Source-Directed Deep Research
Connect the right systems and explicitly direct AI to the evidence.
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 96%
Source-Directed Deep Research treats connected business systems as an evidence layer rather than assuming the AI will automatically choose the right source. First, authorize and enable the relevant systems, such as HubSpot, Gmail, Calendar, or Google Drive. Then explicitly tell the model which sources to inspect and what problem their combined evidence should answer. This creates a clear input-to-output mechanism: selected internal sources provide the evidence, the business question constrains the investigation, and deep research produces a consolidated analysis. When a question spans systems, the prompt should say how to combine them—for example, comparing CRM outcomes with calendar activity. The final report should be reviewed for whether it actually used the requested sources and addressed the stated decision.
Origin
Extracted from Marketing Against The Grain during a demonstration of ChatGPT deep research connected to HubSpot and other workplace systems.
Core principles
- 01Analysis improves when the model can access authoritative internal data.
- 02Connected sources still need explicit routing instructions.
- 03Multiple systems can be combined when the question crosses functional boundaries.
- 04A focused problem statement produces a more useful research report.
How to run it
- 1
Connect the Systems
Authorize the internal systems containing the evidence needed for the investigation. Confirm that each intended source appears as available in deep research.
Pro tip Connect only sources that materially inform the question.
Watch out A connected source may expose broad organizational data, so follow applicable access controls.
- 2
Enable the Sources
Turn on deep research and select the relevant sources before submitting the request. Verify that the CRM or other core system is visibly enabled.
Watch out Do not assume authorization alone means the source is active for the current research run.
- 3
Route the Research
Name the systems the model should search and describe the problem it should solve. If evidence must be combined, state that relationship explicitly.
Pro tip Use language such as “combine the data from HubSpot and G Drive.”
Watch out Leaving source selection implicit can make the model overlook the most important system.
- 4
Review the Report
Check that the report answers the business question and reflects evidence from the requested systems. Refine the prompt if a source or relevant data class was missed.
Pro tip Ask for evidence grouped by source when provenance is difficult to assess.
In the wild
A sales leader directs deep research to combine HubSpot pipeline data with calendar activity. The analysis compares time allocation with deal movement and identifies where the leader or team could spend time differently.
→ The leader receives an evidence-based recommendation grounded in both selling activity and pipeline results.
Common mistakes
Assuming Automatic Source Selection
The model may not reliably infer whether it should search Drive, HubSpot, or another connected system. Explicitly route it to the relevant sources.
Connecting Data Without a Problem
Broad access without a focused question can produce a lengthy report that does not support a clear decision.
Is it for you?
Best for
Knowledge workers investigating questions that require CRM, email, calendar, or document evidence.
Not ideal for
Simple questions that can be answered reliably from one small, already-visible dataset.
From the transcript
“So I think at the moment, what I do is I say go to HubSpot and my G drive to search for this data to…”
“Yeah, or if you want to like look at a combination of data, say it, hey, combine the data from HubSpot and G Drive to…”
From the episode
AI Just Replaced a $100K Sales Analyst for $20/Month