A business owner needs to choose a file-sharing workflow. One option looks easier for staff; another promises more control. A polished AI comparison arrives with a confident recommendation and a long source list. The unresolved question is whether those sources establish the requirements that actually matter.
ChatGPT deep research can help assemble that brief. The operator’s job is to define the decision, inspect the evidence and keep missing information visible. This guide provides a repeatable way to do that, including a fictional office example, a reusable prompt and a claim-by-claim review checklist.
What deep research currently supports
Deep research combines information from the web, uploaded files and eligible connected apps into a cited report. You can review its proposed plan and steer the research while it runs. Availability and usage depend on your account; check the in-product counter. Connected sources retain their permissions, and research uses app read actions, not write actions. Not every app supports it. These details follow OpenAI’s current help guide, checked October 3, 2026.
This article’s briefing and review method is an EAC recommendation. The example below is synthetic: it is not a client engagement, a vendor benchmark or a completed research run. It demonstrates what to request and how to judge the resulting work.
Start with a decision you can describe
“Find the best software” leaves too much room for the assistant to invent your priorities. A more useful question names the workflow, the people affected and the next decision. For example: “Which two file-sharing options should our office test against these requirements before requesting a quote?” That asks for a shortlist and a test plan, rather than treating a report as permission to buy.
Write down your current process first. Where do files start? Who edits them? Who receives them outside the business? What causes rework? Include constraints that would rule an option out. If a requirement is undecided, say so. An unknown retention requirement is information the owner must resolve; the assistant should not quietly fill it with an assumed number.
Separate requirements from preferences. Staff access and external-sharing controls may be required. A familiar interface may be preferred. A tool that meets preferences while missing a requirement should remain a conditional candidate. Ask the report to show that distinction explicitly instead of hiding it inside an overall score.
Before adding internal material, check that it is appropriate for the account and the task. A sanitized requirements sheet may be sufficient. Do not upload customer files merely to illustrate a workflow when a fictional sample would answer the same question.
Choose sources before asking for conclusions
In the source controls, restricting research to specified sites limits the allowed websites. Prioritizing sites while allowing full-web search leaves broader discovery available. Select the setting that matches your brief, rather than assuming a list of preferred domains creates a restriction. The current help guide describes both choices.
For a first vendor shortlist, start with official documentation for the exact product and plan. Product pages can establish what a vendor advertises. Help pages may explain how a setting works. Contract documents and a current quote may be needed for commitments and purchasing terms. These are different forms of evidence; ask the report to identify which it used.
A source restriction can also create a blind spot. Official vendor pages are useful for documented features, but they do not by themselves establish staff usability or independent performance. Keep those questions open for a hands-on test or a separate research pass. If you broaden discovery, label independent reports and user anecdotes separately from vendor documentation.
ELEVATED AI CONSULTING · RESEARCH WORKFLOW
Evidence earns a place in the decision.
Name the next decision
Set evidence boundaries
Review the research scope
Check decisive claims
Resolve unknowns first
Worked example: a fictional office brief
Assume a fictional office has 12 staff members. Staff share draft documents internally and send approved files to outside partners. The owner wants to shortlist two existing candidate products for a trial. There is no approved purchasing decision. The following inputs are invented solely to make the exercise reproducible.
| Requirement | Input for this exercise | What still needs confirmation |
|---|---|---|
| Staff access | 12 individual staff accounts | The exact licensing terms for the chosen plan |
| External sharing | Staff need to send a file to a partner | Required recipient sign-in and owner-approved controls |
| Editing | Staff collaborate on drafts | File types and the office’s existing software |
| Retention | No policy supplied | The owner’s requirement and qualified review if needed |
| Budget | No approved amount supplied | Current quote, taxes and implementation costs |
Replace these inputs with your own approved requirements. Choose the two candidate products and their official domains before running the prompt. If you do not yet have candidates, make discovery a separate first task so the comparison does not mix a broad search with a final recommendation.
Copy this prompt, fill the bracketed fields, and set the source controls to match it:
Prepare a decision brief for a fictional 12-person office comparing
[Candidate A, exact product] and [Candidate B, exact product].
Our next decision is which options deserve a supervised trial.
We are not authorizing a purchase or any account changes.
Use the attached sanitized requirements sheet. Treat missing inputs
as unknown. Do not invent budget, retention or security requirements.
Research current official documentation on these approved domains:
[domain A], [domain B]. Identify the exact plan where possible.
Before researching, show the proposed plan and ask about inputs
that would materially change the shortlist.
Return:
1. A short summary of what is established and what remains open.
2. One row per requirement: candidate, claim, exact source URL,
relevant page section, plan scope, source date if available,
checked date, and status: supported, contradicted or unknown.
