A Research Workflow That Cites Itself

Research fails in two ways: drowning in sources, and trusting a confident wrong answer. This workflow attacks both, and every claim that survives it carries its receipts.

1. Sources into Drive

Everything goes into one folder: PDFs, exported pages, notes. Drive makes the corpus referenceable from any chat, which is what turns "I read that somewhere" into a file the model can quote.

2. Ask with citations

Web search answers with inline citations linking to origin URLs, on every model. The discipline: claims without a source you can click do not enter the notes. For contested questions, ask twice on different models and note where they diverge; divergence is your verification worklist.

3. Fan out the reading

One subagent per source, same brief: claims, evidence quality, contradictions with the rest. The orchestrator merges findings into a matrix with references intact. Fifty sources stop being a week of serial reading; the subagent guide covers the pattern.

4. Verify the spine, then write

Identify the handful of claims the conclusion stands on and re-ask each across two or three models with the sources attached. Keep what survives; flag what does not. Then draft: one model writes from the matrix, and the draft's citations trace back to files in your folder.

The report lands in Drive next to its sources: the deliverable and its evidence, one folder, re-runnable next quarter with an automation.

Good to know

• The deep research workflow page is this loop productized, with scenes of each step.
• Models disagree most on recent events and quantitative claims: exactly where citations matter most.
• For fields with hard accuracy stakes, the workflow narrows checking; it never replaces professional judgment.

Run it once on a question you already answered the hard way and compare: the researchers page shows who lives in this loop daily.