# How would you improve YouTube’s recommendation algorithm?

Aced (formerly Exponent) publishes a public search API over its indexed pages; the guide is at https://www.aced.io/agents/agent-access.

Source: https://www.aced.io/questions/5452/improve-youtube-recommendation-algorithm
Source observed: 2026-11-04T21:22:58.492Z
Last checked: 2026-11-04T21:22:58.492+00:00
Indexed: 2026-11-05T00:43:00.425+00:00

---

# How would you improve YouTube’s recommendation algorithm?

Aced (formerly Exponent) hosts this product-design and product-strategy interview question, tagged **Product Design**, **Product Strategy**, **Product Manager**, **Finance and Strategy**, **Google**, **ADP**, and **Meta**. The question has **5 interviews**, was **last asked 8 months ago** in the rendered details, and has **45 community answers**. You can browse [all questions](https://www.aced.io/questions), [practice with AI](https://www.aced.io/practice/ai/new?type=PM&question=improve-youtube-recommendation-algorithm&src=questions), or [submit your own answer](https://www.aced.io/signup?src=nav&returnPath=/questions/5452/improve-youtube-recommendation-algorithm).

## Question scope

The prompt asks how to improve YouTube’s recommendation algorithm. Community responses consider different recommendation surfaces, user goals, audience segments, discovery, relevance, variety, user control, and measurable outcomes.

## Community answers

### Vraj C. — October 27, 2026

Vraj first clarifies the recommendation surface and goal.

**Recommendation surfaces:**

- Long-form content:
  - Home page
  - Next-video recommendation
  - Content page
- Short-form content:
  - Recommending the next short video

**Possible goals:**

- Increase time spent watching per session.
- Increase the percentage of users who watch a recommended video.
- Increase advertising revenue per session.

**User segments:**

- New users
- Casual users, typically watching 1–2 videos every day
- Power users, watching 2–4 videos per day

**Behavioral segments:**

- Learners seeking to learn about a topic
- Entertainment viewers seeking a specific type of content
- Live-video viewers
- General viewers who want to pass time

Vraj prioritizes long-form content on the hypothesis that users still come to YouTube for long-form content, then prioritizes the next-video recommendation because the user has already started a session with the first video. The proposed initial audience is casual learners.

**Pain points:**

1. The next video is not interesting.
2. The thumbnail looks like the same type of content as the previous video.
3. The next video does not provide incremental learning from the previous video.

**Proposed approach:**

- Rate the top three recommendations.
- For educational content, use topic, subtopic, difficulty level, and whether the previous video belongs to a series as signals.
- Prioritize the next video in an existing series.
- If there is no series continuation, use the previous video’s difficulty and recommend another video on the same subtopic.
- Recommend videos watched by users who watched the next video and completed 75% of it.

**Metrics:**

- Percentage of the recommended video watched.
- Percentage breakdown for the first, second, and third recommendations.
- Percentage of users who finish at least 75% of the recommended video, alongside recommendation rating and number of videos watched.
- Videos watched per session.
- Number of sessions should not decline.

### Malak F. — September 11, 2026

Malak proposes making recommendations more personal without repeatedly showing the same type of video. Relevant signals include what users watch, skip, and search for, and how long they watch. Malak would also give newer or different videos some weight so smaller creators have an opportunity to reach viewers.

The proposed testing approach measures whether users actually enjoy recommended videos rather than measuring clicks alone. The goal is to help users find videos they genuinely want while preserving variety.

### Kartikeya N. — November 30, 2025

Kartikeya frames the answer around **YouTube web**, **long-form videos**, a **six-month timeframe**, and existing recommendations on the Home page and alongside a playing video. The stated target is engagement with recommended videos, specifically improving the percentage of times a recommended video is watched. No other constraints are specified.

Kartikeya connects recommended-video engagement with confidence in the algorithm, the engagement flywheel, and advertising revenue from increased watch time.

