How would you improve a product you use every day?

What the question is screening for

The prompt is deliberately unbounded, and that is the test. There is no correct product and no correct feature, so the only thing left to grade is how you convert an open question into a decision. Interviewers are listening for three signals: that you impose a structure without being handed one, that your empathy attaches to a specific group of people rather than to "users" in general, and that every change you propose is derived from a goal instead of arriving as a standalone idea.

The third signal is the one that separates the round. Two candidates can propose the same feature and score very differently, because one arrived at it by narrowing from a goal and a segment while the other produced it first and reverse-justified it afterwards. Interviewers can hear the difference, mostly because reverse-justification never survives the follow-up "what would make this the wrong thing to build".

Clarify the goal and the business model first

Spend the opening ninety seconds establishing what the product is for and how it makes money. These are two separate statements and you need both. The goal tells you what counts as an improvement; the business model tells you which improvements the company would actually fund.

Say the goal out loud even when it feels obvious, because a stated goal is what you will point back to for the rest of the answer. Then state the revenue mechanism concretely: subscription, advertising, transaction fee, hardware margin, enterprise seats. Ask a clarifying question if you genuinely need one, but do not stall — in an interview you are allowed to state an assumption and move, and doing so reads as more decisive than interrogating the interviewer.

Pick one segment and one unmet need

Now narrow. Name two or three segments the product serves, say which one you are choosing, and say why — usually because it is large, badly served, or strategically important. Then describe that segment in enough detail that its need can be wrong: not "busy professionals" but "someone with a fixed forty-minute commute, half of it underground, who listens on a phone with limited free storage".

Choose a segment you belong to or have watched closely. The empathy in this answer has to be evidenced, and the only evidence available to you in a room with no research is specific observed behaviour. Then state exactly one unmet need. A single need keeps the rest of the answer coherent; three needs guarantee you run out of time before you reach a decision.

Generate options, then choose

Produce three or four distinct ways to address that need, not three variants of one idea. Distinct means they differ in mechanism — automate it, surface it at a different moment, remove the step entirely, let another user do it for you. Then choose one and give the reason: highest impact on the stated need per unit of effort, or lowest risk, or the one that unlocks the others.

Finish with a metric and a risk. The metric should be the behaviour you expect to change, not a proxy that would move for a dozen unrelated reasons, and you should pair it with a guardrail that would catch you harming something else. The risk is the assumption you are least sure of. Naming it yourself is worth more than defending it when challenged.

Worked: Spotify for the underground commute

Goal and model. Spotify's goal is to keep listening sessions happening on Spotify rather than anywhere else. It earns from paid subscriptions and from advertising on the free tier, and its cost base is dominated by per-stream music licensing. That last point matters and I will use it again: podcast and other non-music listening does not carry the same per-stream royalty, so a minute moved from music to podcast is a better minute for gross margin as well as being a longer, stickier session.

Segment and need. I am choosing daily commuters on a metro line with dead zones — my own case. The need is that the listening block is fixed and the connection is not. Today I remember to download episodes some mornings and not others, and on the mornings I forget, the queue stalls in the tunnel and I stop listening for the rest of the journey. The need is reliable audio for a known window, not more content.

Options. First, aggressively auto-download the next unplayed episode of every show I follow. Second, a commute mode that learns the journey window and pre-caches a queue sized to fit it. Third, a resilient player that degrades to a lower bitrate and buffers much further ahead when it detects a lossy connection. Fourth, nothing new in the product: just a notification at 08:15 reminding me to download.

Choice and why. The third option, because it addresses the need without asking the user to do anything or to trust a prediction. Aggressive auto-download is the obvious answer and it is the wrong one for this segment: phone storage is the constraint that made me stop downloading in the first place, so a fix that consumes storage silently trades one failure for another. Deep buffering costs storage only transiently and only for what I am about to hear.

Metric and risk. The metric is the share of commute-window sessions that run to completion without a stall, with total listening minutes for the segment as the confirming measure and music listening minutes as the guardrail so I can see whether I have merely shifted consumption rather than added it. The risk is that connectivity is not the real reason I stop — it may be that I lose interest at minute twenty and the tunnel is coincidence. That is cheap to check before building anything: look at where in the session playback stops for commuters on lossy connections versus clean ones, and if the drop-off is at the same point in both, my whole diagnosis is wrong.

The improvement has to pay for itself

Most weak answers fail in one of three ways, and they are ordered by how quickly they end the round. Opening with a feature list is the fastest — it tells the interviewer there is no goal underneath, and the rest of the conversation becomes them supplying the structure you did not. Choosing a segment you have never been part of is next, because your description of their need stays generic and every follow-up exposes that you are guessing. Both are recoverable if you catch yourself.

The third is subtler and it is where otherwise strong answers stall. Candidates name the business model in the first minute, then never use it again, and end up proposing something that costs the company money and returns goodwill. Naming the model is the checkbox; letting it eliminate an option is the signal. In the example above it does real work twice — it is why moving minutes towards podcasts is worth pursuing at all, and the storage constraint is why the crowd-pleasing auto-download option loses. If you can remove your own idea from the shortlist for a reason drawn from how the company earns, you have demonstrated the thing this question exists to find.

The answer is not the feature, it is the chain: goal, segment, need, options, choice, metric, risk. A strong candidate can be interrupted at any link and defend it from the one before.