NotebookLM Tutorial: Compare Vendor Documents and Check Claims

What you will produce

You will compare a vendor's overview, technical notes and service schedule, then turn the differences into a reviewable evidence log. The goal is not an AI-generated winner: it is a record that another person can verify without trusting your chat history.

Naming note, checked 2 November 2026: Google's NotebookLM help URLs currently redirect to Gemini Notebook Help and the help pages link to notebook.google.com. This guide retains NotebookLM in its title for readers searching by the earlier name. Follow Google's official links if your account shows the newer name. Interface labels and account availability can vary.

Validation boundary: the fixture documents and expected manual findings below are authored for this exercise. They are not real vendor disclosures, and the example findings are not claimed to be actual NotebookLM output. No authenticated Google session was used to certify the model's responses.

Before you start

Use a desktop browser, an eligible Google account with access to the notebook product, and a blank local document or spreadsheet for your evidence log. The web workflow is the reference here; mobile features may differ. This exercise needs only three pasted-text sources and does not require a paid subscription. If your account is ineligible or has reached a limit, stop and use the same manual review method offline.

Use only the fictional text supplied here. Before uploading real proposals, check your organization's confidentiality, data-processing and access rules. Do not upload client records, wallet secrets, API keys or non-public contracts without approval. Sharing the notebook is a separate disclosure decision from sharing your final findings.

Step 1: Create a source manifest

Open the notebook product from Google's official help page. Create a new notebook named Vendor evidence lab - 2026-11-02. Keep it private. In your evidence log record the notebook title, review date, reviewer and the three source IDs below. These IDs make the exercise reproducible even if citation numbers change.

In a real review, also record the original URL or file name, publisher, document date, retrieval time and version. A retrieval date does not prove when a claim became effective. Save an authorized snapshot outside the notebook so later updates do not erase the evidence you reviewed.

Step 2: Add the fictional overview

In the Sources panel choose Add sources, then the pasted-text option. Paste the entire block below and give it the Source A title from the manifest. Open the imported source and confirm that both the first and last sentences are present. If ingestion is still processing, wait before asking questions.

This deliberately broad overview is the first half of the comparison. Do not treat words such as every, instant or continuous as measurable commitments without scope and terms.

FICTIONAL TRAINING DOCUMENT - NOT A REAL VENDOR
Harbor Sample - Overview
Document date: 2026-10-01
A1. Our platform supports USDC payments on every major EVM network.
A2. Payment notifications are instant.
A3. Customers receive continuous support.
A4. All customer data stays in the European Union.
A5. Contact sales for pricing.

Step 3: Add the technical notes and service schedule

Repeat Add sources twice, creating separate sources B and C from the two labelled documents below. Do not paste them into one source. Use the manifest titles and verify the line labels after import. Select only A, B and C in the Sources panel; exclude any discovered sources or unrelated notes.

If a web import is incomplete, open the original page and compare its contents. An imported HTML page does not automatically include linked subpages. For this lab, pasted text removes website-rendering and PDF-extraction ambiguity.

FICTIONAL TRAINING DOCUMENT B - NOT A REAL VENDOR
Harbor Sample - Technical Notes
Document date: 2026-10-15
B1. This pilot supports Ethereum Sepolia only; it does not support mainnet payments.
B2. Webhooks are sent after our confirmation policy is met; delivery time is variable.
B3. Integrators must handle duplicate webhook deliveries.
B4. Primary application storage is in the EU. Operational logs may be processed in the US.
B5. This document does not specify a maximum log-retention period.

FICTIONAL TRAINING DOCUMENT C - NOT A REAL VENDOR
Harbor Sample - Service Schedule
Document date: 2026-10-20
C1. Standard support operates Monday-Friday, 09:00-17:00 UTC, excluding published holidays.
C2. The target first response for priority-one tickets is four business hours.
C3. A first response is not a resolution commitment.
C4. This schedule contains no availability percentage or service-credit commitment.
C5. A separate signed order form determines fees; no fees are stated here.

