NOTEBOOKLM
An ergonomic evaluation of Google's NotebookLM covering usability, cognitive load and stress, followed by a redesign of its sources, notes and audio/video overviews that was tested against the original.
UI/UX
Product Design
Role
Designer
Timeline
4 Week
team
3 Designers
platform
Web
What this is for
NotebookLM is Google’s AI research assistant for your own documents. You add papers, links or notes, and it uses the Gemini 1.5 Pro model to produce summaries, Q&As, reports, podcast-style audio overviews and video overviews based on that material. For students and researchers working through piles of papers, that only helps if they can find the tools, trust the output and stay in control of it.
This project asked how well NotebookLM actually works for those people, and how much mental effort it demands. The aim had four steps: evaluate the usability, understand its cognitive load, identify the key usability challenges, and propose targeted redesigns. Participants included BTech and Masters students, PhD researchers and NotebookLM Discord users.
Process: Analyze → Define → Ideate → Execute
Analyze. Guerrilla testing and a think-aloud pilot study chose a few key tasks that covered NotebookLM’s main features, how often they’re used, and how relevant they are. A baseline came from screen-recorded sessions (task success and efficiency), the System Usability Scale, NASA-TLX for perceived workload, and SASSI for the interactive, speech-based audio mode.
Define. The team mapped the information architecture of the updated UI. They broke five realistic tasks (video, audio, report, literature review, interactive mode) into steps with Hierarchical Task Analysis, mapped errors with SHERPA, ran a heuristic evaluation against Nielsen’s 10 heuristics, and used CREAM to imagine where cognitive errors could cause real harm.
Ideate and execute. All of these insights fed a redesign of the source panel, notes, Audio and Video Overviews, reports and the output library. The old and new UIs were then compared with task success, SUS, galvanic skin response (GSR), eye tracking and NASA-TLX.
A clearer layout, and sources you can manage
The redesign starts with a clear three-panel model: Sources (input) on the left, Chat (process) in the centre and Studio (output) on the right. Users struggled to find where to save a note, so the input bar becomes the place to take notes, and Studio outputs stay visible while you write.
The source panel gets the tools a research tool needs. “Discover” is renamed to Discover source and gains a search bar, filter tags by source type, a pin source option with better contrast, adding from several places in one window, and direct upload from Google Drive. You can customise the discover prompt without losing the sources you’ve already selected. Saved outputs go into an Output Vault with tags, search, pinned notes, and a record of each note’s source and upload time.
Control over what the AI generates
Audio Overview becomes something you can steer. You can add an optional customisation prompt and change sources at any time. Playback adds subtitles, a visual cue for which AI host is talking, topic markers and bookmarks, and a scrubber for full control. You can also join the conversation: subtitles stop while you speak, the hosts answer your question, then carry on from where they paused.
Video Overview gains chapters and bookmarks, and hovering shows where you are. Reports get a Prompt Library with templates you fill in, tone-based suggestions, suggestions drawn from your selected sources, and prompts shared by people the notebook is shared with. Studio features move into the chat window.
The challenge
Error mapping with SHERPA, a method for finding where users slip, make mistakes or give up, surfaced six key problems:
No onboarding at all: users were left guessing how the product works.
Core tools are hard to find: features like Study Guide are hidden, and users don’t realise they exist.
Missing system feedback: there’s no confirmation for long tasks like video generation, so users wait without knowing what’s happening.
No control over generated content: you can’t scrub through long audio or set a summary’s length before it’s generated.
No source management: sources can’t be sorted or filtered, which is a dealbreaker for a research tool.
Confusing layout: users couldn’t form a mental model of the three-panel layout.
Midway through the evaluation, Google shipped a surprise UI update. It fixed most of the pain points from phase 1 and added Video Overview, so the team had to map and evaluate the new interface again.
Results, old UI vs redesign
In the team’s old-vs-new study, task success rose from 93.75% to 98.25%. Average galvanic skin response during tasks, a physical sign of stress, fell 27%, and peak skin-response spikes per task fell 43%. On NASA-TLX, frustration dropped from 58.75 to 55. Not everything improved: mental demand rose slightly, from 40 to 42.
The redesign is a concept. These numbers come from the team’s own comparison of the old and new UIs, not from NotebookLM in production.
Team: Manan Soni, Fuaad Muhammed and Saptarshi.
