Deep research is worth the time when a client question needs many sources compared against each other, not when you need one fact. Write the question brief first, decide which sites it may use, then treat the finished report as a draft with footnotes: every claim you keep goes in a source log you check yourself. The citations tell you where to look. They do not tell you the claim is right.
Editorial research checked September 16, 2026 against OpenAI Help Center articles, read in a browser because they block automated retrieval. Prices are the USD figures OpenAI publishes. This article contains no affiliate links.
Jobs this actually suits
OpenAI positions the feature for investigating “complex questions, compare evidence, and produce results with citations or source links”, and is explicit about the alternative: “Use search for quick facts, and use deep research for depth and thoroughness.” For a solo consultant that maps to a short list of jobs.
- Market or competitor scans where you need twenty sources compared, not one summary.
- Policy or regulation background where you must show where each statement came from.
- Vendor or technology landscapes where the client will ask why an option was excluded.
- Pre-proposal context on an industry you have not worked in for a year.
It suits none of these: a single price check, anything where the answer must be current to the day, and anything you cannot verify afterwards because the underlying sources are closed to you.
Before you buyChatGPT Plus vs Claude Pro vs Google AI ProWhich AI subscription to pay for: plans, data settings and a five-step choice.See the plans →Step 1: write the question brief before you start
The tool asks clarifying questions and, in Chat, “will create a proposed research plan, which you can review and modify”. That plan is only as good as what you gave it. Write four lines first and paste them as the prompt.
| Field | What to write |
|---|---|
| Decision | The client decision the research supports, in one sentence. |
| Question | The specific question, with the market, region and time window named. |
| Out of scope | What you do not want researched, so the report does not wander. |
| Evidence standard | What counts as a source here: regulators, vendor documentation, filings, named studies. |
| Output | Structure, length, and that every claim must carry a source link. |
Ask for disagreement explicitly. A report that lists where sources conflict is more useful to a client than one that smooths the conflict away.
Step 2: decide the source scope
By default, deep research in Chat can access “The public web” and “Files you upload”. You can narrow that. OpenAI documents going to “Sites > Manage sites in the ChatGPT prompt window”, where you can restrict research to only the sites or domains you enter, or choose “Prioritize these sites, but allow full-web search” to weight them while still searching more widely.
For client work, a restricted list is usually the right starting point: regulators, standards bodies, the vendors in scope and the client’s own public material. Then run a second, wider pass if the narrow one leaves gaps. Connected apps are a separate question. OpenAI states that availability depends on plan, region, workspace settings and the app itself, that “Starting deep research does not grant access to additional files or apps”, and that deep research “uses available read actions from connected apps. It does not use app write actions as part of research.” Source: Deep research in ChatGPT.
Step 3: run it, and steer it
You can start a task by typing the deep research command in the chat window or selecting it from the tools menu. While it runs, OpenAI says you can follow progress and “interrupt at any time to refine the focus, including adjusting which sources it can access.”
Interrupt when you see it drifting into the out-of-scope list, leaning on one publisher, or answering a broader question than you asked. Steering mid-run costs less than discarding the report afterwards. The finished report opens in a fullscreen view with a table of contents, a sources used section and an activity history showing how the research progressed, and can be downloaded as Markdown, Word or PDF.
Step 4: build the claim-to-source log
This is the step that turns a generated report into something you can put your name on. Work through the report and move every claim you intend to keep into a log. Anything that does not survive this pass does not go to the client.
| Claim | Source | Date on the source | Checked | Confidence |
|---|---|---|---|---|
| The statement as it will appear in your deliverable | Publisher and link you opened yourself | Publication or last-updated date | Your initials and the date you opened it | Direct evidence, inference, or vendor claim |
| Example: the regulator’s filing deadline is March 31 | The regulator’s own page | Shown on the page | Yes | Direct evidence |
| Example: the market is consolidating | Two trade articles | Both from this year | Yes | Inference, label it as such |
Three columns matter more than the rest. The date on the source, because a correct citation to an outdated page is still wrong for the client. The checked column, because the only claims you can defend are the ones you opened. And confidence, because a vendor’s description of its own product is not independent evidence.
