Summarize Without Losing the Argument

Ask any AI tool to summarise a paper and you reliably get the findings. What goes missing is everything that made the findings mean something: the sample the authors worked with, the conditions they said their result holds under, the hedge in the discussion section, and the position they were arguing against in the first place.

That loss is invisible until you cite the paper and someone who read it asks you a question. The fix is not a better model. It is asking for the argument explicitly, in a form the tool cannot satisfy by listing conclusions.

Six things a paper contains that a default summary throws away

Before touching a tool, know what you are protecting. A research argument has parts, and a summary is lossy in a predictable direction:

• The claim survives. It always does.
• The warrant, meaning why the evidence supports the claim rather than something weaker, usually goes.
• Scope conditions go. "In adults over 65 with no prior diagnosis" becomes "in adults".
• The authors' own hedge goes, because hedges read as filler to a summariser optimising for confidence.
• The counterposition goes. Papers are written against something, and summaries rarely name what.
• Numbers lose their units and their intervals. A point estimate arrives without the confidence interval that made it interesting or unremarkable.

Every prompt below exists to force one of those six back into the output.

Five places a summary appears, and only two carry clickable citations

Gemini Notebook, the tool Google renamed from NotebookLM on 16 July 2026, produces summaries in several places that look interchangeable and are not. The differences decide whether you can defend a sentence later.

Sources for those figures: the free-tier limits come from Google's Upgrade Gemini Notebook table and its FAQ, which states free accounts "Have daily limits of 50 chat queries and 3 audio generations." The 1,000-note figure is from the notes documentation: "You can create up to 1,000 notes per notebook."

The citation preservation is the part worth building your habit around. Google documents it precisely: "When you save a response as a note, the original format, including tables and clickable inline citations, gets saved." So the durable summary is a chat answer saved with Save to Note. A Report might be the prettier object, but the help centre never says it carries citations, and you cannot defend what you cannot click.

The prompt that makes the reasoning survive

Select one source with its checkbox, then ask for structure rather than for a summary. This one is written to fail loudly rather than fill gaps:

Summarise this source as an argument, not as a list of findings. Give me exactly these fields:
CLAIM: the main claim, in the authors' own words where possible.
AGAINST: what position or prior result the authors are arguing against. Quote the sentence where they set it up.
EVIDENCE: the specific evidence offered, with sample size, study design, and any effect size or interval as reported.
SCOPE: the conditions the authors say the claim holds under, quoted.
HEDGE: the strongest limitation the authors state about their own work, quoted.
FALSIFIER: what result, according to this paper, would count against the claim.
Write "not stated in the source" for any field the source does not contain. Do not infer, do not generalise, and do not use knowledge from outside this source.

Three design choices do the work. Asking for quotes forces the tool back to the text instead of paraphrasing toward the middle. The AGAINST field recovers the counterposition, which is the single most-dropped element. And the explicit "not stated in the source" escape hatch gives the model somewhere to go that is not invention.

Google's own guidance points the same direction: "To get more focused summaries in the chat, ask specific questions about the information you're looking for. When multiple sources are selected, mentioning source names in your query helps Gemini Notebook narrow its search."

Set length and style before you ask, in Configure Chat

Two settings sit under Configure Chat in the Chat panel, and both change how much argument survives. The chat documentation lists them:

• Conversational style. Default is "Best for general purpose research and brainstorming tasks." Learning Guide is "Best for educational content, helping you grasp new concepts and skills efficiently." Custom is "Best for choosing a specific style ('Respond like a PhD student') or suggesting a role."
• Response length. Default, Longer, or Shorter. Then select Save.

Longer is the correct setting for argument-preserving work, for the dull reason that hedges and scope conditions are the first casualties of a length budget. Learning Guide is the wrong setting for it, since simplification is exactly the operation you are trying to prevent.

A Custom style is worth writing once and leaving in place. Something like: "Respond as a methods reviewer. Always report sample size, study design and stated limitations. Never smooth over disagreement between sources." Note that Google documents no effect of the style setting on grounding, only on tone and register, so this is a nudge and not a control.

Audio Overview has two formats built for disagreement

Most people generate the default Audio Overview and never see the format picker. There are four, documented in Generate Audio Overview in Gemini Notebook, and two of them are aimed at argument rather than at recap:

• Deep Dive, the default, where "Two hosts unpack and connect topics from your sources in a lively conversation."
• The Brief, where "A single speaker delivers the key takeaways of a document in under two minutes."
• The Critique, for critical feedback, where "Two hosts provide a constructive evaluation of material like an essay or design doc."
• The Debate, for exploring multiple perspectives, where "Two hosts engage in a formal, back-and-forth debate on the topic."

The Debate over a set of papers that disagree will surface tensions a written summary flattens. You can also "Input a prompt to provide instructions to focus on specific topics or adjust the expertise level", and set length to Shorter, Default or Longer, which the docs note is English only.

Then there is the caveat Google prints on the same page, and it is the reason audio is a thinking aid rather than a citable artifact: "Audio Overviews, including voices, are AI-generated and may contain inaccuracies or audio glitches." Google also frames the format as "designed to be an objective reflection of your source content, rather than subjective opinions from the AI hosts", which is a design intention rather than a guarantee.

Have the notebook critique your own summary

The notes documentation lists quick actions you can run over selected notes. One of them is genuinely useful here: you can "Ask Gemini Notebook to give you constructive feedback on your prose or argument." Others gather all notes into one, produce an outline "organized around topics", or build a study guide.

So the loop is: write your summary as a note, select it, ask for feedback on the argument, and see whether the tool can find the paper's hedge in a place your summary left blank. That is a cheap check and it costs no chat query.

Keep the result in the notebook. If you export it, know what you are giving up: "Any modifications made within the newly created Docs or Sheets file will not synchronize with your original Gemini Notebook note, and sharing permissions will not carry over."

The Reports gap in Google's own help centre

Reports are metered on the Upgrade page at 10/day on the free tier, 20/day on Plus, 100/day on Pro and 1K/day on Ultra. Flashcards, Quizzes and Mind Maps get the same allowance and each has its own help article.

Reports do not. The Gemini Notebook help centre index runs to sixteen articles covering chat, sources, notes, Mind Maps, Audio Overviews, Video Overviews, Flashcards and Quizzes, Infographics and Slide Decks, and there is no article documenting what a Report is, which report types exist, how long one runs, or whether it carries citations back to your sources.

Which means the artifact most likely to look like a finished literature summary is the one Google explains least. Generate them if they help you think. Do not paste one into a draft, because you have no documented way to check where any sentence in it came from.