Reading time: 12 minutes
ChatGPT deep research is a mode where the AI plans a search, reads many sources and delivers a report with citations. Gemini Deep Research and the research modes in Perplexity and Claude follow the same logic. So only a brand that the agent finds and can read makes it into the report.
In short
- According to OpenAI, ChatGPT deep research reads and analyzes many sources before it delivers a report with citations. In the API, the company mentions hundreds of sources.
- In a Peec study across more than 10 free accounts, ChatGPT deep research searched only on Bing, never clicked and read about 5,700 characters per page on the first read.
- Gemini deep research uses Google Search as its default source. Perplexity and Claude have their own bots, with different robots.txt rules.
- To get into the report, a brand must allow the bots, put the answer at the top of the HTML and earn mentions in sources the AI treats as official.
What is ChatGPT deep research?
ChatGPT deep research is a long-research mode. You describe the result you want. Then ChatGPT proposes a plan and searches the web, files and connected apps. Finally, it delivers a structured report, with citations or links to the sources. So it takes minutes, not seconds.
The OpenAI Help Center describes the process in steps. Before it starts, ChatGPT may ask clarifying questions and shows the plan, which you can edit. Then, during the research, you can follow the progress and interrupt it to change the focus. Finally, the report opens in full screen, with a table of contents, a sources section and an activity history.
Since February 2026, users can also focus the research on specific sites. Under Sites > Manage sites, they can restrict the search to the domains they enter, or only prioritize them while the search still covers the web. So a buyer can, for example, ask for a report based only on review sites.
How to use deep research in ChatGPT
According to OpenAI, there are three ways to start:
- Command: type /Deepresearch in the message box.
- Tools menu: click + and choose Deep research.
- Side menu: select Deep research.
Usage of ChatGPT deep research varies by plan, and the counter in the product shows how many tasks remain. On the pricing page, the Free plan lists limited deep research. However, OpenAI warns that availability also depends on the country.
How many sources does deep research read, and which sites does it prefer?
OpenAI does not set a number of sources for ChatGPT. In the API documentation, however, it says its deep research models can find, analyze and synthesize hundreds of sources. So that number applies to the API models, not to the product.
The most detailed picture comes from Peec, which sells a tool for tracking brands in AI. The company recorded ChatGPT deep research in more than 10 free accounts, with about 20 requests per account, covering research and supplier comparisons. According to the study, published in June 2026, every logged search used Bing, with no Google queries. OpenAI, however, does not publish where the results come from.
So the numbers show how deep research filters sources:
- Bing snippet: the summary of each result has at most about 285 characters, which is almost all the agent sees before it opens the page.
- First read: each opened page yields a window of about 5,700 characters (median), never the whole page.
- Depth: once it knows what it needs, the agent reads only the top 2 or 3 results of each search.
- Re-reading: the most common trigger is finding the exact word the agent was looking for on the page.
As for preferences, the logged reasoning shows the agent looking for sources it calls official and reliable. On the other hand, it treats forums and user content as background, not as proof. This reasoning is what the agent writes itself, not a rule OpenAI published. In addition, Peec warns that the study is observational and that the numbers change with each version.
How is deep research different from regular search and agent mode?
Regular search answers fast, with a summary and links. ChatGPT deep research takes longer, reads many sources and delivers a report. An agent, in contrast, acts on sites: it opens pages in a browser, clicks and fills in forms. In ChatGPT, that role now belongs to ChatGPT Work, because the old agent mode is gone.
OpenAI sums up the difference between the first two modes like this:
“Search quickly pulls recent web information and returns a short summary with links. Deep research takes more time to read and analyze many sources, then produces a detailed, documented report.”
OpenAI, Help Center (2026)
| Mode | Role | What it means for a brand |
|---|---|---|
| Search | Quick summary with links | Be among the cited sources |
| Deep research | Report with many sources | Be read and cited in the report |
| ChatGPT Work | Long tasks across sites and apps | Site open to OpenAI’s browser |
The technical difference also matters. In the Peec study, ChatGPT deep research used only three commands in a text browser: search, open and find inside the page. So it never clicked, although it opens the links it reads. The cloud browser in Work, on the other hand, clicks buttons and fills in forms, according to OpenAI. The case of regular search inside an answer is covered in the guide on whether ChatGPT searches the internet.
Does Gemini deep research work the same way?
It works in a similar way, but on a different base. Gemini deep research, called Deep Research, uses Google Search as its default source. Users can swap or add sources such as Gmail, Drive and files. It also builds an editable plan before it starts, and it can take up to 10 minutes, or longer for complex reports.
According to the Gemini Help Center, all users can use the Thinking model for deep research. However, Google AI Pro and Ultra subscribers can use the Pro model and get higher limits. On the other hand, Google’s limits page warns that, without a plan, the feature may be unavailable during periods of high demand. On its overview page, Google also says Deep Research browses hundreds of websites.
So the consequence for a brand is direct. Gemini deep research starts from Google Search, so the page must be in Google’s index to be found there. In ChatGPT, Peec saw Bing in every deep research search. Other studies show a more mixed origin, as the guide on how to index your site on Bing explains.
ChatGPT or Gemini deep research: which one cites more sources?