Step 5: check dates and quotes by hand
- Open every source you cite. Not the summary of it, the page.
- Find the date. Publication, last-updated, or the period the data covers. Put it in the log.
- Match every quotation to the page. Paraphrase if the wording has drifted, and keep quotations short.
- Check numbers at the source. Units, currency, time period and whether a figure is a forecast.
- Look for a newer version. Regulations and pricing pages change without changing their address.
- Drop what you cannot verify. A thinner report you can defend beats a fuller one you cannot.
What it costs and what limits you will hit
| Plan | Price | Deep research usage | Training on your content |
|---|---|---|---|
| Plus | $20 per month, monthly billing only | Plan-dependent allowance shown in the in-product counter | May be used unless you turn training off in Data Controls |
| Pro $100 | Named tier | 5x the Plus usage allowance | Same as Plus |
| Pro $200 | Named tier; new sign-ups paused since September 10, 2026 | 20x the Plus usage allowance | Same as Plus |
| Business, Enterprise, Edu | Per seat | Uses the Work or Codex allowance or credits | Not used for training by default |
OpenAI does not publish a task count per plan. It says usage “varies by plan”, that “Your in-product usage counter shows your remaining tasks”, and that for plans with a fixed monthly allowance it “resets every 30 days from the date of your first use.” Read your own counter before promising a client a research-heavy week. Availability is also not universal: asked whether deep research works in all countries, OpenAI answers “No. Availability depends on your plan and your country or territory.” Source: About ChatGPT Pro tiers.
Confidential client material
Deep research conversations “follow the same data handling and privacy settings as regular ChatGPT conversations.” On consumer plans that means, in OpenAI’s words, that “For Plus and Pro, conversations may be used to improve models unless you turn off training in Data Controls”, while “Business, Enterprise, and Edu content is not used for training by default.” Its consumer data FAQ is blunter still: “Please do not enter sensitive information that you would not want reviewed or used.” Source: Data Usage for Consumer Services FAQ.
- Research the question, not the client. Name the market and the problem, not the client, their pricing or their unannounced plans.
- Upload only what you may share. Files you upload are a documented default source; a signed brief usually is not something to upload.
- Check the engagement terms before any client document goes near the tool.
- Know the retention rule. Results follow the conversation’s retention policy, and deleting the chat also deletes its deep research outputs under that policy, so export what you must keep.
If you are formalising how client material is captured in the first place, our client intake forms guide covers the brief that this research should be based on.
The final human review
- Every kept claim appears in the log with a source you opened and a date.
- Conflicting evidence is shown as a conflict, not resolved silently.
- Inferences are labelled as inference, and vendor statements as vendor statements.
- Nothing confidential from another client appears anywhere in the text.
- The recommendation is yours, written by you, and traceable to the logged claims.
- The method note tells the client how the research was produced and what was not checked.
That method note matters commercially. Clients are entitled to know how the work was done, and a report that states its own limits is easier to defend than one that implies certainty it never had.
Frequently asked questions
Does deep research check whether its sources are correct?
No. It produces citations or source links so you can verify the information. Verification remains yours.
Can I stop it searching the whole web?
Yes. In Sites, then Manage sites, you can restrict research to the sites you list, or prioritise them while still allowing a full-web search.
How many research tasks do I get?
OpenAI does not publish a number. Usage varies by plan and is shown in the in-product counter, resetting every 30 days from your first use on plans with a fixed allowance.
Is it available everywhere?
No. OpenAI says availability depends on your plan and your country or territory.
Can I upload a client’s confidential brief?
Treat that as a contract question first. On consumer plans, conversations may be used to improve models unless training is turned off, and OpenAI advises against entering sensitive information.
Sources and method
Behaviour, controls, limits and plan facts were checked on September 16, 2026 against the OpenAI Help Center articles linked beside each claim. The claim log and review steps are our editorial method, not an OpenAI feature. Plans, allowances and availability change; check your own account before committing to client work.
Read next
