AI at Work
Gemini Deep Research for Work: Audit the Research Plan, Sources, and Final Claims
A revised fictional supplier-research plan and ten-claim audit turn source qualifications, unsupported forecasts and missing evidence into a bounded review memo.
Audit a Gemini Deep Research report at three points: the plan before research starts, the sources behind its consequential statements, and the claims retained in the final memo. Specify the business decision and permitted materials first. Treat the cited report as a draft. Open the sources and check whether they support each consequential statement’s wording, date, scope, and comparison.
This LandOffer.ai guide uses an entirely fictional supplier investigation to show a plan revision and a ten-claim audit. The vendors, source extracts, generated-looking statements, people, and decisions are invented teaching materials. They were not produced by a Gemini session or obtained from a supplier. Current Google documentation supports the feature descriptions; no authenticated research, internal connection, upload, purchase, or sharing was performed.

Establish what Deep Research can use in your account
Google's Deep Research instructions, checked October 25, 2026, describe a research plan that you can edit before selecting Start research. Google Search is included by default; source selection can include uploaded files, notebooks, and supported connected services. Availability in your account is a separate question from whether your organization permits particular information to be used.
For work or school accounts, Google's Connected Apps guidance identifies Workspace edition, administrator enablement, and activity settings as prerequisites for connections. It also warns that a response can use outdated information. Check the active work account and approved source scope before beginning. A personal account with an available connector does not establish permission to process company material.
For a public supplier scan, private mail may add no necessary evidence. Start with the minimum materials needed to answer the question. If your task requires an internal requirements document, establish that the document and the selected service are approved for that use; omit unrelated customer details and confidential negotiations. Do not connect another service merely because a report looks incomplete.
This article does not compare plans, promise a particular report allowance, or confirm that every source option appears for every reader. Those facts can vary. Record the account context, permitted sources, decision question, and exclusions before starting this investigation.
Turn a broad request into a researchable decision
Fictional operations lead Leena needs to decide whether two invented knowledge-search suppliers, Alder and Cove, merit a limited evaluation. She is preparing questions for an internal review, not selecting a supplier or approving a contract. Her requirements are employee sign-in, a documented export route, stated storage-location options, and a way to investigate how access restrictions affect results.
Her initial request is: “Research Alder and Cove and recommend the best enterprise search tool.” That invites a broad ranking without specifying the organization's needs or what evidence would justify a recommendation. Without those limits, a favorable statement about an Enterprise edition could be mistaken for evidence about the edition your team would evaluate.
Leena rewrites the task around a smaller decision: “Identify which requirements have current public evidence and which require written answers or a controlled evaluation. Do not rank overall quality or recommend a purchase.” She records an October 20 evidence cutoff for the fictional packet. An older page may still be relevant, but the memo must preserve its date and avoid describing it as a current commitment without confirmation.
The cutoff is part of the example, not today's date or a Gemini setting. A real investigation should record the actual retrieval date separately from each document's publication, update, and effective dates. A page updated recently can still quote an older policy or describe a future feature.
Worked asset: the research plan before and after review
The following plans are author-written simulations. They illustrate what Leena would change; they are not screenshots or evidence of a real model's behavior.
The initial plan proposes: survey general reviews, compare feature lists, estimate savings, inspect any available internal messages, and recommend a winner. Leena rejects it because neither the evidence boundary nor the decision standard is defined.
| Plan area | Initial proposal | Leena's completed revision and reason |
|---|---|---|
| Decision | Choose the best supplier | Prepare an evidence memo for deciding whether a limited evaluation is worthwhile; purchase remains outside scope |
| Materials | Search reviews and relevant company messages | Use the two suppliers' public product, plan, help, and policy pages plus a synthetic requirements sheet; no Gmail, Drive, or confidential material |
| Comparison | Combine all listed features | Keep supplier, edition, region, and date attached to each feature; an Enterprise statement does not establish Standard availability |
| Evidence | Include supporting links | Preserve the exact supporting passage and qualification for each consequential claim; label marketing, documentation, and inference separately |
| Economics | Estimate productivity savings | Record published pricing conditions if available; omit savings and total-cost forecasts without workload, price, and effort evidence |
| Output | Ranked recommendation | Ten-claim audit, unresolved questions, and a conditional next-step memo; no contact, sharing, trial enrollment, or commitment |
The revised prompt requests conflicting evidence as well as favorable passages. It asks for “not found” when the selected materials do not establish a requirement and for separate treatment of announced features. The revised plan records the requested evidence standard. The report still needs inspection to determine whether it followed that standard.
In the documented interface, review the proposed plan with Edit plan before starting. Confirm the actual source selection as well as the wording: a sentence saying “public sources only” is not a substitute for inspecting which sources are selected. Google documents that Search can be deselected when research should use only other selected sources. Keep the configured source scope with your review record.
Read the evidence behind each important sentence
Check the original page rather than accepting a report's description of it. Read enough surrounding text to identify exclusions, plan names, conditions, effective dates, and whether a statement is an aspiration or an available capability. A citation attached to a paragraph may support only one sentence.
Use distinct dispositions. Retain a supported statement at its actual scope. Narrow wording that overgeneralizes. Mark a contradiction when the cited source says the opposite. Keep an unanswered question open. Reserve “not found” for the materials reviewed; it does not prove the feature does not exist.
