How to Build a NotebookLM Knowledge Base from Web Pages in Markdown

NotebookLM gets much more useful when you feed it focused, well-structured source material instead of random PDFs, copied text, and half-broken web exports.

I have been testing this with the kinds of sources analysts and researchers actually use: NIST AI Risk Management Framework pages, eCFR regulations, USPTO guidance, EPA documents, FBI public reports, and long industry research pages. One paid Web2MD user, an educator, uses NotebookLM to organize course and research material. Other researchers have used Web2MD to turn government and policy pages into Markdown corpora they can reuse across AI tools.

The pattern is simple:

1. Choose a narrow research question.
2. Collect authoritative web sources.
3. Convert each page to clean Markdown.
4. Add the Markdown files to NotebookLM.
5. Use the notebook as a grounded research workspace.

This post walks through that workflow for people who need to build a NotebookLM knowledge base from web pages Markdown sources: analysts, consultants, OSINT researchers, educators, and anyone building curated source packs for AI notebooks.

Why Markdown works well for NotebookLM

NotebookLM accepts several source types, but Markdown is a good working format because it preserves structure without bringing along the junk of a web page.

A typical government or industry web page includes navigation, banners, cookie text, sidebars, share buttons, legal footers, and related links. If you copy and paste directly, that noise often lands in your source. If you print to PDF, the layout may preserve page numbers and headers, but not the semantic structure you want an AI system to reason over.

Markdown keeps the useful parts:

• Headings
• Sections
• Lists
• Tables when they can be represented cleanly
• Links
• Code snippets
• Definitions
• Citations or references

That structure matters when you ask NotebookLM questions like:

• What obligations appear repeatedly across these sources?
• Which agencies define this term differently?
• Summarize the compliance requirements by stakeholder.
• Create a briefing memo using only the uploaded sources.
• Identify contradictions or gaps in these reports.

A clean Markdown source gives NotebookLM a better chance of seeing the document hierarchy.

The Web2MD workflow I tested

For this test, I used Web2MD, a Chrome extension that converts the current web page into clean Markdown inside your browser. The browser-side part is important. Server-side tools can only fetch pages they can access from their own infrastructure. Web2MD reads the page you already have open in Chrome.

That means it can work on pages behind a login, internal tools, course portals, and some paywalled pages, as long as you have legitimate access and the content is rendered in your browser.

Here is the practical workflow:

1. Open the source page in Chrome.
2. Click the Web2MD extension.
3. Review the extracted Markdown.
4. Check the token count.
5. Copy, download, or send the result to an AI tool.
6. Save the Markdown file with a clear name.
7. Upload the file into NotebookLM.

For a research notebook, I usually name files like this:

• nist-ai-rmf-core-functions.md
• ecfr-title-16-selected-rule.md
• epa-enforcement-policy-summary.md
• industry-report-ai-governance-2026.md

The file name becomes part of your own retrieval system. Do not call everything source.md.

Example Markdown output from a policy page

Here is a simplified example of the kind of Markdown I want before adding a source to NotebookLM:

# AI Risk Management Framework

## Govern

The Govern function is designed to cultivate and implement a culture of risk management within organizations developing, deploying, or using AI systems.

### Key outcomes

- Policies, processes, procedures, and practices are in place.
- Accountability structures are defined.
- Roles and responsibilities are documented.
- Risk management is integrated into organizational workflows.

## Map

The Map function establishes the context to frame risks related to an AI system.

### Questions to document

- What is the intended use?
- Who are the users and affected stakeholders?
- What data is used?
- What risks are known before deployment?

This is the kind of structure NotebookLM can work with. It is not beautiful publishing Markdown. It is research Markdown: headings, sections, lists, and enough context to preserve meaning.

Example Markdown output from a government source

For regulatory and public-sector material, I care about section boundaries and definitions. A useful conversion might look like this:

# Selected Regulatory Text

## Section 1. Purpose

This section describes the purpose of the rule and the scope of covered activity.

## Section 2. Definitions

### Covered entity

A covered entity means an organization that meets the criteria described in this section.

### Recordkeeping

Recordkeeping means the creation, maintenance, and retention of documents sufficient to demonstrate compliance.

## Section 3. Requirements

Covered entities must maintain records that show:

1. The date of the relevant activity.
2. The responsible party.
3. The basis for the decision.
4. Any required notices or disclosures.