**Target user segments:**

- Students
- Working professionals
- Homemakers
- Kids

The answer prioritizes students and working professionals, especially for educational content, and identifies these pain points:

- Difficulty discovering new videos or topics without searching or receiving a shared link — high acuteness and high frequency, **P0**.
- Recommended videos may be poor, causing users to waste time, particularly with long-form videos — high acuteness and high frequency, **P0**.
- Some creator content is paywalled and available only to channel subscribers — medium acuteness and low frequency, **P2**.
- A cluttered Home page with many types of videos makes it hard to know where to begin — high acuteness and medium frequency, **P1**.

**Proposed solutions:**

1. **Discover mode — P0:** Add a Home-page mode that increases randomness while retaining relevance, helping users discover educational content and new creators rather than only trending or highly engaged content. The answer rates this as high impact and medium effort.
2. **Improved semantic and contextual signals — P1:** Use generative-AI capabilities to analyze the semantics and context of the entire video’s audio and video content, rather than relying mainly on metadata. The answer rates this as high impact and high effort and suggests that Gemini capabilities could strengthen the approach.
3. **Video and comment summaries — P1:** Provide a generative-AI summary alongside the video description and summarize useful discussion in the comments while removing noise. The answer suggests piloting this in one geography before broader rollout and using the summaries as recommendation signals. It also identifies a possible downside: users might read summaries instead of watching videos. The answer mentions possible monetization through ads between summaries.

**Success metrics:**

- North-star metric: percentage of times a recommended video was watched.
- Countermetrics: creator satisfaction scores and cross-category watch time.
- Ancillary metrics: watch time, DAU/WAU stickiness, ARPU, and the percentage of users subscribing to YouTube Premium.

### Aekagra S. — September 8, 2025

Aekagra identifies three underlying problems:

- Recommendations are not specific enough, frustrating users and causing them to leave YouTube.
- Recommendations can overemphasize one set of videos, producing content fatigue and failing to include other user interests.
- Users feel they have insufficient control over their feed.

**Goals:**

- **Engagement:** Increase watch time per user, session length per user, and videos consumed per session.
- **Retention:** Reduce fatigue-related churn and increase repeat visits.
- **Trust:** Give users transparency and control to reduce frustration with the algorithm.

The proposed strategy moves from a reactive model optimized primarily for watch time toward a proactive, user-centric system that prioritizes satisfaction, variety, and user agency.

**Context-aware recommendations:**

- Time of watch: distinguish a utility request, such as fixing a leaky faucet at 2 AM, from entertainment discovery.
- Playback speed and rewatches: use slowed playback or repeated segments of a tutorial as signals of specific interest.
- Video chapters: use chapter jumps to identify the viewer’s primary interest.
- Comments and engagement: use comments and shared portions as micro-signals of value.
- Video-attribute tagging: match users using detailed attributes such as “Japanese cuisine, street food, vlogs, under 10 minutes,” based on creator tagging with verification or machine-learning analysis of video content.

**Variety and fatigue reduction:**

- Add a **Diversity Score** to the recommendation objective.
- Introduce a **Variety** shelf with:
  - **New for You:** Adjacent-interest discovery from topics, creators, or genres the user has not engaged with before.
  - **From Your Subscriptions:** Missed content from subscribed creators.
  - **Trending Near You:** Locally relevant picks.
- Add decay and refresh logic so videos that users have seen and ignored lose weight, allowing newer and more diverse related content to appear.

**User controls:**

- **Manage My Interests:** A settings dashboard where users can view and adjust the topics identified for them.
- **Temporary Mute:** Temporarily mute a topic, such as Olympic recommendations for two weeks.
- **Why am I seeing this?:** Explain the signals behind each recommendation. The answer’s example explains that a user might see a *Game of Thrones* fan-theory video after watching *The Lord of the Rings* videos and because similar viewers liked it.

**Prioritization:**

1. **Phase 1 — Specificity:** Improve signals and add decay logic as quick engagement improvements.
2. **Phase 2 — Diversity shelves or curated sections:** Balance novelty with familiar binge viewing.
3. **Phase 3 — User controls:** Increase trust, with the expectation that adoption may be lower because mainly power users configure them.

### Priya A. — July 25, 2026

Priya begins by clarifying the objective:

- Target retention rather than engagement or time spent.
- Start with existing customers rather than trying to bring back new customers.
- Include both long-form and short-form videos.
- Start with the United States rather than a global audience.

