Cortexa AI Glossary · Asking well
What is deep research?
From Cortexa Learn, by Cortexa Consulting. Last checked .
A button that reads for several minutes and hands back a long report. What happens while you wait.
The button
Some chatbots now have a button called deep research. You ask a big question, say which kind of heat pump suits an older house, and it tells you this will take a while. You go and make coffee. When you come back there's a long report, organized under headings, with a footnote on nearly every claim. Pages of it. What was it doing all that time?
An agent that reads
Deep research works like an artificial intelligence (AI) agent: a tool that takes a series of steps on its own toward a goal. It plans what to look for, runs a search, reads what comes back, and then decides what to search for next, again and again. Some tools show you the plan first. Google's Gemini app, for example, lets you edit it before the research starts. When OpenAI launched its version in ChatGPT in February 2025, it said a run could take anywhere from five to thirty minutes.14
Many readers at once
Some versions split the work. Anthropic has described how its research feature does it. A lead agent plans the research, then sends several helper agents to explore different parts of the question at the same time. Each helper reads and reports back, and the lead pulls their findings together. A last step goes through the report and matches its claims to the sources, so each one can carry a citation.3
What you get back
What comes back reads like a briefing: an overview, sections, a summary, and citations beside the claims. OpenAI's guide for developers says its deep research models can find, analyze, and combine hundreds of sources into one report. Some versions can also read files you give them, alongside the open web. That's far more reading than most of us would do on a first pass, and it happened while you were busy with something else.2
Where it helps
Used well, it gives you a head start on jobs like these.
- A first pass on a subject that's new to you.
- A reading list of sources to start from.
- A map of the main arguments, so you can see where people disagree.
In each case it hands you a place to start, and the deciding stays with you. It's also a quick way to learn which questions you didn't know to ask.
Where it slips
There's a catch. A long, polished report can hide a mistake well. OpenAI's own launch post says deep research can sometimes make up facts or draw the wrong conclusion, and can struggle to tell solid information from rumor. One wrong figure on page six looks just as confident as the right ones around it. And a footnote only shows where a claim came from. The page it points to may not say what the sentence says.1
Open what you'll rely on
You don't need to check everything. Use the report as a map, and check the parts you'll lean on. Pick the two or three claims your decision depends on, open the sources behind them, and read those lines yourself. Which claim in your last long AI answer would you check first?