Use an Audio Overview without losing the qualification
A lively two-host discussion can make a cautious finding sound settled. Before you share the overview, listen for the caveat as well as the conclusion.
What you'll make
a focused Audio Overview whose key claim you’ve checked against the source, with a short note of the timestamp and passage.
What you'll need
Gemini Notebook (formerly NotebookLM) in your work account, edit access to a notebook with a few sources you’re allowed to use, and about twenty minutes. The beginner’s guide to NotebookLM and Gemini Notebook covers setting one up.
Example result.
Claim checked: 2:14 — "Attendees reported higher confidence than non-attendees."
Source: induction questionnaire, para 1 (18 attendees vs 12 non-attendees; attendance not random).
Qualification: kept — "That's an association; the survey can't show the induction caused it."
OK to share with: the onboarding working group, via the shared notebook.1. Decide what the listener needs and what must survive
Choose the audience and one question the overview should answer. Then write down the qualifications that must survive: an unconfirmed date, a small sample, an exception, or an association that doesn’t show cause. Read the key passage yourself. Keep the overview private while you check it.
Why this works: you can only listen for a caveat you’ve named in advance.
What you should see: one question and one or two qualifications written down.
2. Generate a focused overview
Open the notebook. In the Studio panel, select Audio Overview and use its options before generating:
- Format: Deep Dive (default, two hosts), The Brief (one speaker, under two minutes), The Critique or The Debate. For a first check, The Brief is quickest to verify.
- Length: Shorter, Default or Longer (English only).
- Prompt: paste the instructions below.
Explain [topic] to [audience] using only the selected sources. Focus on [question]. Keep these limits clear: [qualifications]. Distinguish what the source found, what it didn't establish and what is still unknown. Don't turn an association into a cause or a proposal into a decision.The brackets are the topic, the audience, and your question and qualifications from step 1. If you don’t see the Audio Overview options, the feature may be limited on your account or the mobile app; make a written briefing from the notebook instead, and don’t claim you reviewed audio you couldn’t generate.
Why this works: naming the limits points the hosts at exactly the detail broad summaries drop. It doesn’t guarantee accurate speech, so listening is still the job.
What you should see: an overview generating in the background in the Studio panel.
3. Listen and check the claims that matter
Play it. Pause at any statement that would change a decision and note the timestamp, the words and the source passage. Check numbers, comparisons, conditions and recommendations. You can open sources and ask questions while it plays. Google notes that generated audio can contain inaccuracies or glitches, so a sudden voice switch isn’t a source statement.
Why this works: the hosts’ confidence is a performance; the timestamp and passage are what you can check.
What you should see: a note like the example result, for each claim that matters.
Check the result
- Every claim you’d act on has a timestamp and a matching passage.
- Each qualification you named in step 1 is said out loud.
- Numbers and group sizes match the source.
- Nothing turns a finding into a cause, or a proposal into a decision.
The usual failure: the caveat drops out of the conversation. “Attendees reported higher confidence” becomes “the induction boosted confidence”, because it makes better listening. Spot it at the timestamp where the finding is discussed. Then generate a new overview with this prompt, and delete the old one once the new one checks out (both need edit access):
The previous overview overstated [claim] at [timestamp]. The source says [passage], including [qualification]. Focus this overview on that distinction. Keep the supported finding, state its limit plainly and don't repeat the overstatement.Fill in the claim, the timestamp, the source wording and the qualification. If it still won’t hold the caveat, share a checked written briefing instead.
Worked example: association isn’t explanation
Fictional example. The source:
In a voluntary questionnaire, 18 respondents who attended the induction reported
higher confidence than 12 respondents who did not. Attendance was not randomly
assigned. The questionnaire does not establish that the induction caused the
difference, and respondents may differ in other ways.A Deep Dive overview said at 2:14: “So the induction really works: it increased confidence.”
- The failure. “Increased” is a causal claim the source explicitly rules out, and the 18-versus-12 groups and non-random attendance had gone.
- The repair. A new overview in The Brief format, with the correction prompt, said: “Attendees reported higher confidence. That’s an association; the survey can’t show the induction caused it, and the two groups may differ in other ways.” The old overview was deleted.
- Why it matters. The small groups and self-selection aren’t background detail; they’re why the finding is cautious. The example result records the check.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
When you can’t listen at work
If you can’t play audio, ask the notebook for a written briefing with the same focus and limits, then check it against the passage the same way. It’s the same job without the timestamps.
Share the checked version only
Share from the audio player with Share, which needs the notebook to be shared with the recipient; or share the whole notebook so others find it in Studio; or Download it. Public links aren’t available for Workspace Enterprise or Education accounts, and deleting an overview breaks its share link. Share a notebook with the right sources and audience covers who sees what. For checking written answers, see Ask a notebook question and check the cited passage.