Check a research report before relying on it
The research report recommends a supplier, cites eleven sources and reads beautifully. Before anyone buys on it, check whether the claims driving the recommendation rest on an old price, the wrong plan or a certificate that doesn't exist yet.

What you'll make
a corrected brief that keeps only the claims its sources support, narrows the ones they partly support, and lists the evidence still needed before the decision.
What you'll need
a normal chat in Claude or ChatGPT (Gemini or Copilot work too), the report with its cited sources or the fictional AlderDesk pack below, someone who can open those sources, and about 30 minutes. You don’t need to start a new deep-research task to check one.
Example result.
Recommendation: pause. Don't buy AlderDesk Pro yet.
Evidence still needed:
- A current quote: the £18 offer ended on 31 December 2024 and had a five-seat minimum.
- Which plan includes CSV audit-log export (the feature page says Enterprise only).
- A completed security certificate: the supplier says it is "preparing for" one.
1. Keep the report and its sources together
Save the original report, its date and its source links before you change anything. Write down the decision it’s meant to support and your must-have requirements. Gather the cited pages or extracts with their dates and page or section. Label anything you can’t open.
Then pick the few claims that could change the decision (prices, required features, dates, security) and split combined sentences. “Pro costs £18 per user monthly with no minimum” is two claims: a price and a minimum. One can be right while the other is wrong.
Why this works: a combined sentence lets a correct price lend credibility to a wrong “no minimum”.
What you should see: a short list of single claims, each with the source it cites.
2. Ask what each source actually supports
Paste or attach the report and sources. If you give a link, check the assistant really opened it: a URL is not the page.
Audit [report] for this decision: [decision and requirements]. Use only
the source material below. First say what you could read, including
inaccessible links. For each important claim give its exact wording,
the source and date, the short supporting passage and its location, a
status, and any correction. Status is one of: supported, partly
supported, contradicted, unverified. Split combined claims. Missing
evidence is unverified, not false. Check product, plan, period and
geography. Don't invent prices, certificates or results, or search
outside these sources.
REPORT: [paste or attach]
SOURCES: [extracts, files or links]
[report] names it; [decision and requirements] say which errors matter most.
Why this works: the exact wording stops the model softening a claim before judging it, and the passage makes each verdict checkable.
What you should see: a ledger: one row per claim, each with a passage and one of the four statuses.
3. Open the decision-critical sources yourself
For each claim that could change the decision, open the source and read the passage the ledger quotes. Is it the same product, plan, date and country? Does it support the claim, or only discuss the topic? For prices, check billing period, minimum seats, tax and whether the offer is still valid. For a study, check who ran it and what was measured.
Why this works: another model agreeing with the ledger isn’t the same as reading the page.
What you should see: each important claim confirmed, narrowed or downgraded by you, not just by the ledger.
4. Keep only the conclusion the evidence can carry
Rewrite the report using only claims marked supported in the reviewed
ledger. Narrow partly supported claims to what the source shows. Put
unverified claims that matter for the decision in an "Evidence still
needed" section. Keep facts apart from recommendations, link each fact
to its source and location, and add a short change log.
Keep the original. Check the source links still work in the saved version. Where the decision has legal, financial or safety consequences, involve a qualified reviewer.
Why this works: the recommendation changes only where the gaps affect the decision, which is the honest result.
What you should see: a shorter brief, an “Evidence still needed” section and a change log.
Check the result
- Each claim you keep matches its passage for the same product, plan and date.
- Expired offers, “coming soon” features and “preparing for” certifications aren’t presented as current facts.
- A study described as independent really was run by someone other than the supplier.
- Unverified decision-critical claims appear in “Evidence still needed”, not deleted.
The usual failure: the source is real and on topic, but supports less than the report says: a different plan, an expired price, a certification in preparation. Spot it by reading the scope words around each quoted passage (“Enterprise only”, “offer withdrawn”, “preparing for”).
Claim [claim] cites [source], which says [what it says]. Rewrite the
claim to match that scope exactly, or mark it unverified. Then say
whether the recommendation still holds without it.
[claim] is the report’s wording; [source] and [what it says] come from the passage you read.
Worked example: the source supports less than the report says
Fictional example. The report recommends AlderDesk Pro for 12 staff. The full report and source cards and answer key are optional practice files; don’t search for the invented supplier.
Claim: Pro costs £18 per user monthly with no minimum.
Archived price page, 1 June 2024: Pro £18 per user per month; minimum five
seats. Offer withdrawn on 31 December 2024.
Claim: Pro includes CSV audit-log export.
Feature page, 2 October 2026: CSV audit-log export is available only on
Enterprise. Pro offers a usage summary.
Claim: AlderDesk is independently security-certified.
Security page, 30 September 2026: Preparing for an external certification
assessment. No completed certificate supplied.
Claim: An independent study proves every team saves 28% of work time.
Supplier blog, 12 September 2026: Our internal trial with 12 volunteers found
28% less drafting time on one standard response. Review time not measured.
- Price and minimum: unverified. The offer has ended, and it had a five-seat minimum, so “no minimum” was wrong even then. Today’s price is unknown.
- Audit export: contradicted. The current feature page puts it on Enterprise. Certification: unverified. “Preparing for” isn’t certified, and the gap doesn’t prove the product is insecure.
- The study: the failure in its clearest form. “Independent” is contradicted (the supplier ran it), and “every team saves 28% of work time” isn’t established. The repair prompt narrowed it to: “The supplier reports 28% less drafting time on one standard response in an internal trial with 12 volunteers. Review time wasn’t measured.” The rewrite prompt then produced the example result.
Count the minutes the audit took and the corrections it found: that’s the real cost of the polished report.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Where to start a new deep-research task
If you need fresh research rather than a check, review the plan and sources before it runs, then audit the report the same way.
| Product | Where to start |
|---|---|
| ChatGPT | + › Deep research in the message box |
| Claude | + › Research (paid plans; web search must be on) |
| Gemini app | Add Files › Deep Research, then Sources and Edit plan before Start research |
| Gemini in Workspace | Ask Gemini › Tools › Deep Research, then Customize before Start research; the report saves as a Google Doc |
Give a new research task a boundary
A bounded brief produces a report that’s easier to check.
Research whether [supplier and product] meets these requirements for a
small pilot: [requirements]. As of [date], use public primary sources.
Separate supplier claims, independently established facts and your
inferences. For price, give currency, billing period, minimum seats,
tax and required plan. Give a URL, section and access date for each
important claim, and mark missing information.
Audit a long report
Check every decision-critical claim, plus a stated sample of the rest, and say what the sample covered: a checked sample doesn’t certify the whole report. Trace “research shows” links back to the original study, including who paid for it and what was compared. Two articles repeating one press release are one evidence chain. The claim ledger template keeps it organised. For gathering evidence on several competitors from scratch, see competitor research you can check.