ChatGPT Deep Research: How It Works and Cites

ChatGPT deep research is an agent inside ChatGPT that plans a multi-step investigation, reads many web pages, files and connected sources, and returns a structured report with citations. OpenAI launched it on February 2, 2025, describing a task that “accomplishes in tens of minutes what would take a human many hours.” It is slower than a normal search answer on purpose, and that makes it useful for market and competitor research, and worth understanding if you want your pages to be among the sources it cites.

This guide covers how deep research works according to OpenAI’s own pages (checked September 22, 2026), what the limits are, a workflow for using it on marketing questions, and what makes a page citable by an agent that reads dozens of sources before it writes.

On this page

• What is ChatGPT deep research?
• How a deep research task runs
• Deep research limits and plans
• When to use search vs deep research
• Using ChatGPT deep research for marketing research
• How does deep research choose what to cite?
• Becoming a source ChatGPT deep research cites

What is ChatGPT deep research?

Deep research is ChatGPT’s mode for questions that need many sources combined into one documented answer. OpenAI’s help article on deep research describes it as helping you “accomplish complex online tasks by reasoning, researching, and synthesizing information into a documented report.” It can work with uploaded files, search the public web or specific sites, and use enabled ChatGPT apps.

At launch, OpenAI’s announcement said it was powered by a version of its o3 model “optimized for web browsing and data analysis,” able to read text, images and PDFs, and to pivot “as needed in reaction to information it encounters.” Today the help page says it runs on “the latest models” by default, with legacy models still selectable.

The difference from ordinary ChatGPT search is depth, not kind. Search runs a few queries and writes a short answer. Deep research runs a plan, reads far more, and hands back a report you can check line by line.

It is not unique to OpenAI. Perplexity’s help center describes its own Research mode, which “performs dozens of searches automatically, reads hundreds of sources” and delivers a report “in just 2-4 minutes” (How does Perplexity work?). The products differ in speed, source controls and output format, but they share one property that matters to site owners: they read many more pages per question than a normal answer does, and they cite what they use. For a business, that means a single research task can compare you against every competitor it finds, using whatever each of you has published.

How a deep research task runs

A deep research task moves through six stages, and you can steer it at three of them. The diagram below follows OpenAI’s documented flow.

A few details from the help page matter in practice:

• Starting it. Type /Deepresearch in the prompt, choose Deep research from the tools menu (+), or pick it from the sidebar.
• Sources. By default it reads the public web and your uploaded files. Under Sites > Manage sites you can restrict it to specific domains, or choose “Prioritize these sites, but allow full-web search.”
• Apps. It can use supported connected apps such as document stores, but only their read actions. Connections do not override the provider’s own permissions.
• Output. Every report includes “citations or source links,” a table of contents, a sources-used section and an activity history.

Deep research limits and plans

Deep research limits depend on your plan, and OpenAI has changed them several times, so treat any number as dated. The timeline below lists the changes OpenAI has published.

The practical rule: check the usage counter in ChatGPT rather than trusting a number from a blog, including this one. The help page says fixed monthly allowances reset “every 30 days from the date of your first use,” not on the first of the month. Availability also “depends on your plan and your country or territory.”

When to use search vs deep research

Use search for quick facts and deep research for decisions that need many sources. OpenAI’s help page puts it the same way: search “quickly pulls recent web information and returns a short summary,” while deep research “takes more time to read and analyze many sources.” At launch OpenAI said a task may take 5 to 30 minutes.

Using ChatGPT deep research for marketing research

Deep research works best for marketing when you give it a decision, a scope and an output format, then verify what comes back. Here is a workflow that holds up for a competitor or market scan.

1. Write the decision, not the topic. “Help us decide which three competitors to benchmark our pricing page against” beats “research our competitors.”
2. Fix the scope. Market, geography, date range, and company size. Deep research follows constraints well when they are explicit.
3. Pick sources on purpose. Use Manage sites to prioritize primary sources such as competitors’ own sites, regulators or official documentation. Restrict fully only when you are sure the answer lives there.
4. Ask for a table with a citation per row. It makes the report easy to audit.
5. Answer the clarifying questions carefully. The plan is built from them. Edit the plan if it drifts.
6. Verify before you use it. OpenAI’s launch post says deep research “can sometimes hallucinate facts” and “may struggle with distinguishing authoritative information from rumors.” Open every source behind a number you plan to repeat.

An example prompt for a local service business:

Research how the five highest-rated residential roofing companies in Tampa, Florida present their services online. For each, find the services listed, whether prices are published, the review platforms linked, and any certifications claimed. Prefer each company’s own website. Output one table with a source link per cell, then a short list of gaps none of them cover.

What you get back is a draft, not a finding. Treat the report the way you would treat a junior analyst’s first pass: useful for structure and for surfacing sources you had not seen, weak wherever it had to infer. Three habits help:

• Keep the activity history. It shows which searches ran and which pages were read, so you can spot a question it never asked.
• Re-run with different constraints. A second task restricted to primary sources can disagree with the first on details, and the disagreements are where your own checking should go.
• Save the export. Markdown or Word files make it easy to annotate, verify and share the parts you trust.

Two cautions. Deep research can’t give you reliable search volumes or rankings, which come from keyword tools, not the web. And it reflects what is published, so a competitor with a thin website may look weaker than it is. For using ChatGPT on other SEO tasks, see ChatGPT for SEO, and for the classic version of this exercise, SEO competitor analysis.

How does deep research choose what to cite?

OpenAI does not publish how deep research ranks or selects sources, so any precise answer is a guess. What the documentation does say narrows it down.

• It can only cite what it can read. OpenAI’s help article on accuracy lists “lack of access” as a limit: the model “may not be able to obtain relevant information from a given website due to technical issues, paywalls or preferences set via robots.txt” (Does ChatGPT tell the truth?).
• It searches like ChatGPT search, at greater depth. OpenAI does not say which of its crawlers deep research uses. Its bots page says OAI-SearchBot decides whether a site appears in ChatGPT search answers, and ChatGPT-User handles user-initiated actions. Keeping both unblocked is the safe setting.
• It rewards passages that can be quoted. A report with a citation behind each claim needs sources that state claims plainly. That is the same retrieval-then-selection logic described in retrieval vs citation, stretched over more steps.

Becoming a source ChatGPT deep research cites

To become a source ChatGPT deep research cites, make your pages reachable, readable and specific enough to support a sentence in someone else’s report. The diagram below turns that into five checks.

In practice:

• Check access first. Test that your CDN or firewall is not challenging OpenAI’s agents; the full list of user agents is in robots.txt for AI crawlers.
• Publish the numbers people research. Pricing ranges, timelines, specifications, service areas. A research agent comparing vendors cannot quote what you keep behind a sales call.
• Write sections that stand alone. Lead with the answer and keep the claim and its evidence together, as covered in content chunking for AI.
• Get the same facts repeated elsewhere. A research agent that cross-checks sources favors facts it can confirm. Coverage in trade press and directories does that work, which is the argument of digital PR for AI search.

None of this guarantees a citation, and OpenAI does not offer a way to request one. If you want a structured view of where your brand is and isn’t cited across ChatGPT and other engines, our GEO service starts by testing the questions your buyers research and mapping which sources win them.