**User-intent segments:**

- Routine viewers who want familiar content.
- Explorers who want to discover something new.
- Context-driven viewers whose intent changes by time or situation.
- Shared-account users whose preferences differ within one account.

**Problems and solutions:**

| Problem | Proposed solution |
|---|---|
| Repetition: users repeatedly see the same creators and topics | Controlled exploration recommendations |
| Interest drift: users’ interests change over time | Periodic explicit preference refresh |
| Context blindness: the same recommendation strategy is used regardless of time or context | Context-aware recommendations based on individual user patterns |
| Shared accounts: one account may represent multiple users | Profiles or automatic preference segmentation |

**Prioritization:**

- **P0 — Context-aware recommendations:** High potential impact and low user friction. Use existing behavioral data to understand how each individual’s preferences change by time and context.
- **P0 — Periodic interest refresh:** Directly address interest drift by occasionally asking users to confirm or update their interests.
- **P1 — Controlled exploration:** Potentially improve discovery and long-term engagement, but with a higher risk of irrelevant recommendations. Test with a small percentage of the feed first.
- **P1/P2 — Shared profiles:** Potentially valuable for households, but requiring more user effort and product complexity.
- **Experiment — Shorts-to-long-form signals:** Test whether Shorts engagement predicts interest in long-form content before creating a major recommendation dependency.

## Related questions

- [How would you improve Airbnb's recommendation algorithm for guests?](https://www.aced.io/questions/1288/improve-airbnb-recommendation-algorithm)
- [How would you improve YouTube?](https://www.aced.io/questions/666/how-to-improve-youtube)
- [How would you improve Rover?](https://www.aced.io/questions/2607/improve-rover-strategies)

## Related courses

### [Product Management Interview Prep](https://www.aced.io/courses/pm)

![Product Management Interview Prep](https://images.ctfassets.net/x78yjrjc11pq/4PFBEhtIRyKwPm14eaBAFm/ad92a8c4dd1dc51108b3a855de64473b/product_management_track_compressed_063e6d4b35.jpeg?w=800&h=800&fm=webp&q=75)

Prepare for product-management interviews at Anthropic, OpenAI, Meta, Apple, Google, and other companies. The course covers AI product-sense rounds, vibe-coding prototypes, metrics questions, AI-fluency interviews, product strategy, product sense, analytical and execution interviews, leadership and drive, and technical product management interviews.

### [Product Skills Course](https://www.aced.io/courses/pm-skills)

![Product Skills Course](https://images.ctfassets.net/x78yjrjc11pq/6bSjuhpNfU7emXQdz776t5/389a7c3fc4c4006965c8471b06554b8a/user-insight_a3d886ea91__281_29_4333187956.jpeg?w=800&h=800&fm=webp&q=75)

Learn the fundamentals of product management and the basic skills a product manager needs.

## Access and actions

- [Upgrade](https://www.aced.io/upgrade?src=questions) to unlock detailed company statistics for this question.
- [Practice with AI](https://www.aced.io/practice/ai/new?type=PM&question=improve-youtube-recommendation-algorithm&src=questions).
- Submit an answer through the question’s answer editor; submitting marks the question complete.

---

[Read this page’s summary](https://www.aced.io/agents/api/public/v1/page-tldr?url=https%3A%2F%2Fwww.aced.io%2Fquestions%2F5452%2Fimprove-youtube-recommendation-algorithm&trace=t_679ff9ab98f3)

## Ask a follow-up about this page

Need a more specific answer? Ask the exact question you have. It returns a fresh, cited answer from Aced (formerly Exponent)’s knowledge base. No account or API key required.

Open this URL with your own question, URL-encoded. Include your question in q. All research inputs are optional; no account or setup.

`GET https://www.aced.io/agents/agent/faq?q={question}&trace=t_679ff9ab98f3`

[Tool guide and examples](https://www.aced.io/agents/agent-access) · [OpenAPI](https://www.aced.io/agents/api/public/v1/openapi.json) · [Page TLDR form](https://www.aced.io/agents/agent-tools/page-tldr?url=https%3A%2F%2Fwww.aced.io%2Fquestions%2F5452%2Fimprove-youtube-recommendation-algorithm). Opening a form does not execute a tool.