3. Separate labels for documented facts, estimates and your inferences.
4. Conflicting or inaccessible sources; do not resolve them by guessing.
5. A conditional shortlist and a small trial checklist.
6. Questions for the owner or vendor before a purchasing decision.
Do not treat a feature description as proof of compliance or a guarantee.
Do not send messages, edit files in connected apps or make a purchase.
Review the proposed plan before it starts. It should cover both candidates, all requirements and a way to expose missing evidence. If the plan is mostly a tour of product marketing, redirect it toward the individual requirements. If it introduces a third candidate or a new use case, decide whether that belongs in this brief before allowing the scope to expand.
The intended deliverable is a comparison you can inspect. It is acceptable for the report to conclude that neither option is ready for selection because the required inputs are incomplete. That is a useful finding when it identifies exactly what to resolve next.
Audit the claims that drive the decision
Read the recommendation first and identify the claims that could change it. Then open the cited pages for those claims. A source list is a route back to evidence; it does not establish that every sentence in the report is correct. OpenAI’s deep research introduction discusses incorrect facts and inferences as limitations. The review below is designed to catch those failures where they affect your decision.
Check the exact product, plan and feature scope. “External sharing supported” is less useful than an explanation of the recipient’s access, the owner’s controls and the plan to which the documentation applies. Check page dates where available, and record your own checked date separately. An undated page should remain undated rather than acquiring a date from the report’s creation time.
Here is a synthetic audit example. It deliberately uses placeholders instead of pretending to report live vendor findings:
| Draft claim | Evidence in the exercise | Reviewer’s decision |
|---|---|---|
| Candidate A meets external-sharing needs | A hypothetical help page describes a sharing feature; recipient requirements are missing | Narrow to “sharing feature documented”; requirement remains unknown |
| Candidate B is cheaper for this office | No current quote or approved budget was supplied | Remove the price conclusion; request comparable quotes |
| Candidate A is compliant with our retention needs | The office has no stated retention requirement | Remove the compliance conclusion; resolve the requirement first |
| Staff will prefer Candidate B | No staff trial has occurred | Label as an untested hypothesis; include a usability exercise |
Notice how each correction changes the claim, rather than merely adding “verify this” at the end. A useful review leaves the report more precise. If you cannot trace a decisive sentence to evidence, remove it from the recommendation until the source or requirement is available.
Turn the brief into a small supervised trial
A research report can suggest what to test. For this fictional office, a trial could use sample documents to check a few ordinary tasks: invite an approved staff tester, collaborate on a draft, share a sample with an authorized external tester, change access and confirm what the recipient can still open. The owner should choose the permitted environment and testers before any account or sharing action.
Record observations against the original requirements. For each task, write the expected behavior, the observed behavior and any unresolved question. A successful sample exercise establishes what happened in that configuration. It does not prove every plan, file type or future workflow will behave the same way.
Keep the research brief separate from the trial log. The former records what sources say; the latter records what your testers observed. If they disagree, preserve both and investigate the difference. That may reveal a plan restriction, a configuration issue, stale documentation or a misunderstanding in the brief.
For work that continues after the initial research, our OpenAI Dots guide covers assigning an ongoing responsibility. If an eventual workflow sends customer messages, use the separate customer-email test plan before enabling those actions.
A checklist before you share the brief
Use this review before presenting the report as decision support:
- ☐ The brief names the next decision, affected people and required inputs.
- ☐ Requirements and preferences are separate; missing inputs remain unknown.
- ☐ Source settings match the intended scope.
- ☐ Decisive claims have a source URL, relevant section and exact product or plan.
- ☐ Source dates and checked dates are recorded separately.
- ☐ Facts, estimates, inferences and trial observations have distinct labels.
- ☐ Conflicting or inaccessible sources remain visible.
- ☐ Prices and purchasing terms have current, comparable evidence if used.
- ☐ Compliance, suitability and staff preference are not inferred from feature lists.
- ☐ A named person owns the remaining questions and the final decision.
For a single factual lookup, a quick search may be enough. Use a deeper brief when the decision requires several sources and tradeoffs. Stop the research when the next step is owner clarification, a vendor answer or a controlled trial; another page of prose cannot supply missing requirements.
A good first run is one bounded decision, two candidate options and a brief you can check. Save the requirements, sources and corrections together so the next reviewer can understand why the shortlist changed. If you want help shaping that process, tell EAC about your project.

Sam Irizarry
Sam Irizarry is the founder of Elevated AI Consulting. His experience includes enterprise AI operations, personally delivered AI training, website development and workflow implementation.
Learn more about us →