Step 4: Ask for a bounded claim comparison

Paste this prompt into Chat after confirming all three sources are selected. Do not request web research or additional sources. If your current interface offers tools that can search beyond the notebook, keep this exercise confined to A, B and C and check the origin of every cited passage.

Read the response as a draft. A neat table is not evidence of accuracy. If it combines unrelated claims into one row, ask it to split them so each row can have its own verification status.

Use only sources A, B and C selected in this notebook. Do not browse or infer missing terms. Compare network support, notification timing, duplicate handling, support hours, response versus resolution, data location, retention, availability and pricing. For each row return: claim; source title and line label; short supporting excerpt with a source citation; limitation or conflicting passage; status (supported, qualified, conflicting or not stated); follow-up question. Keep marketing wording separate from a technical or contractual commitment. If the sources do not establish a fact, write not stated.

Step 5: Open every material citation

Select a citation in the answer to open the supporting passage; hovering may show a preview. Compare the sentence with the full source around it, including qualifiers, dates and exclusions. Then open the original source text, not just the model's paraphrase. Record the source title and line label in your log; a floating citation number alone will not survive every export or revision.

Check whether each passage actually supports the specific claim. B1 establishes the scope of this pilot, not the vendor's entire future offering. C2 concerns first response during business hours, not four elapsed hours or resolution. B4 contradicts an absolute all-data-in-EU assertion, but does not tell you which personal data appears in logs. Mark that follow-up unresolved.

If a citation opens the wrong passage, provides only a source-level reference, or cannot be found in the original, mark the claim unverified. Ask a narrower question naming the source and line, or read it manually. Do not invent a quote to complete a row.

Step 6: Compare against the manual answer key

These are expected human findings from the fictional text, not a required model response. Your wording can differ. Your review should preserve the distinctions in this table. In particular, missing retention and price information must remain missing rather than becoming zero, unlimited or included.

Step 7: Save a reviewer-owned evidence log

Create one row per claim in your own evidence log using the fields below. Write supported only when the cited source supports the precise scope you record. Use qualified for a narrower supported version, conflicting for inconsistent statements, and not stated when the source set is silent. A second reviewer should be able to find every passage from the source ID and line label.

You may save a useful chat response with Save to note, but keep your checked log and source manifest outside the chat as well. If you export, reopen the exported file and test whether citations still resolve. Preserve explicit source titles and excerpts even when interactive citations do not survive.

claim_id | precise_claim | source_id | document_date | location
short_excerpt | conflict_or_qualification | review_status
reviewer | checked_at | vendor_question | response_due | resolution

Example authored row, not model output:
EU-01 | All data stays in EU | A + B | 2026-10-01 / 2026-10-15
A4 / B4 | US log processing disclosed in B4 | conflicts with all-data wording
conflicting | your name | your review date | Which log data is processed in US?
not assigned | unresolved

Step 8: Rerun after changing the evidence

As a controlled test, deselect A and ask the same question using B and C only. The pilot network scope should still be available, but the broad marketing assertions should no longer be presented as sourced facts. If they persist, start a fresh chat or notebook with only B and C and check again; never use model recall as a substitute for the selected documents.

For real documents, record when a source changes and repeat affected checks. Current Google help describes automatic updates for connected Drive sources, while uploaded or pasted files require your own version handling. Existing generated outputs should not be assumed to update just because the underlying source changed. Keep the review snapshot separate from the live notebook.

Troubleshooting and stop conditions

Stop the review if you cannot access the original evidence, if a document's permission or confidentiality status is unclear, or if the output cannot be traced to selected sources. The right outcome can be an unresolved question, not a completed scorecard.

Acceptance checklist

You have three separately titled sources, a dated manifest, six or more reviewed claim rows, and a follow-up question for each gap or conflict. Another reviewer can reproduce the Sepolia scope, support-hours conflict and US-log exception directly from the fixture. No unsupported price, uptime or retention value has been filled in.

The evidence log supports further diligence, not legal, compliance or security approval. Real vendor claims require current documents, relevant contractual terms and appropriate specialist review.

Continue the research workflow

• Compare AI research tools
• Browse practical tutorials