Prompting Gemini Deep Research

TL;DR: Gemini Deep Research isn't one thing. It's a mode in the Gemini app that plans and runs a multi-step research task into a cited report, and separately a preview agent in the Gemini API. Both differ from an ordinary Gemini prompt and from NotebookLM. Here's what each surface documents, and how to scope a prompt so the report stays useful.

What is Gemini Deep Research, and where does it actually live?

Two documented surfaces, not one, and they are easy to conflate because they share a name and, per Google's own account, the same underlying infrastructure.

The consumer surface is inside the Gemini app. Google's support article describes it plainly: "You can conduct in-depth and real-time research on almost any subject with Deep Research in Gemini Apps. By default, Gemini includes Google Search as a source for your research. You can change or add other sources, like your personal Gmail or Drive, for your research. You can also upload files and add NotebookLM notebooks" (support.google.com/gemini/answer/15719111, accessed September 3, 2026).

The developer surface is a separate, API-only product: the Gemini Deep Research agent. Google's developer documentation calls it exactly that and states it is currently in preview: "The Gemini Deep Research agent autonomously plans, executes, and synthesizes multi-step research tasks... Preview: The Gemini Deep Research agent is currently in preview. The Deep Research agent is exclusively available using the Interactions API. You cannot access it through generate_content" (ai.google.dev/gemini-api/docs/interactions/deep-research, accessed September 3, 2026). It ships in two named versions. Google's own supported-versions list describes deep-research-preview-04-2026 as "Designed for speed and efficiency, ideal to be streamed back to a client UI." It describes deep-research-max-preview-04-2026 as "Maximum comprehensiveness for automated context gathering and synthesis" — the kind of thing you'd point at a nightly batch job rather than a live chat.

A third, business-only surface exists too: Gemini Enterprise ships its own Deep Research agent over a company's internal data, documented separately from both of the above. Most people searching for "Gemini Deep Research" want one of the first two, so that's this post's focus.

Google's own framing connects all of them: "When you build with the Deep Research agent, you are tapping into the same autonomous research infrastructure that powers research capabilities within some of Google’s most popular products like Gemini App, NotebookLM, Google Search and Google Finance" (blog.google, "Introducing Deep Research and Deep Research Max", published April 21, 2026, read here). One engine, several front doors.

How do you start a Deep Research report in the Gemini app?

The documented flow, step by step, straight from Google's support article: open the text box on gemini.google.com, click Add Files, then Deep Research. Optionally add files or choose sources — Gmail and Drive are only available once Google Workspace is connected. Google Search is on by default; deselect it to limit research to only the other sources you picked. Enter your question, click Submit, and Gemini returns a proposed research plan. You can click Edit plan to change it before anything runs, then click Start research.

On timing, Google is specific: "It usually takes about 5-10 minutes to generate the report since Gemini analyzes many sources. For more complex reports, it may take longer" (same support article). You can leave the chat while it runs; Gemini notifies you when the report is ready.

Two model choices are documented, not one: "Google AI Pro and Google AI Ultra users can generate reports using Pro for even higher quality. All users can use Thinking for their reports." And one visual feature is gated specifically to the top tier: with a Google AI Ultra plan, reports "can include animations and visuals" — charts, diagrams, interactive simulators — but this "isn’t available if you include any Google Workspace services, like Gmail and Drive, as a source." Choosing your own files over the open web costs you the visuals, at least for now.

What changes when you prompt the Deep Research agent through the API?

Everything about the interaction shape. The consumer app is a guided, synchronous-feeling flow; the API agent is explicitly asynchronous. Google's docs are direct about this: research tasks "involve iterative searching and reading and can take several minutes to complete. You must use background execution (set background=true) to run the agent asynchronously and poll for results or stream updates."

A request looks like this, using the Interactions API rather than the ordinary generate_content call:

POST https://generativelanguage.googleapis.com/v1beta/interactions
{
  "input": "Research the history of Google TPUs.",
  "agent": "deep-research-preview-04-2026",
  "background": true
}

You then poll the returned interaction ID until its status is completed or failed. This is a genuinely different prompting shape from a normal Gemini call: your "prompt" is just the input field naming what to research, and the actual work happens across many internal search-and-read steps you don't write yourself.

The API also exposes a step the consumer app doesn't document in the same way: collaborative planning. Setting collaborative_planning: true on the first call makes the agent return a proposed plan instead of a full report, which you can then approve, edit, or reject across further turns before the actual research runs — a more explicit version of the app's "Edit plan" button, built for a program to drive rather than a person clicking a button.

Google's own announcement of the Deep Research and Deep Research Max versions also documents composability that the consumer app doesn't expose directly: "Combine the full suite of Gemini API tooling. Run Deep Research with Google Search, remote MCP servers, URL Context, Code Execution and File Search simultaneously — or turn off web access entirely to exclusively search over your custom data" (blog.google, published April 21, 2026). Turning off web access is the API's version of deselecting Google Search in the app, just with a wider set of tools it can be combined with instead.

How is this different from NotebookLM?

By default, they start from opposite places. NotebookLM answers only from sources you explicitly loaded into a notebook — a closed corpus you control, and prompting NotebookLM well means writing for that constraint rather than fighting it. Deep Research defaults the other way, to the open web via Google Search as its first source, with your own files as something you add on top, not the starting point.

