Turn business sources into a research brief you can inspect.

Start with a decision, not a folder of links

“Research client portals” invites a broad summary. “Which portals fit a five-person client team, what billing commitment do they require, and which facts are still missing?” creates a job someone can review. Name the audience, criteria and date before collecting sources.

Keep fact collection separate from judgment. A source may state an annual price, while another states a monthly price with different seat limits. The checker deliberately treats these as text. It does not turn them into a cheapest-vendor ranking or infer which features matter to your business.

Give each source an identity and a useful excerpt

The source library uses source_id, title, source_date, state and source_text. Dates use YYYY-MM-DD. State is current, superseded or draft. Include enough original wording to check the claim in context; an isolated headline may omit a condition that changes its meaning.

The review date and freshness window are yours to choose. Older does not automatically mean false, and a recent date does not make a claim reliable. These checks create a review queue. Duplicate source IDs stay ambiguous rather than silently selecting the first or latest document.

• Keep archived documents visibly superseded instead of deleting the history.
• Preserve an untouched copy of every approved source outside the working packet.
• Use fictional data to learn the workflow; obtain permission before importing company or customer documents into an AI product.
• Each CSV supports up to 10,000 data rows and 1,000,000 UTF-8 bytes. Source excerpts above 50,000 characters are retained for review but not searched.

Record observations with quotes—and leave room for unknowns

The findings CSV uses finding_id, option, criterion, value, source_id and quote. Give each observation a unique finding ID. Use the same option and criterion labels when two sources address the same question. A blank value means unknown; it is not converted into zero, no, or a negative feature claim.

Copy the source wording into quote and put your interpretation in value. Exact matching catches missing or altered wording, but it cannot tell whether your interpretation follows from that quote. Two different values with current traceable quotes under the same option and criterion are flagged as a possible conflict. Different plans or scopes can also produce that flag, so read both passages before deciding.

The worked answer: six findings need attention

Northstar Studio and all three portal vendors in this case are fictional. The packet contains five original source documents and twelve findings. F03, F05, F06, F07, F08 and F10 pass the mechanical checks. They still require a reader to verify the meaning and relevance of their quotes.

F01 and F02 record two current Alder Portal prices: USD 79 and USD 69 per month for the same five-seat plan. Both quotes exist, so the problem is a possible conflict—not a missing citation. F04 leaves Alder SSO availability unknown. The checker must preserve that gap.

F09 cites Cedar’s superseded USD 85 price; its current sheet says USD 95. F11 claims 24/7 support using a quote that is absent from the source, which actually describes weekday support. F12 cites a source that is not in the packet. Those issues do not disappear because the resulting memo sounds confident.

• Ask Alder to confirm which current price applies; do not silently prefer the newer FAQ.
• Keep Birch’s USD 720 annual billing commitment visible when comparing it with monthly plans.
• Replace the unsupported support statement with an accurately quoted, reviewer-checked statement, and request evidence for audit logs.
• Do not approve a vendor solely from this practice packet. It leaves business requirements and factual questions unresolved.

Give AI the evidence and the unresolved questions

After inspecting the queue, use the source packet, evidence matrix and audit with an approved AI tool. Ask for a short memo separating source-backed observations, unresolved questions and proposed next actions. Require source IDs and finding IDs in the draft, and tell the assistant not to choose between conflicting facts without reviewer guidance.

The downloaded sample memo is an authored answer, not a NotebookLM run or a model benchmark. You can try the same task yourself and compare its output with the answer key. A useful test includes the bad evidence as well as the clean findings: check whether the assistant preserves the uncertainty rather than filling it in.

Reuse the workspace without losing the source boundary

Google announced native notebook copying on August 17, 2026, including sources and Studio materials when copying is permitted. Copies do not stay synchronized with the original, and some source permissions still apply. This resource provides files and instructions; it does not provide a native copied notebook or access your account. Google’s copying announcement.

Current Google Help documents reports exported to Docs and data tables exported to Sheets, with citations in the exported table workflow. Review both the output and its sharing permissions. Recheck source dates and unresolved findings when you reuse a packet for another client or reporting period. Google’s notebook guide.