Steal This Research Playbook from Uber (Conjoint Analysis Case Study)
โWhen Uber built UberPool and ExpressPool, its researchers ran a four-step playbook to work out which carpooling trade-offs riders would accept: identify trade-offs with user interviews, measure them with MaxDiff analysis, optimise them with conjoint analysis, then implement them with usability testing and a pilot launch. The core finding: extra discounts mattered far more to riders than the annoyance of walking a block or waiting a few minutes, so Uber leaned into cost savings and let riders trade small inconveniences for a cheaper ride. This is a real, published case study (from a 29-page 2018 paper by Uber researchers Jenny Lo and Steve Morseman), and itโs one of the clearest real-world examples of conjoint and MaxDiff being used together on an actual product decision.โ
The UberPool research playbook
Each stage of this playbook built on the insights of the step before it:
Identifying Trade-Offs โ User Interviews
Measuring Trade-Offs โ MaxDiff Analysis
Optimising Trade-Offs โ Conjoint Analysis
Implementing Trade-Offs โ Usability Workshops
And Uber wants you to steal it. They published this case study as an academic paper specifically to "assist readers in executing similar research methods."
I love seeing teams share their knowledge like this. We've recently published research playbooks from Dogo, Labster & OpinionX (yes, a case study about our own research), and we're keen to keep this anti-gatekeeping movement going. If your team has an interesting research project you can post publicly, get in touch through The Full-Stack Researcher. Together we'll turn it into a research playbook and share it with the newsletter's 42,000+ user researchers.
Step 1: Identifying trade-offs
UberPool expands access to Uber's services by driving down the price per user through shared rides. But as the screenshot above shows, the trade-off was painful: unreliable arrival times and unpredictable detours for extra pickups along the way.
To fix this, the UberPool team first needed to understand how people think about their transportation choices. They interviewed 23 people in Chicago and Washington DC, mixing new Uber users, experienced riders, and people who'd never tried Uber before. Each 2.5-hour interview took place in the person's home and began with a "travel mapping" exercise to lay out their typical trips: school runs, work commutes, grocery shopping.
^ Real examples of artefacts used or created by interviewees during the โtravel mappingโ exercise.
Next, the researchers dug into each individual trip from the mapping exercise. Using Ulwick's Eight Process Steps (define, locate, prepare, confirm, execute, monitor, modify, and conclude), they unpacked each trip step by step. As interviewees described their trips, the researchers looked for three key influences:
Functional โ what the person needed to accomplish (price, timing, logistics).
Emotional โ how the person wanted to feel, or avoid feeling (emotions, frustrations).
Aspirational โ how they want to be perceived by others (status, values, identity, image).
Functional influences dominated the interviews, especially efficiency: route planning, waiting time, arrival reliability. These were the non-negotiables shaping people's everyday travel decisions. When the researchers compared these needs against the UberPool experience, the gap was obvious. Unlike a regular Uber trip, where drivers can reroute to avoid traffic, UberPool's algorithm locked trips to a fixed route. Riders had little control over other important factors too, like detours, pickup time, co-passengers, or car size.
That raised a critical question. What if Uber gave riders more visibility or control over these inefficiencies? Any new information or flexibility would cost Uber something, so it would only be worth it if riders truly valued the changes.
Step 2: Measuring trade-offs
The researchers listed all the functionality factors shaping the UberPool rider experience, then removed price (tested separately later) and anything they couldn't directly control in the product. That left six factors:
Expected Arrival Time
Fixed Route
Number of Stops
On-Demand Availability
On-Trip Duration
Walking
The challenge was working out how much each factor actually mattered to riders. In scenarios like this, academic research warns against simple rating (1-10 scale) or full ranking polls (1st to last), because both tend to produce flat, clustered results where everything ends up looking similarly important. So they ran a MaxDiff Analysis survey instead.
^ Example of a MaxDiff Analysis survey hosted for free on OpinionX
Their MaxDiff survey showed four factors at a time and asked participants to pick the most and least important factor from that set. After each vote, another random set of four appeared, repeating until the participant had voted on five sets in total.
Once complete, MaxDiff automatically calculates a relative importance score for each factor, showing how much riders value it compared to the others. These scores fall between -100 (always "least important") and +100 (always "most important") using a simple scoring formula. Uber used a more advanced regression-based algorithm, but the outcome is similar as long as you have strong sample sizes.
Within the same survey, participants also answered questions about their UberPool usage, commuting habits, and broader transport needs. The team recruited riders who'd used UberPool in the last 30 days in Boston, Washington DC, New York City, Chicago, and San Francisco. 5000 randomly-selected riders were invited per city. In total, 3000 completed the survey, a 12% response rate with a median completion time of just five minutes.