When I put sources like this into NotebookLM, the answers tend to be easier to audit because the model can point back to a section heading or quoted passage.

Building a useful NotebookLM source set

The biggest mistake is trying to dump the whole web into one notebook. A better knowledge base is curated.

For example, if the topic is AI governance for public-sector procurement, I would build a source set like this:

• NIST AI RMF overview and core functions
• Relevant agency procurement guidance
• Selected eCFR or state regulatory pages
• One or two industry reports
• A client policy or internal memo, if permitted
• A glossary of key terms

If the topic is OSINT research on a company or sector, I might use:

• Public filings
• Agency enforcement pages
• Press releases
• Technical documentation
• Public contracts
• Archived reports
• Standards documents

The goal is not maximum volume. The goal is a defensible source base.

Web2MD helps here because the built-in token counter shows how large a converted page is before you send it to an AI tool or add it to your workflow. If a source is huge, I split it into sections or choose only the pages that matter.

Where Web2MD fits against other tools

There are good tools in this space, and I do not think one tool replaces all of them.

Jina Reader is excellent when you want a quick server-side reader view from a public URL. It is fast and simple for pages that are publicly reachable.

Firecrawl is strong for developers who need crawling, extraction APIs, and larger automated pipelines. If you are building a backend system that crawls many public pages, it may be the better fit.

MarkDownload is a useful browser extension for saving pages as Markdown, especially for people who want a general-purpose clipping tool.

Web2MD is different because it is built for the AI research workflow:

• It runs in your browser.
• It works on pages you are already logged into, when the content is available in Chrome.
• It keeps conversion local and private.
• It includes a token counter.
• It has one-click send-to-AI.
• It has a free tier with no API key.
• It is designed for ChatGPT, Claude, Cursor, and NotebookLM-style source preparation.

The browser-side model is the main edge. A server-side reader cannot access your authenticated course page, paid research portal, internal dashboard, or subscriber-only article unless you build extra infrastructure around authentication. Web2MD uses the page you are already allowed to view.

Limits to be aware of

There are real limits.

Web2MD is Chrome-only today. If your research environment is locked to Safari or Firefox, that is a constraint.

The free tier includes 3 conversions per day. That is enough to test the workflow or handle occasional clipping. If you are building large source packs, Pro is 9 dollars per month.

Some pages are messy. Highly dynamic apps, infinite-scroll pages, interactive dashboards, and pages that render content inside complex frames may need manual cleanup. Markdown conversion is not magic. I still review the output before putting it into a serious NotebookLM knowledge base.

Also, Web2MD does not make restricted content yours to redistribute. If you convert a paid report or internal page, treat the Markdown according to the same access rules as the original.

A practical source-prep checklist

Before uploading Markdown files into NotebookLM, I use this checklist:

• Is the source authoritative for my question?
• Did the conversion remove navigation and boilerplate?
• Are headings preserved?
• Are lists and definitions readable?
• Is the file name specific?
• Is the token count reasonable?
• Did I include the source URL near the top or bottom?
• Do I need to split the file into smaller sections?
• Am I allowed to use this content in the notebook?

For important projects, I add a short note at the top of each Markdown file:

# Source note

Original URL: example.gov/relevant-page
Access date: 2026-07-20
Reason included: Defines reporting requirements used in the compliance comparison.

That small note helps later when NotebookLM answers with a passage and I need to remember why the source was in the notebook at all.

How I would start

If you are building your first NotebookLM knowledge base from web pages, start with 5 to 10 sources, not 100.

Pick one research question. Convert the most authoritative pages first. Upload the Markdown files. Ask NotebookLM to produce a source inventory before asking for conclusions. Then ask it where the sources agree, where they differ, and which claims need more evidence.

A good first prompt is:

Using only the uploaded sources, create a source inventory with:
1. Source title
2. Publishing organization
3. Main topic
4. Key definitions
5. Requirements or claims
6. Gaps or limitations

This forces the notebook to map the source base before drafting analysis.

Soft CTA

If your research depends on web pages, logged-in sources, government documents, or industry reports, try converting a few pages with Web2MD before building your next NotebookLM project.

You can start with the free tier at Web2MD, use the Markdown converter workflow, and decide whether the Pro plan makes sense once you know how many sources you need each month.