Ishita Sarkar | Portfolio

Invisible Interactions

Recommended for you

Evaluating Recommendation Systems Across YouTube, Instagram, and Spotify

Duration: 1 week • Solo Project

4:38

This study focuses on how UI design affects users’ trust in AI-curated playlists and content.

How do different platforms visualize or hide algorithmic logic?

Recommendation systems are shaping what people watch, listen to, and engage with every day, subtly influencing taste, opinions, and even moods. Yet, users often don’t understand why they’re seeing certain content or how to control it.

Complex recommendation logic tied to watch time, engagement, and satisfaction. Great for studying “explainability” and control granularity.

Video/Visual media

Youtube

Audio Streaming

Spotify

Visual Social Media

Instagram

How clearly does the interface communicates why something is recommended?

Are personalization settings easy to find and understand?

How easy is it for users to adjust, reset, or refuse recommendations?

Do users feel their actions (like/dislike) are reflected in future recommendations?

Transparency

Control

Trust & feedback loops

Consistency & visibility

Dimensions for Evaluating Algorithmic Trust

Types of

Recommendations

Type

Home screen

Video sidebar/next play

Shorts tab, home

Home, watch page

Personalized mix of interests

While watching a video

Scroll behaviour

Tap on recommend topic

Watch history, Subscriptions, Trends

Current video, topic similarity, engagement

Shorts engagement, past likes

Watched content themes

Placement

Data used

Trigger

YouTube

Home feed

Up next / Autoplay

Shorts Recommendations

Topic-based suggestions

Type

Placement

Data used

Trigger

Mix of playlists, artists, and albums

Personalized algorithmic playlists

Query and genre trends

Current track

Social graph

Personalized mix of interests

While watching a video

Scroll behaviour

Tap on recommend topic

Listening patterns

Dynamic, scrollable carousels

Auto-play of related tracks after ending

Displays what friends are playing

Home feed

Made For You (Daily Mixes, Discover Weekly)

Search Tab Suggestions

Now Playing Queue

Friend Activity Feed (Desktop)

Spotify

Type

Home feed between followed posts

Reels tab

Explore tab

Shopping tab / in-feed ads

Feed scroll area

Reels tab / Explore page

Explore page

Shopping tab / ads section

Engagement patterns, comments, view duration

Watch time, audio reused, content similarity

Topic & content attributes, accounts you don’t follow

Behavioural data, search history, purchase intent

Placement

Data used

Trigger

Instagram

Suggested Posts (in-feed)

Suggested Reels

Explore page / “For You” content

Shopping / Ads recommendations

Heuristic Evaluation

Through this evaluation we are trying to see if the algorithmic cues explained in user-friendly language.

Can users influence or reset their recommendations and if the interface teach users how recommendations work over time.

Through this evaluation we are trying to see if the algorithmic cues explained in user-friendly language.

Can users influence or reset their recommendations and if the interface teach users how recommendations work over time.

See detailed evaluation (updating soon..)

Spotify

Instagram

YouTube

1. Visibility of System Status

2. Match Between System & Real World

3. User Control & Freedom

4. Consistency & Standards

5. Error Prevention

6. Recognition Rather Than Recall

7. Flexibility & Efficiency of Use

8. Aesthetic & Minimalist Design

9. Help Users Recognize, Diagnose & Recover

10. Help & Documentation

Poor

Heuristic

Medium

Good

Pattern in recommendations

Why was this video recommended?

Pattern Type

Indirect / Deeply Nested Information Disclosure

No direct information for a particular video. Longer flows of clicks, long scrolls to reach desired info, finding it needs effort.

View

YouTube

Spotify

Pattern Type

Transparency exists but is hidden under secondary menus or vague headers. It follow a consistent naming pattern for personalized content.

View

Implicit feedback system. Interaction is passive, indirectly training the algorithm.

Instagram

Pattern Type

Just-in-time transparency approach, easily visible

Present in the post itself for each and every post. one click required to reach the option.

View

Agency & feedback

Can users react to what they are shown?

Can users customize/shape what they see?

Can users understand and control how the system uses their data?

Reactive Control

Proactive Control

Not Interested

Remove from Watch history → Refresh

Topic Tab

Not Interested

Remove from Watch history → Refresh

Topic Tab

Meta Control

Feedback

YouTube

The “Remove” button

The are placed in the kebab menu at the side of every video. The options have icons to represent their meaning.

Deleting from watch history and refreshing removes videos related to that topic from appearing again.

Users can select what gets stored as their YouTube history and regulates their feeds and recommendations.

On clicking Not Interested there is an option to undo as well as to give further feedback.

On deleting a video from watch History:

Settings → History → Manage All History → Controls → YouTube History

Tab at the top of home feed lets you modify your feed according to the topics

There are occasional customizations of feed available through these kinds of exercises.

Depends totally on the user to do them or not, can be easily missed.

There are occasional customizations of feed available through these kinds of exercises.

Depends totally on the user to do them or not, can be easily missed.

Can users react to what they are shown?

Can users customize/shape what they see?

Can users understand and control how the system uses their data?

Reactive Control

Proactive Control

Meta Control

Feedback

Spotify

Option to un-include a song from your preferences

Song → Kebab menu → Exclude

If it is a public playlist or a song not in Liked/Playlist, the option to hide it from the playlist or queue is available

Songs can be added to the queue that is playing

Once muted or hidden the songs by the artist becomes unclickable and faded.

After hiding/muting, a pop-up appears to confirm that the sog won’t be playing anymore.

Muting the artist will completely remove them from playing

Undo option is available

Can users react to what they are shown?

Can users customize/shape what they see?

Can users understand and control how the system uses their data?

Reactive Control

Proactive Control

Meta Control

Feedback

Instagram

The kebab menu in the reel/post lets user immediately remove or hide content that they don’t like.

If it is a reel/post from a new account that they haven’t interacted with, they can click “interested” to make it more visible.

The kebab menu in the reel/post lets user immediately remove or hide content that they don’t like.

If it is a reel/post from a new account that they haven’t interacted with, they can click “interested” to make it more visible.

Clicking “not interested” open up options like muting the user, muting words and reporting a post as being “uncomfortable”.

Preferences can be set according to levels of comfortability

Interested and Not interested posts data is stored and saved for user to access.

Interested and Not interested can be undone or deleted.

Feedback informs the user of the action the app will be taking as well as till wha duration.

SUS: System Usability Score

Average Usability Score

Task

Task Description

Evaluation Dimension

Description

Youtube

Spotify

Instagram

Find out why a particular video/song/post was recommended to you.

Transparency

How clearly the platform explains why something was recommended.

Try to remove or dismiss a recommended video/song/post you don’t like.

User Control

How easily users can influence or change recommendations.

Try to adjust or influence future recommendations (e.g., clear watch history, dislike, “not interested,” etc.).

Feedbck

Whether the interface provides acknowledgment or visible response to user input.

Observe how the interface responds after feedback (e.g., confirmation, visual change).

Trust

The degree to which users find the recommendations credible or appropriate.

54.3

59.3

Perceived Usability

Find out why a particular video/song/post was recommended to you.

UX Honeycomb

Evaluation of YouTube, Spotify, and Instagram’s recommendation systems across five UX parameters, highlighting differences in usability and overall user satisfaction.

YouTube

Spotify

Instagram

Usable

Findable

Desirable

Credible

Useful

Documentation in progress….