product sense · standard
"How would you improve the Uber app?"
How would you improve the Uber app?
The most common mistake on this question is treating Uber as a single-sided consumer app. Uber operates in 70+ countries across 10,000+ cities, with a two-sided marketplace and a growing autonomous vehicle layer that has materially changed where the real product constraints are. The interviewer is checking whether you understand that before you propose a single feature.
Scope first, always
The prompt is intentionally broad. A strong candidate narrows it immediately, and does so in a way that signals marketplace fluency rather than framework compliance.
Before choosing a user segment, identify which side of the marketplace and which product surface (Rides, Eats, or Connect), then ask whether you are in an AV-deployed market. In Austin, Atlanta, Dallas, and Las Vegas, Uber has AV partners on the platform: Waymo in Austin and Atlanta, Avride in Dallas, Motional in Las Vegas. Waymo trips on Uber show 30% higher utilization and 25% faster ETAs vs first-party AV operators in comparable cities. Driver supply is no longer the binding constraint in those markets. MOIA (VW’s autonomous fleet) deploys on Uber in Los Angeles by end of 2026, with 15 cities targeted across Uber Autonomous Solutions.
In non-AV markets, driver supply and matching efficiency still drive most of the friction. Pick your context explicitly. It tells the interviewer you understand what you are actually solving.
Structure a strong answer
strong
"Let me scope before I propose anything. Core Rides product, rider side, mid-size US city without AV deployment. Non-AV markets still have real driver supply friction, and any improvement applies across most of Uber's global footprint.
Goal: 30-day repeat ride rate among infrequent commuters. Not DAU, not NPS. This segment takes 3 to 8 rides a month, owns a car, and uses Uber situationally: airports, nights out, bad weather. They churn when the experience disappoints on price or wait. One bad surge and they open Lyft next time. Acquisition cost is high; one-time riders do not pay back.
Core pain point: surge pricing and wait-time uncertainty arrive simultaneously, right when the rider needs to commit. They do not know whether waiting 8 minutes produces a cheaper ride or a worse one. This is not a pricing problem; it is an opacity problem. The behavioral outcome is predictable: they open a competitor tab, and sometimes they leave.
Proposed improvement: a 15-minute price and availability forecast on the booking screen, with a simple commitment prompt: "Lock in now / Wait and likely save X% / Wait and risk Y% increase." Uber already runs demand and supply prediction models at this granularity. This is not a new model; it is surfacing an existing signal at the moment of friction rather than hiding it behind algorithmic output the rider cannot interpret.
Viable: it targets the exact decision moment where churn is triggered, without subsidizing rides or changing pricing mechanics. Lovable: it meets the rider at the booking screen with one actionable signal, no notifications, no tutorials, no behavioral change required.
Primary metric: 30-day repeat ride rate for infrequent commuters in pilot cities. Secondary: ride completion rate from the booking screen. Guardrail: average fare in pilot vs control, confirming we are not training riders to hold out for lower prices. The main risk: a wrong forecast erodes trust faster than no forecast. Calibration is the engineering quality bar I would build into acceptance criteria before any pilot.
In AV markets, this gets more powerful. AV supply is more predictable than human-driver supply, so forecast accuracy improves. Austin and Atlanta are better versions of the same feature. That extensibility is part of the prioritization argument."
weak
"I would improve the rider experience by adding saved destinations, better ETA accuracy, fare splitting, and an in-app safety button." No user specificity, no marketplace awareness, no connection between any proposed feature and a metric that matters to Uber's business. The fatal problem: this answer works equally well for Lyft, Grab, or Bolt in any city in any year. Nothing in it reflects how Uber actually operates in 2026. The interviewer can tell this is a rehearsed framework, not product thinking. The common tell: the candidate mentions DAU and NPS as success metrics without explaining why those are the right measures for the specific problem they named.
The Uber One layer
Uber One is the cross-platform membership covering Rides, Eats, and Connect. Subscriber economics are meaningfully different from transactional riders: subscribers have higher LTV, use the platform across more surfaces, and are more price-stable. If you scope an improvement to the Rides product without acknowledging the membership layer, your segmentation is incomplete.
A senior answer notes that the improvement question bifurcates by membership status. Uber One members are less likely to churn over surge opacity because they have a benefits layer that buffers price uncertainty. The infrequent commuter churn problem is most acute in the non-subscriber population, which is also where a retention improvement has the highest LTV uplift. This distinction tells the interviewer you understand subscriber economics, not just user behavior.
The Jam Session format
Uber’s “Jam Session” round is a collaborative brainstorm with Uber employees, not a solo presentation. Candidates who have only prepared a framework walkthrough lose energy when the format becomes conversational and the Uber employee starts pushing on specifics. Knowing the AV deployment map, the Uber One layer, and the marketplace mechanics by market type is what lets you engage as a peer rather than a candidate performing a script.
The 2026 reframe
Before AI-era routing and AV partnerships, improving Uber meant improving supply: more drivers, better dispatch, lower wait times. That problem is being structurally solved. Uber collected 3M+ hours of robotaxi-specific driving data via its NVIDIA Cosmos partnership for L4 model training. Waymo on Uber delivers measurably better utilization and ETAs than human-driver markets. The frontier has moved.
The 2026 version of this question is a viability and lovability problem. Viable: which rider problems are large enough that solving them creates measurable LTV, not just an engagement tick. Lovable: does the improvement meet riders in the actual moment of friction, without obnoxious push notifications or onboarding flows, just the right signal at the right time. The candidate who answers this as a dispatch-and-supply optimization problem is answering 2019’s question.
See the CIRCLES framework for the full answer structure, feasibility is free for the broader 2026 lens, and lovable, not just usable for the distinction between features that pass the bar and features that merely clear it.