Deep Research

Deep Research is source-grounded investigation for professionals and teams. Ask a work question, attach library, Drive, or team material when you have it, say what the work needs to produce, and the agent investigates — filling gaps from the open web — then returns a circulatable artifact: a brief, comparison table, cited report, slide deck, or training pack.

It is not a web-search chatbot. The job is a work artifact out: a cited review, a comparison, a one-page brief, or training the team can use. Attach sources when they exist; start from a question when the evidence is public.

Starting a Research Session

Open Research from the Create menu or from Home's Research composer. Both take you to Research, where past investigations sit under a composer.

1. Type the question, or attach sources from your library, a team library, Google Drive, or an upload. Sources are optional when the evidence is public.
2. Choose what to Produce: Brief, Table, Cited report, or Slides. You can pick more than one.
3. Optionally start from a job — cited evidence review, comparison table, one-page brief, or training pack from your sources.
4. Set evidence class, length, and model, then start.

If you still have an old /home?create=research bookmark, that guided source flow still works.

Starting from AI Chat

You can also start a research session from any AI chat. Ask the AI to investigate a work question — for example, "research these vendor claims against our RFP" or "map this year's regulatory changes for leadership." The AI asks what you want produced, then runs the investigation in the background. When the work is ready, it appears inline in the conversation.

Evidence filters

Before starting, you can favor specific evidence classes. Choose one or more from the Evidence control:

• Papers & journals — Peer-reviewed journals, working papers, and scholarly articles.
• Handbooks & manuals — Professional handbooks, codes, manuals, and textbooks used as reference works.
• News & industry — Recent reporting and industry analysis.
• Government & policy — Official publications, statutes, agencies, and public data.

Leave all filters off to search across all source types.

Linking Your Library

Sources are first-class. Click the attach control on the composer to add:

• PDFs and other documents you have uploaded
• Pages you have written
• Library files — saved notes, outlines, CSVs, code, and other text/data files
• Folders from your workspace or a team library
• AI Meeting Notes (with their transcripts)
• Podcasts (with their transcripts)
• Previous research reports

You can also pick Google Drive files directly — Google Docs, Sheets, Slides, PDFs, Word docs, PowerPoints, or text files — without downloading and re-uploading. See Connections.

Attaching Files Directly

You can also attach a file directly into a research session. Deep Research reads PDFs, PowerPoint, Word, and text files, not just images. Drop them into the composer the same way you'd attach an image.

Linked and uploaded sources are primary. The agent researches the open web to fill gaps, not to replace what you attached. Linking a previous report is useful for extending an earlier investigation instead of starting over.

Clarifying Questions

If you already picked a Produce chip or named the artifact in the prompt, preparation can skip the output question. Otherwise the AI asks what you want produced — a cited report, slides, a one-page brief, a comparison table, a training outline, or more than one of those. Options are tailored to your request, and you can pick several or type your own. It may also ask a short follow-up about audience, angle, or timeframe when that would change the result.

If the questions are not relevant, skip them to jump straight into research. A cited report is the default when you skip and did not already name an output.

Report Length

Choose how long the finished work should be from the length dropdown next to the prompt:

• Short — One-page brief.
• Medium (default) — Decision memo.
• Long — Full investigation.

AI Model

Choose which model plans the investigation, analyzes sources, and writes the result from the model picker next to the prompt. Paid plans can pick a specific model; free users can browse the options with locked models shown behind a lock badge. Scholarly remembers the model you pick for Research separately from AI Chat and your other creation tools. See Choosing an AI Model.

Research Plan and Progress

Once research begins, a progress checklist appears showing each step the AI is taking. Each step shows its current state: pending, in progress, or completed.

You can watch in real time as the AI reads your sources, searches the web, evaluates evidence, and extracts key information.

Code Execution

The AI can run Python code, analyze data, and create visualizations during research. This happens automatically when the work involves data processing or when charts and graphs would strengthen the result.

Results: a Verdict and Files

A finished investigation opens on Results — built around what you walk away with rather than a long document:

• Verdict — The answer in a few sentences, with a confidence level, a count of claims where credible sources disagreed, and the questions the evidence could not settle.
• Files — The real deliverables the agent produced: a brief, a comparison spreadsheet or CSV, a findings file, charts, and the working data behind every number it computed (including the analysis script, so you can rerun the math with your own assumptions). Every file opens in its own tab, and each can be saved to your library as a Library file.
• Full report — The complete cited narrative is one click away from the Files list and opens in a new tab. It keeps inline source pills, citations, embedded images, and charts.

The agent creates these files automatically on every run — you do not have to ask. Naming a specific artifact in the prompt or on the Produce row still steers exactly what gets made.

Evidence

The Evidence tab lists every web source and linked document the investigation drew on, so any claim can be traced back to where it came from.

Report Timeline

The sidebar shows a clickable timeline of all reports in the session. Click any entry to jump directly to that report. Linked content from your library is also visible in the sidebar.

Follow-Up Questions

After the initial report is delivered, you can ask follow-up questions in the same session. The AI builds on its previous findings, so follow-ups are faster and more targeted than starting from scratch.

Actions from Research

After a report is complete, you can branch directly into other workflows from the actions above the report:

• Copy report -- Copy the report text to your clipboard.
• Create Page -- Turn the report into an editable page.
• Create Flashcards -- Generate flashcards from the findings.
• Download report -- Save the report to your device.

To expand or refine the report itself, ask a follow-up question in the same session instead -- see Follow-Up Questions above.

Sharing

Share a completed research report with others. Recipients can view the full report including all sections, citations, and visualizations.

Archive and Restore

Archive research sessions you no longer need from the home page. Archived sessions are hidden from your main content list but not deleted. Restore them at any time to bring them back.

Research Index

The research page opens to a composer for a new investigation, with jobs you can start from and recent sessions underneath. Click any past session to jump back in, scan a previous report, or ask follow-up questions. Deleted sessions disappear right away — you don't need to refresh.

Limits

Deep Research draws from your shared AI creation allowance — the same pool that powers flashcards, podcasts, video lectures, slides, and other AI tools. On Free, those are lifetime free credits and do not reset. You can spend your free creations however you like instead of being capped on a separate research limit. Paid plans raise the allowance. See Plans and Limits.