Deep Research is the mode where a chatbot stops answering instantly and instead drafts a plan, fires off dozens of chained searches and hands back a long, sourced report. It pays off for questions that require crossing dozens of pages — comparing vendors, getting up to speed on an unfamiliar field, or mapping what has already been published on a topic. For a single fact, it is wasted time and quota. Notably, none of the three major providers publishes how many reports you can request per month, which says a lot on its own.
- Method over speed: Deep Research plans, searches in multiple rounds and decides what to inspect next based on what it finds, unlike a standard search that reads a few pages and summarizes.
- Quotas are opaque: ChatGPT, Gemini and Claude all gate the feature behind vague plan tiers or daily caps, and only Google publishes a time estimate of 5 to 10 minutes per report.
- Verification is mandatory: A 20-page report with dozens of links looks rigorous even when it is not, and almost nobody opens the links. Correcting the plan and checking the three claims your decision depends on is the cheapest safeguard.
What actually changes with Deep Research
The difference is one of method. A normal search queries a handful of pages and summarizes; Deep Research plans, searches across several rounds and decides what to look at next based on what it is finding. OpenAI frames it plainly: search brings recent information and returns a short summary with links, while Deep Research takes longer because it reads and analyzes many sources and produces a documented report.
All three follow a similar blueprint, with nuances that matter:
- ChatGPT proposes a research plan you can review and change before it starts, lets you follow progress and interrupt to redirect it. It can work with the public web, files you upload, connected apps and specific sites: under Sites > Manage sites you can restrict it to certain domains or merely prioritize them.
- Gemini uses Google Search as its default source, and you can add Gmail, Drive, your own files or NotebookLM notebooks. It also generates a plan, with a button to edit it before pressing start. Users must be over 18.
- Claude calls the feature simply Research. Anthropic describes it as a series of searches that build on one another and explore different angles of the question. It requires web search to be enabled and can use your connectors, such as Gmail, Google Calendar or Google Docs.
One Claude detail worth knowing: if you launch research from the mobile app, it only searches the web and ignores your connectors, and Anthropic’s own help documentation acknowledges you will see no notice that they were skipped.
Timing and quota realities
Only Google gives a time figure: a report usually takes between 5 and 10 minutes, and longer if the research is complex. You can leave the conversation meanwhile and Gemini notifies you when it finishes. OpenAI and Anthropic publish no duration in their help pages; OpenAI merely recommends search or normal chat if you are in a hurry.
Limits follow a similar pattern. As of late October 2026, ChatGPT offers the feature on Free and Go plans in a “limited” form, on Plus as “expanded” and on Pro as “maximum”; a counter in the app shows remaining tasks, and where a fixed monthly allowance applies it renews every 30 days from first use. It is not available in every country. Gemini makes it available to all users, with Google AI Pro and Ultra getting the Pro model and a higher cap, under a daily limit on research runs and simultaneous research, with a warning when you approach it. Claude restricts it to paid plans — Pro, Max, Team and Enterprise — with the same limits as a normal conversation, but consuming them faster.
Google’s limits page adds a detail that explains why quota disappears so quickly: it classifies Deep Research as a high usage-cost task, on par with video generation, because it chains searches and several rounds of reasoning. Its general limits reset every 5 hours up to a weekly ceiling.
The takeaway from that table: “limited,” “expanded” and “maximum” are adjectives, not limits. If you plan to pay for a tier with this feature in mind, the only reliable number is the counter the app shows once you are inside.
How to audit the report before trusting it
A report with citations is not a verified report. Citations tell you where the model says it got each item; confirming the source actually says that remains your job. Here is how to use each tool:
- Fix the plan before it starts. This is the cheapest moment to prevent errors: if the plan includes angles you do not care about or misses sources that matter (an official body, a specific database), change it there. ChatGPT and Gemini allow this explicitly.
- Narrow the sources when the topic demands it. For a legal or health question, restricting research to official domains in ChatGPT, or unchecking Google Search in Gemini to work only with your documents, cuts noise from blogs and aggregators.
- Use the sources section, not just the text. In ChatGPT the final report opens full-screen with an index, a sources used section and an activity history showing how the research progressed. Check whether sources are primary or outlets repeating other outlets.
- Open the three claims your decision depends on. You do not need to check all forty citations, but you do need the figures and conclusions you will act on, and that the linked page truly says what the report attributes to it. A real link does not guarantee a real sentence.
- Distrust studies and laws you did not know. If the report cites a paper or rule that is unfamiliar, find the original document before repeating it.
For preserving the work: OpenAI lets you download the report in Markdown, Word and PDF, and if you delete the chat, you delete the report. Gemini exports to Google Docs, but you will only find older reports if Keep activity is enabled.
Deep Research shifts the work from searching to reviewing, and that shift only pays off if you actually review. The most common mistake is not using it for the wrong questions, but treating the report as a finished product when it is a very well-documented first draft. Used as a starting point that gets cross-checked, it saves hours. Used as a final answer, it turns a model’s error into your error, complete with a bibliography.