How to Brief Executives With Gemini's Deep Research

Nobody wants the Sunday night version of you: hunched over six browser tabs, a CRM export and a competitor's pricing page, trying to turn it into a slide your VP will read on Monday. Gemini Deep Research, the research agent inside Gemini Enterprise, is built for exactly that job. Rebuilt on Gemini 3.1 Pro, it now reads your internal data stores, arbitrary files you hand it, and remote MCP servers in the same pass, then hands back a cited report instead of a pile of open tabs.

Here is how to point it at a real briefing instead of a demo.

What you'll have by the end

A cited research report, a short audio summary for the walk to the meeting, and a Canvas version you can turn into a one-page visual or export straight to Google Docs for sign-off. Built from your own CRM exports, competitor decks and internal wiki pages, not generic web results.

Point Deep Research at the data that actually matters

In the Gemini Enterprise console, open Deep Research and pick your sources before typing a word of the prompt: your connected data store, and optionally Google Search for outside context. If you're scripting this instead, useful if you want the same briefing regenerated every sprint, the first call to streamAssist generates the research plan against a specific data store:

curl -X POST \
  "https://discoveryengine.googleapis.com/v1/projects/<PROJECT_ID>/locations/global/collections/default_collection/engines/<APP_ID>/assistants/default_assistant:streamAssist" \
  -H "Authorization: Bearer $(gcloud auth print-access-token)" \
  -H "Content-Type: application/json" \
  -H "X-Goog-User-Project: <PROJECT_ID>" \
  -d '{
    "query": { "text": "Summarize Q3 churn risk in our top 10 enterprise accounts and how it compares to the competitor moves in our tracked data store." },
    "agentsSpec": { "agentSpecs": { "agentId": "deep_research" } },
    "toolsSpec": {
      "vertexAiSearchSpec": {
        "dataStoreSpecs": { "dataStore": "projects/<PROJECT_ID>/locations/global/collections/default_collection/datastores/<DATA_STORE_ID>" }
      },
      "webGroundingSpec": {}
    }
  }'

The response includes a session ID and a draft research plan. webGroundingSpec is optional; drop it if the briefing shouldn't leave your own data at all.

Hand it the primary sources directly, don't make it guess

This used to be your job. Deep Research now accepts a combination of PDFs, CSVs, images, audio and video as grounding input, plus connections to arbitrary remote MCP servers, so you can point it at a specialized data stream, like a support ticket system or a competitive intelligence feed, without exporting anything first. Upload the CRM export, the competitor's latest pricing PDF and last quarter's board deck as file inputs so the agent reasons over your numbers, not a summary of your numbers.

Review the plan before it runs

Every session starts with a plan, a list of sub-questions it intends to chase, and it waits for you to edit it. This is the step people skip and regret. If the plan reads "research competitor pricing trends" instead of "compare our Q3 discount rate against the three deals we lost to Acme," fix it now. A vague plan produces a vague report, and executives can smell one from across the room.

Start the research and let it stream

Once the plan looks right, send the second streamAssist call using the session ID from the first response:

curl -X POST \
  "https://discoveryengine.googleapis.com/v1/projects/<PROJECT_ID>/locations/global/collections/default_collection/engines/<APP_ID>/assistants/default_assistant:streamAssist" \
  -H "Authorization: Bearer $(gcloud auth print-access-token)" \
  -H "Content-Type: application/json" \
  -H "X-Goog-User-Project: <PROJECT_ID>" \
  -d '{
    "query": { "text": "Start Research" },
    "session": "<SESSION_ID>",
    "agentsSpec": { "agentSpecs": { "agentId": "deep_research" } },
    "toolsSpec": {
      "vertexAiSearchSpec": {
        "dataStoreSpecs": { "dataStore": "projects/<PROJECT_ID>/locations/global/collections/default_collection/datastores/<DATA_STORE_ID>" }
      },
      "webGroundingSpec": {}
    }
  }'

For anything genuinely complex, like a full account health review across a dozen data sources, switch to Deep Research Max, the variant tuned for extended reasoning instead of speed. It takes longer but chases down more of the sub-questions instead of stopping at the obvious answer.

Turn the report into something an executive will open

The report lands with inline citations back to your source documents, which is what makes it defensible when someone asks where a number came from. From there, use the Create menu in Canvas to turn the report into an infographic or a one-page web view, or export it straight to Google Docs if your exec prefers to comment inline before a decision meeting. Attach the audio summary too; half your VPs will listen to it in the elevator instead of reading the doc.

Gotchas

• Deep Research on Gemini Enterprise is rolling out generally available with allowlist access, so confirm your org has it turned on before promising a live demo in a leadership meeting.
• MCP connections give an agent read access to whatever system you point it at, so route new connections through the same security review as any other integration, especially for CRM or support data.
• The plan-review step is not optional in practice; the default plan is almost always broader than the question your executive actually asked.
• Citations point back to your source documents, not to a guarantee of accuracy, so skim the top two or three before presenting them as fact.

Ship one of these before your next leadership sync instead of building the deck by hand. It's the difference between saying churn is up and showing the four accounts, the numbers, and the source documents, which is the sentence that actually gets you budget.