in the paper and information the model guessed from context.

If the first result is too broad, do not keep adding columns to the same prompt. Generate a small table first, check it, then add the fields that matter. Ten useful columns beat twenty columns filled with paraphrases.

Decide what each column means before generating the table. “Study design” might contain cross-sectional survey, randomized trial, or qualitative interviews. “Outcome measure” should name the actual scale, test, or variable rather than saying “well-being was assessed.”

Some papers will not fit neatly. One study may report results separately for managers and employees; another may combine everyone. Keep that difference visible. Add a note in the cell rather than forcing both papers into the same label.

This is also why one row should usually represent one study, not one PDF. A paper can report several studies, and the same study can appear in more than one publication. If the unit is unclear, ask the notebook to identify the study before it fills the row.

Check the cells that could change your interpretation of the study:

Open the cited passage and compare it with the cell. Pay extra attention to tables and figures. A generated answer may capture the surrounding explanation while missing a subgroup label, footnote, or denominator inside the table.

Do a spot check on at least one easy paper and one awkward paper. The awkward paper is more revealing: mixed samples, several time points, or multiple outcome measures expose whether your column definitions are holding up.

First, search the PDF for the field yourself. The information may use a different name or appear in a supplement. If it really is absent, keep “not reported.” Missing information is a finding about the source, not an empty cell that must be filled.

Sometimes the gap tells you that your paper set is too thin. Suppose three studies use self-reported well-being and none uses clinical measures. You may want to search for studies using a different outcome, then run the same comparison fields across the new papers.

This is where WisPaper Search can continue the workflow. Search for the missing combination in natural language, save promising papers to My Library, and ask Library QA the same questions you used as table columns. Now your search and comparison are using the same criteria instead of drifting apart.

Use it to spot patterns, not to write the paragraph for you. Sort the studies by design, population, or outcome measure. You may discover that the papers appear to disagree because they studied different groups, or that a confident conclusion rests on one small sample.

Then go back to the papers behind the pattern. A useful synthesis sentence names the relationship you found: “Studies using weekly well-being scales reported short-term improvements, while the two longitudinal studies found no durable change.” The table helped you see that split; the papers provide the evidence for writing it.

Gemini Notebook can export a data table to Google Sheets, with citations placed in a separate tab. Changes made in the exported sheet do not sync back to the notebook, so treat the sheet as your working copy. Add your own screening decisions and notes there.