Direct answer
Run pre-NotebookLM source preparation when the source set is scattered, duplicated, stale, oversized, poorly named, or context-heavy. The goal is to create a source-preserving export and, where useful, a clearly labelled AI research expansion before NotebookLM starts generating derivative reports.
Pre-NotebookLM preparation is not busywork. It prevents the source set from becoming noisy, unsafe or too compressed before synthesis begins.
The real problem
Most knowledge-pack projects start with messy material: chat exports, transcripts, prompts, rough notes, screenshots, research links, partial drafts, revised instructions and previous outputs.
NotebookLM can help analyse sources, but it still depends on what you give it. If the upload set contains duplicates, stale instructions, contradictory versions, rights-unsafe material or unclear file roles, the reports downstream may blend things that should stay separate.
Source preparation protects the work before the AI system starts summarising it.
When to run it
Run pre-NotebookLM preparation when there are many source files, long conversation exports, repeated prompt versions, mixed raw and AI-generated material, unclear filenames, sensitive source details, old drafts that may conflict with newer decisions or source files that exceed practical upload limits.
It is also useful when you need to preserve decisions, caveats, contradictions, source roles or version history before asking NotebookLM to generate derivative reports.
What the stage should produce
The main output is usually a source-preserving master export. This consolidates useful supplied material without flattening it into a shallow summary.
Where broader context is useful, the stage may also produce an AI deep research expansion. That expansion should be clearly labelled as supplemental context, not primary evidence.
What to clean before upload
Before NotebookLM, review the source set for duplicates, superseded versions, private notes, rights-sensitive material, unclear filenames, missing dates, source-role confusion and low-value files that would add noise.
The aim is not to delete useful nuance. The aim is to make the source set readable, traceable and bounded.
What not to do
Do not compress valuable sources into a thin summary just to make the upload easier. Do not mix original source material and AI-generated research without labels. Do not treat a generated expansion as equal authority to supplied material. Do not upload process prompts as source files unless they are part of the actual knowledge base.
If the material is too large, split it by source role or topic rather than flattening it.
Example
If a project has twenty conversation exports, five prompt versions, three draft packs and a set of research notes, Stage 01 can create a clean source-preserving export before NotebookLM is asked to produce evidence ledgers, claims tables or workflow maps.
The result is a better downstream analysis because the source base is clearer.
When this does not apply
If you have one small, clean, low-risk source file, you may not need a separate pre-NotebookLM stage. But as soon as source identity, rights, versioning, size or contradictions matter, preparation becomes useful.
Soft next step
Before uploading to NotebookLM, make a simple source inventory: what each file is, where it came from, whether it is current, whether it is original or AI-generated, and what role it should play.