Question-Depth Routing Rule
Route factual lookups to fast models and complex investigations to deep research
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 98%
This rule routes questions according to the depth of investigation needed. A direct, current fact that can be answered from a straightforward lookup belongs with a fast model because additional research adds latency without meaningful value. A question involving causes, construction methods, participants, regulations, conflicting evidence, or other interconnected details belongs with deep research, which can search and synthesize across many sources. The user first states the decision or knowledge need, then estimates how many independent facts and relationships must be established. Consequential results still require citation review. The framework optimizes speed for simple work while preserving depth for genuinely complex investigations.
Origin
Extracted from Marketing Against The Grain
Core principles
- 01Match tool cost and latency to the depth of the question
- 02Use fast models for direct factual answers
- 03Use deep research when the answer requires synthesis across sources
- 04Judge complexity by investigation depth, not by topic prestige
How to run it
- 1
Frame the question
Write the request as specifically as possible, including the decision the answer will support.
Pro tip Separate a factual lookup from the deeper explanation it may inspire.
Watch out A vague question makes routing unreliable.
- 2
Test for directness
Ask whether a single authoritative fact or simple response can adequately resolve the request.
Pro tip Favor the fast model when the answer is likely immediate and uncontroversial.
Watch out Do not mistake a short question for a simple investigation.
- 3
Assess investigation depth
Look for multiple actors, causes, regulations, documents, historical context, or conflicting sources.
Pro tip If several linked follow-up questions are inherent, use deep research.
Watch out Deep research cannot compensate for inaccessible or nonexistent evidence.
- 4
Route the task
Send direct questions to the fast model and substantive investigations to deep research.
Pro tip Escalate if the fast answer reveals unexpected complexity.
Watch out Do not use deep research merely because a topic feels important.
- 5
Verify the output
Inspect sources and assumptions whenever the result informs a consequential action.
Pro tip Open the most decision-critical citations rather than sampling randomly.
Watch out Tool selection does not guarantee correctness.
In the wild
A user asks who won yesterday's Cubs game and receives an immediate answer from a fast model. A second question asks why a stadium structure was built, how it was constructed, who worked on it, and what permits were required, so the user routes that investigation to deep research.
→ Each question receives the appropriate balance of speed and evidentiary depth.
Common mistakes
Using deep research for every lookup
The added latency and output volume provide little benefit for a direct factual question.
Using fast answers for layered investigations
A quick model may provide a plausible surface response without establishing the linked evidence the question requires.
Routing by word count
A short prompt can conceal a complex regulatory or causal investigation, while a long prompt may request one simple transformation.
Is it for you?
Best for
People choosing between a fast conversational model and an autonomous research mode.
Not ideal for
Tasks requiring private data, physical inspection, or professional judgment unavailable to either tool.
From the transcript
“your everyday questions, you know, if it's a simple question, just use 2.0 flash.”
“If you really do need something that is not surface level”
“that level of depth in the question that you have, there's no product that can do that besides deep research.”
From the episode
Google's Secret AI Advantage (Why DeepMind Will Dominate)