product improvement cases

“How would you improve X?” is the most common product case study question. It is also the one candidates botch most consistently — because they skip straight to features.

The interviewer does not want a feature list. They want to see whether you can take an ambiguous prompt, narrow it to a specific problem, identify who you are solving for, pick the right metric, and then — only then — propose a solution that has trade-offs you have actually thought through.

Here are three fully worked cases. Each follows the same structure: clarify the scope, segment users, pick a metric, propose a solution, and name the trade-offs. Study the structure more than the specific answers.

The framework in 60 seconds

Before the cases, here is the skeleton you should internalize:

Most candidates spend 80% of their time on step 6 and skip steps 1-3 entirely. That is backwards. The interviewer has already seen a hundred feature specs. They have not seen enough people who can frame a problem.

Case 1: Improve Netflix recommendations

User segments

Netflix users are not a monolith. For recommendations, there are at least three meaningful segments:

1. Browsing-heavy users — spend 10-20 minutes scrolling before selecting anything. High browse-to-watch ratio. Often leave without watching.
2. Habitual watchers — know what they want (continuing a series, rewatching a comfort show). Recommendations barely matter to them.
3. Shared-account users — a family of four on one profile, or a couple who watches together. The algorithm sees contradictory signals and produces confused recommendations.

Which segment to pick: Segment 3 — shared-account users. Why? Because Netflix already spent $1 million on the Netflix Prize to improve collaborative filtering for individuals. The algorithm is mature for single-user signals. But shared profiles produce noisy data that degrades recommendation quality for everyone on that account. And Netflix has publicly struggled with this — the profile system exists, but adoption is low in many households, especially in markets like India where families share a single TV.

The pain point

When a parent watches a kids’ movie with their child on Saturday, the homepage fills with animated content on Monday. The algorithm cannot distinguish between “I chose this for myself” and “I chose this for the room.” The recommendation quality drops, the user scrolls more, and eventually they stop trusting the suggestions.

Metric

Recommendation click-through rate (CTR) for shared profiles — specifically, the percentage of recommended titles that are clicked and watched for more than 5 minutes. This filters out accidental clicks and mismatches.

Secondary metric: browse time before play. If recommendations improve, this number goes down.

Solution: Contextual viewing modes

Instead of relying solely on profiles (which many users ignore), introduce a lightweight “Who’s watching?” prompt that infers context from signals:

• Time of day — Kids’ content at 4 PM vs. thriller at 11 PM.
• Device — TV in living room vs. phone in bed.
• Viewing pattern — If someone watches 3 episodes of a Korean drama, then switches to Peppa Pig, the system flags a context switch instead of blending the signals.

The system builds micro-profiles within a single account — without requiring the user to manually switch profiles. Each micro-profile feeds separate recommendation models.

MVP scope: Start with time-of-day signals only. Split recommendations into “daytime” and “evening” preference clusters for shared profiles. Run an A/B test against standard recommendations.

Trade-offs

The key insight the interviewer wants: You did not say “add a better algorithm.” You identified a specific failure mode (mixed signals from shared accounts), proposed a specific mechanism (contextual micro-profiles), and named a specific risk (inference errors). That is product thinking applied to an improvement question.

Case 2: Improve LinkedIn feed

Clarifying the scope

“Improve LinkedIn” is impossibly broad. LinkedIn is a jobs platform, a professional network, a content distribution system, a learning platform, and an advertising business. You must pick one surface.

For this case: the LinkedIn feed — the main content experience that users see when they open the app.

User segments

1. Job seekers — open the app to check job alerts, but scroll the feed out of habit. Feed content is rarely relevant to their search.
2. Content creators — post regularly, want distribution and engagement. Their problem is reach, not consumption.
3. Passive professionals — the largest segment. Open LinkedIn 2-3 times a week. Scroll for a few minutes. Rarely post. Looking for something useful but often leave feeling like they wasted time.

Which segment to pick: Segment 3 — passive professionals. They are the majority of LinkedIn’s DAU but have the weakest engagement loop. If you improve the feed for them, you increase time-on-platform and ad impressions, which is LinkedIn’s core revenue driver.

The pain point

The LinkedIn feed mixes three types of content with no separation:

1. Professional content — industry insights, career advice, skill-building
2. Social content — congratulations posts, work anniversaries, “I’m humbled to announce”
3. Engagement bait — polls asking “Do you agree?”, rage-bait about hustle culture, reposted motivational quotes

Passive professionals want type 1. They tolerate type 2. They actively dislike type 3 — but type 3 gets the most engagement signals (comments, reactions), so the algorithm promotes it.

The result: the feed feels noisy. The signal-to-noise ratio is low. Passive professionals reduce their visit frequency because the content does not feel worth their time.

Metric

Feed quality score — measured by a post-session survey (sampling 1% of sessions): “Did you find something useful in your feed today?” Binary yes/no.

Supporting metric: session frequency for passive professionals (currently 2-3x/week — target 4-5x/week).

Why not engagement time? Because you can increase engagement time by showing more outrage bait. The metric must align with the outcome you want, not just the thing that is easiest to measure.