An academic test compared them. In DeepResearch Bench, from the University of Science and Technology of China, experts created 100 PhD-level research tasks across 22 fields. A valid citation, in the test, is one that the source actually supports. With the April and May 2025 versions, the result was this:
| Tool | Valid citations per task | Accuracy rate |
|---|---|---|
| Gemini 2.5 Pro Deep Research | 111.2 | 81.4% |
| OpenAI Deep Research | 40.8 | 78.0% |
| Perplexity Deep Research | 31.3 | 90.2% |
So Gemini cited far more, while Perplexity was right more often. However, all three products have changed since then. So use the test as a sign of style, not as a current scoreboard.
What about Perplexity and Claude research?
They also read many sources, but with other bots and other access rules. Perplexity calls the feature Research and says it runs dozens of searches and reads hundreds of sources. In Claude, research chains several searches across the web and connected tools, and it exists only on paid plans.
According to the Perplexity Help Center, most tasks finish in under 3 minutes. Free users get limited access, while Pro users get extended access. At the end, the report can become a PDF, a document or a Perplexity Page.
At Anthropic, Claude research is available on the Pro, Max, Team and Enterprise plans, on the web, desktop and mobile. It requires web search to be on and follows the same usage limits as regular chats, although it uses up the limit faster. On mobile, however, it searches only the web, without connectors.
Does deep research follow robots.txt and paywalls?
It depends on the company, and the documentation does not always say which bot each mode uses. OpenAI does not name the bot for ChatGPT deep research. The only public clue is Peec’s log, in which pages blocked for OAI-SearchBot came back empty. Paywalls, logins and slow servers can also drop a source from the report.
| Tool | Documented bots | What the company says |
|---|---|---|
| ChatGPT | OAI-SearchBot and ChatGPT-User | robots.txt may not apply to ChatGPT-User |
| Gemini | Google-Extended token | Controls training and grounding in Gemini apps |
| Perplexity | PerplexityBot and Perplexity-User | Perplexity-User generally ignores robots.txt |
| Claude | Claude-SearchBot and Claude-User | The bots follow robots.txt |
The data comes from OpenAI’s bot documentation and the equivalent pages from Google, Perplexity and Anthropic. None of the four companies says, on those pages, which bot the long-research mode uses.
On paywalls, OpenAI itself warns that ChatGPT may not be able to read a site because of technical issues, paywalls or robots.txt rules. In the Peec study, social networks with login, large marketplaces and slow servers were also left out, and the report gave no warning. Since the log comes only from Peec, treat OAI-SearchBot as a clue, not as a rule OpenAI confirmed. The rule for each bot, with a sample file, is in the guide on robots.txt for AI.
How does a brand get into the supplier report?
A brand gets in when the agent finds the page, reads the answer at the top and sees the brand name in sources it considers reliable. This matters in B2B buying, because AI already influences the supplier shortlist. In a G2 survey of 1,076 buying decision makers, 54% named AI chatbots as a source that influences the software shortlist.
G2 sells marketing services to software vendors, and the survey covers chatbots in general, not only deep research. It also left out Latin America, so it does not describe Brazil.
Changes on your site
- Allow the search bots: OAI-SearchBot, PerplexityBot and Claude-SearchBot in robots.txt, and their IPs in the CDN and firewall. Peec tested only OAI-SearchBot. On Google, anyone who wants to appear in Gemini apps should not block Google-Extended.
- Put the answer at the top of the HTML: according to Peec, only about 33% of the first read was content on pages with 60 or more links, versus 78% on pages with fewer than 20.
- Write in text what sits in tabs and filters: price, specifications, industries served and certifications. After all, ChatGPT deep research does not click.
- Use the buyer’s exact words in titles and text, such as “ERP for manufacturing”, because re-reading starts with a term search inside the page.
- Describe images in the alt text: Peec saw the agent pull a rating from a badge’s alt text and skip images without a description.
- Use descriptive internal links to price, specifications and comparisons, since the agent follows the links it reads.
Changes outside your site
Since the agent looks for sources it describes as official, it also helps to be mentioned by third parties it recognizes. So work on review sites, trade press and associations that appear in reports for your category. Which sources those are, and how to earn a presence in them, is in the guide on the sources ChatGPT cites most.
Then, test. Run a deep research in ChatGPT and Gemini with the request your buyer would make. In ChatGPT, open the sources section and the activity history, and note which domains the AI read and which it cited. Finally, compare with the competitors that made the report. The brand’s other routes, from training to search, are in the guide on how ChatGPT works.
Frequently asked questions
Yes, with a limit. OpenAI’s pricing page shows limited deep research on the Free plan, and usage varies by plan. The counter inside ChatGPT shows how many tasks remain, but availability also depends on the country. Gemini deep research also works without a paid plan, but it may be unavailable during periods of high demand.
Yes, but the site may drop out of other answers too. In Peec’s log, ChatGPT deep research respected the block on OAI-SearchBot, the bot that puts a site in ChatGPT search. OpenAI does not confirm this link. Even so, blocking that bot removes the page from ChatGPT search and, from what Peec saw, from the report too.
From a few minutes to tens of minutes. Perplexity says most tasks finish in under 3 minutes. Google says Gemini deep research can take up to 10 minutes, or longer for complex reports. In the API, OpenAI warns that deep research models can take tens of minutes.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- OpenAI, Deep research in ChatGPT (2026).
- Peec, how ChatGPT deep research reads a site, from session logs (2026).
- OpenAI, deep research documentation in the API (2026).
- Google, Use Deep Research in Gemini Apps (2026).
- OpenAI, overview of OpenAI crawlers (2026).
- DeepResearch Bench, academic comparison of research agents (2025).
- Perplexity, What is Research mode? (2026).
- Anthropic, Use research on Claude (2026).