That said, Google's own consumer documentation blurs the line on purpose: you "can also upload files and add NotebookLM notebooks" as sources inside a Deep Research report. So a NotebookLM notebook can become one input to a broader Deep Research task, but a Deep Research report is not a NotebookLM notebook, and the reverse doesn't hold — NotebookLM does not run open-web Deep Research on its own. If your actual need is answer only from these specific documents, nothing else, that is NotebookLM's job by design. If it's read the open web plus these documents and write me a report, that's Deep Research with sources added.

How do you scope a Deep Research prompt so it doesn't wander?

The two real levers are sources and the plan, and both are worth using deliberately rather than accepting the defaults.

On sources: if you want the report grounded only in things you control, deselect Google Search and add exactly the files, Gmail, Drive folders, or NotebookLM notebook you want it to read. Google's own next-generation announcement frames the equivalent API option the same way, calling out that you can "turn off web access entirely to exclusively search over your custom data" when that's the goal.

On the plan: don't accept the first proposed plan by default. Read it, and edit it before you click Start research (or, on the API, before you approve a collaborative_planning turn). A plan that's too broad produces a report that's shallow everywhere; a plan you've narrowed to specific sub-questions produces one that's deep where you actually need it.

Research [topic]. Before you start, propose a plan that breaks this
into no more than five sub-questions. Prioritize [named source type,
e.g. peer-reviewed studies, SEC filings, vendor documentation] over
general web results, and flag anywhere the sources disagree instead
of picking one silently.

Restrict this research to the files and notebook I've added — do not
use Google Search as a source. If you can't answer something from
those sources alone, say so instead of filling the gap from the web.

Where should the context go in a long Deep Research prompt?

This is one place where the two vendors publish opposite advice for their own models, and it's worth knowing before you paste a wall of background material into either one.

Google's own Gemini prompting guide is explicit: "When providing large amounts of context (e.g., documents, code), supply all the context first. Place your specific instructions or questions at the very end of the prompt." It follows that with a transition tip, telling you to "use a clear transition phrase to bridge the context and your query" once the data block is done, giving an example opener along the lines of Based on the information above (ai.google.dev/gemini-api/docs/prompting-strategies, accessed September 3, 2026).

OpenAI's own prompt engineering guide describes the "Context" component of a prompt — the same kind of supporting material Google's guidance is about — the other way round: "Give the model any additional information it might need to generate a response, like private/proprietary data outside its training data, or any other data you know will be particularly relevant. This content is usually best positioned near the end of your prompt, as you may include different context for different generation requests" (developers.openai.com/api/docs/guides/prompt-engineering, accessed September 3, 2026).

Line the two up and they point opposite directions for the same block of material. Google's structure is context block, then question, in that order, with the question last. OpenAI's structure puts context near the end of the whole prompt, after identity and instructions — closer to context, then nothing, since the context effectively closes the prompt rather than a question following it. Neither guide is wrong; each is written for that vendor's own model and its own documented prompt structure, and neither claims to describe the other's. The practical move for a Deep Research brief with a large block of background material: follow the placement the vendor whose model you're actually prompting documents, rather than assuming one universal rule carries over from the other.

How does Gemini Deep Research compare to ChatGPT Deep Research, honestly?

Both products do the same basic job, a structured, multi-source, cited research report instead of a chat reply, and both are documented in enough detail to compare fairly on a few specific points, without guessing at the rest.

The gaps in that table are as informative as the matches. OpenAI's help center explicitly documents mid-run interruption: "You can follow progress as it runs and interrupt at any time to refine the focus, including adjusting which sources it can access." Google's consumer support article doesn't describe an equivalent control. That doesn't mean Gemini definitely lacks it; it means Google hasn't published one, which for a fast-moving feature is worth treating as a real gap rather than filling in an assumption either way.

On timing specifically, OpenAI's current help article is worth flagging on its own: it states usage "varies by plan" and points to an in-product counter, with no minute-range figure anywhere on the page as of this writing — a change from an earlier 2025 launch post that did quote a range. Google's page, by contrast, states a specific range today. Neither fact says one product is faster; it says one vendor currently publishes a number in this specific spot and the other doesn't.

Copy-paste Deep Research prompt starters

Three you can adapt directly, one for a broad first pass, one scoped to your own sources, and one that forces the plan to stay narrow.

Research [topic] and produce a report covering: current state, the
main disagreements between sources, and what's changed in the last
six months specifically. Flag anything you couldn't confirm.

Using only my added sources — do not use Google Search — summarize
what these documents say about [topic]. If the documents don't
cover something I ask, tell me that directly instead of guessing.

Before running, propose a plan with exactly four sub-questions on
[topic]. Wait for me to approve or edit the plan before you start
the actual research.

Honest note on where a prompt tool fits here

We build a prompt generator, not a research agent, and we don't run web searches or produce cited reports ourselves. What a prompt can genuinely improve on this page is narrow: naming the sub-questions before you accept the default plan, stating explicitly whether the open web should be in scope, and asking for the disagreement between sources instead of a single flattened summary. If you find yourself writing the same scoping instructions into every Deep Research prompt, that's the same problem Gemini Gems solve for ordinary Gemini chats, just applied to a report instead of a reply — worth saving once you've settled on the wording rather than retyping it. What no prompt can do is make either product's citations accurate on their own, or resolve which vendor's page is currently right when two of that vendor's own pages disagree. That part stays yours: read the plan before you approve it, and open the sources before you repeat what the report says about them.