Solution: Intent-based feed modes

Give passive professionals a toggle at the top of the feed:

• “Catch up” (default) — the standard algorithmic feed, but with engagement-bait signals down-weighted.
• “Learn” — filters to long-form posts, articles, and shared links from people in your industry. No polls, no congratulatory posts, no reshares without commentary.

The key design choice: “Learn” mode is not a separate tab (tabs get ignored). It is a toggle that transforms the existing feed. One tap. The user stays in the same scroll experience but the content quality goes up.

MVP scope: Launch “Learn” mode to 5% of passive professionals (identified by posting frequency < 1/month and session frequency 2-4x/week). Measure feed quality score and session frequency against a control group.

Trade-offs

The key insight: You did not propose a feature that LinkedIn already tried and killed (Stories). You identified a specific tension (engagement algorithm promotes low-quality content) and proposed a mechanism that aligns the business model (more sessions = more ad revenue) with user value (useful content). That alignment is what makes a PM answer compelling.

Case 3: Improve WhatsApp groups

Clarifying the scope

WhatsApp is messaging, voice calls, video calls, status updates, payments (in India), and business messaging. For this case: WhatsApp groups — the most used and most complained-about feature in markets like India, Brazil, and Nigeria.

User segments

1. Group administrators — created the group, manage membership, set rules. Their pain is moderation.
2. Active participants — post regularly, reply to threads. Their pain is information overload when they are in 10+ groups.
3. Silent members — in the group but rarely post. Open the group, scroll past 200 unread messages, mute it again. Their pain is noise. They stay because leaving feels socially awkward.

Which segment to pick: Segment 3 — silent members. They are the majority in most groups. Their experience is so poor that many mute all groups permanently, which means they miss the 5% of messages that actually matter to them. If you can surface the signal from the noise, you retain users who are one step away from disengaging entirely.

The pain point

A family group in India generates 50-100 messages per day. Of those, maybe 3-5 are important (a wedding date, a medical update, a school notice). The rest are good-morning images, forwarded memes, political forwards, and prayer messages. There is no way to find the important messages without scrolling through everything.

WhatsApp’s design treats every message as equal. There is no hierarchy, no summary, no way to mark something as important except pinning (which only admins can do, and only 3 messages at a time).

Metric

Group unmute rate — the percentage of muted groups that users unmute after the improvement. If silent members start unmuting groups, it means the signal-to-noise problem is being addressed.

Supporting metric: messages read per group visit for muted groups (currently near zero for silent members).

Solution: Smart group digest

For muted groups, generate a daily digest that surfaces:

• Messages with high reply counts (the group thought it was important)
• Messages from admins (usually announcements)
• Messages that mention you by name
• Shared documents, locations, and events (structured content, not chat noise)

The digest appears as a single notification: “3 things you might have missed in Family Group.” Tap to see a condensed view. No AI-generated summary — just the actual messages, filtered and ranked.

MVP scope: Start with admin messages + high-reply messages only. Two signals, simple ranking. Test in India (highest group density) with groups that have been muted for 30+ days.

Trade-offs

The key insight: You did not say “add threaded replies” (Slack already did that and most consumer users ignore threads). You identified a specific user (silent members), a specific behavior (muting), and a specific mechanism (on-device digest) that works within WhatsApp’s privacy constraints. The constraint awareness is what elevates the answer.

Common mistakes in improvement cases

I have evaluated thousands of product improvement cases across Pragmatic Leaders cohorts. The same mistakes appear again and again.

1. Jumping to features without qualifying “improve.”

“Improve” means nothing until you specify: improve what? For whom? Measured how? The first two minutes of your answer should be entirely about narrowing scope. If you start with “I would add a dark mode,” you have already lost.

2. Picking the largest user segment instead of the most interesting one.

“I’ll focus on all users” is not a segment. “I’ll focus on power users because they are the most engaged” is lazy — power users are already engaged. The interesting segments are the ones where a specific pain point, if solved, changes behavior. Silent members who unmute. Browsers who finally click play. Passive scrollers who start visiting more often.

3. Confusing engagement metrics with value metrics.

Time spent on app is not always a good metric. If users are spending more time because they cannot find what they want, you have a search problem, not an engagement win. Match your metric to the outcome you actually want to produce.

4. Proposing solutions that already exist.

Before you suggest “add user profiles” to Netflix, check whether Netflix already has profiles. Before you suggest “add a mute button” to WhatsApp, know that muting already exists. This is basic product awareness. If you are doing a case on a product, use it for a week first.

5. Ignoring trade-offs.

Every improvement has a cost. More personalization means more data collection. Better feed filtering means some content creators lose reach. A new feature for one segment means engineering time not spent on another. If you present a solution with no downsides, the interviewer knows you have not thought hard enough.

6. Making it about technology instead of user behavior.

“We could use ML to improve recommendations” is not a product answer. It is an engineering suggestion. The interviewer wants to hear: which user, which behavior, which pain point, which outcome. The technology is a means, not the answer.

Where to go next

• Master the general framework: How to Approach Any PM Case Study
• Practice product design cases: Product Design Cases
• Sharpen your metrics thinking: Metrics & Analytics Cases
• Try India-specific challenges: Indian Market Cases