Google's response guidance warns that Gemini can present inaccurate information, including claims about its own sourcing and freshness. Asking the assistant whether it verified everything is therefore insufficient evidence. Open the relevant reference and record what you checked.
For inaccessible or missing documents, preserve the gap. Do not replace an unavailable policy with a review article and retain the same certainty. A marketing page can establish what a vendor claims, while a documented procedure can establish a narrower supported workflow. Neither by itself proves behavior under your configuration or supplies contractual terms.

Worked asset: a completed ten-claim audit
All entries below belong to the invented packet. S1 is Alder's October 12 plan page; S2 its October 15 export guide; S3 its October 16 storage FAQ; S4 Cove's October 10 product page; S5 its October 14 administration guide; S6 its September 30 preview note; S7 an Alder marketing example updated October 8. No real company makes the fictional statements quoted here.
| Draft claim | Source evidence in the fictional packet | Audit decision and final wording |
|---|---|---|
| 1. Alder includes SAML on every plan | S1 lists SAML under Enterprise, not Standard | Narrow: “S1 documents SAML for Enterprise.” Standard entitlement remains unestablished |
| 2. Alder exports every attachment | S2 describes CSV export of record fields and explicitly excludes attached files | Contradicted: “The documented export excludes attachments.” Test completeness separately if an evaluation proceeds |
| 3. Alder stores all data in the EU | S3 offers an EU primary-storage option for specified Enterprise tenants; it does not describe every processing location | Narrow to the stated option and tenant scope; all processing locations remain an unanswered question |
| 4. Cove is always faster | S4 claims rapid search but supplies no comparable workload or latency result | Remove the ranking; this packet contains no comparative speed evidence |
| 5. Cove searches every file type | S5 lists supported text, PDF, and presentation inputs with size limits | Replace with that documented list; other formats are unestablished, not necessarily unsupported |
| 6. Cove's new connector is generally available | S6 labels it a limited preview requiring invitation | Correct: “The September 30 note describes an invitation-only preview.” Current access requires confirmation |
| 7. Alder will save our team ten hours weekly | S7 describes a marketing example from another organization, with no matching workload | Remove the forecast; retain only the fact that a marketing example exists if it is relevant |
| 8. Neither supplier changes existing access rules | S5 describes Cove filtering results by indexed permissions; S1 contains no equivalent Alder explanation | Split: retain Cove's documented mechanism as a vendor description; Alder remains unanswered and neither implementation is tested |
| 9. A trial proves migration is reversible | S2 documents a partial export, not restoration or rollback | Remove the conclusion; identify restoration and missing attachments as evaluation questions |
| 10. Both satisfy our requirements | Claims 2, 3, 5, 6, 8, and 9 expose exclusions or gaps | Replace with “The packet is insufficient to establish full suitability for either supplier.” No purchase recommendation |
This audit resolves the wording of all ten draft statements, including statements that must be removed. It does not resolve every underlying supplier question. That distinction lets a reviewer finish the evidence task without falsely declaring the procurement decision ready.
Several rows also show why a single positive label is inadequate. Cove's permission-filtering description is relevant, but indexed permissions may not answer Leena's question about timing after a restriction changes. Alder's partial export is concrete evidence, but its usefulness depends on whether attachments are necessary to the intended workflow. Ask how quickly Cove applies a permission restriction and whether Alder can preserve the attachments your workflow requires; neither question is answered by the cited extract.
Produce a bounded memo and preserve its open questions
Leena's final fictional memo says: “Do not select either supplier from this packet. Alder documents Enterprise SAML and a limited export, but attachment preservation and processing-location questions remain. Cove documents particular input formats and permission filtering; its preview connector and access-change behavior need clarification. Request scoped answers before proposing a controlled evaluation.”
She assigns herself to prepare those questions for procurement owner Omar. Omar accepts a review of the question list, not permission to contact suppliers or enroll in trials. Technical lead Sofia accepts a thirty-minute review of the proposed access-change check, replacing an optional demonstration-polish slot. Neither review has occurred at the memo's cutoff.
The memo retains its revised plan, source identifiers, dates, ten dispositions, and decision limits. That allows another reviewer to reconstruct why a favorable sentence was narrowed or removed. If a later document changes an answer, update the affected claim and decision rather than silently substituting a new page behind the old citation.
Google documents report copying and export options, including Export to Docs. Treat export as a separate destination decision: the generated content and any permitted internal material still need an appropriate audience. For this exercise nothing was exported or shared. Check the retained claims before distributing an exported document.
For your next research task, choose the three statements most likely to change the decision and open their original sources first. For each, save the supporting passage, its date and edition, the wording you retain, and any unanswered question. Then audit the remaining material in the same record. Find more workplace guides through LandOffer.ai; finish with a reviewable evidence memo whose confidence matches its sources.
Sources and Further Reading
- Google Gemini Apps Help: Use Deep Research — plan review, selected sources, report retrieval and export; public instructions checked October 25, 2026.
- Google Gemini Apps Help: Connected Apps with a work or school account — account prerequisites, administrator controls and potentially outdated responses.
- Google Gemini Apps Help: Learn about responses — limitations of generated information and descriptions of sourcing.