The MaxDiff results were crystal clear. On-Trip Duration (the total time spent in the UberPool) was the most important factor, validating the interview findings. Next came On-Demand Availability (how quickly a car would be ready) and Expected Arrival Time. At the other end, Walking (the distance between a rider's actual location and the pickup or dropoff) ranked least important overall. Walking would also have been the cheapest variable for Uber to adjust, so these results instantly sparked some new ideas.
The team started sketching a new concept called the PerfectPool: a shared trip where multiple riders with similar start and end points travel together without any mid-route pickups or detours. PerfectPool optimised the most important factors (duration and arrival time) but required riders to walk before and after their trip. Which raised a big question: would riders trade some walking for a faster, more reliable UberPool experience?
Step 3: Optimising trade-offs
By now the researchers knew walking distance itself wasn't important to riders, but at what distance would required walking make an UberPool ride unattractive? To find out, they needed a way to test different scenarios and measure how preferences shifted as walking distances increased. That's where Conjoint Analysis comes in.
^ Example of a free Conjoint Analysis survey on OpinionX
Conjoint analysis shows people multiple versions of a product with different combinations of features, asks them to pick their preferred version, and looks at the patterns in their choices across many votes. For PerfectPool, the "features" were Availability, Walking, Trip Duration, Potential Delay, and Price. Each factor held a range of options (say No Walking, Walk 1-Block, Walk 2-3 Blocks). Here's an example of the inputs they added to their conjoint survey:
This table of categories and options is most of the work that goes into building a conjoint survey. Once you've decided on these, you drop them into a survey platform that supports conjoint analysis, like OpinionX (which lets you test conjoint surveys for free). After adding your categories and options, you make a few survey design decisions:
Profiles Per Set โ how many profiles appear together on screen (Uber chose 2).
Sets Per Participant โ how many times someone picks their preferred profile before the survey moves on (Uber went with 7 sets; more advice here).
"None" Option โ lets participants indicate they aren't interested in any of the profiles (Uber included this).
Matching Rules โ rules to prevent unrealistic combinations (Uber skipped this and allowed full random, which is best practice and makes for easier setup).
With 5 categories of 3-5 options each, there are 900 possible profiles the survey tool could display to participants.
The survey went out to 18,000 UberPool riders by email on August 3rd, 2017. As an incentive, riders could win one of five $500 Amazon gift cards. When sending the invites, the research team linked existing rider data (city, product usage, lifetime billings, and survey variables) so they could compare results by user segment later. A total of 1,934 completed the survey (10.7%) with a median completion time of 5.5 minutes. The overall results showed the relative importance of each option. Comparing those scores against the "None" choice showed whether each option increased or decreased the likelihood that a rider would choose UberPool:
^ This chart uses wording based on internal language used by the UberPool researchers rather than the labels shown to participants on the conjoint survey, so it may look a little confusing at first, but it is very insightful!
On this graph, the slope of each line shows how much that factor influenced rider preference. As expected, price (purple line) had the biggest effect. Walking (yellow line) showed the steepest negative slope, meaning the further people had to walk, the more likely they'd reject UberPool. Trip variance (potential delays, the blue line) had the smallest effect on ride uptake.
The curve for ETA (driver availability) was especially interesting. People preferred immediate rides, but they were open to booking in advance within reason: the red curve dropped sharply where preferences plummeted somewhere between 15-30 minutes of waiting time. That revealed a big opportunity. Riders could be asked to wait a bit longer upfront (within reason) in exchange for bigger discounts, giving Uber extra time to find other riders on roughly the same route to match together without any mid-ride pickups.
Filtering the conjoint scores to compare segments, they also found that:
Riders with 2+ years on Uber were less likely to opt in.
New riders were more price-sensitive than veteran users.
People with short commutes were more sensitive to walking requirements.
These segmented insights were key to fine-tuning the next stage: product design and pilot launch for the newly named ExpressPool service.
Step 4: Implementing trade-offs
The conjoint analysis showed that extra discounts had a far greater positive effect on rider preference than the negative effect of inconveniences like walking or waiting. Uber's leadership used this to justify tasking the UberPool team with finding efficiency wins that could deliver the lowest possible cost to riders.
i. Valuing inconveniences
To put exact numbers on how much of a discount was needed to offset inconveniences like walking or waiting, the team used a scenario simulator. Scenario simulators are unique to conjoint analysis. You create two competing profiles and the simulator predicts which one each participant would have picked, with results like "70% for Profile1 vs 30% for Profile2". This let the researchers set Profile 1 as a default UberPool ride (no discount, no walking, no waiting) and Profile 2 as an ExpressPool option (10% discount, 1-block walk, 15-minute wait). Once the simulator predicted a 50/50 split, they knew the trade-offs were balanced.
^ OpinionX surveys come with an easy-to-use automated simulator for analyzing conjoint results (the screen recording above is from OpinionX).
ii. Usability testing
Uber already displayed waiting time, but showing walking distance before a ride was confirmed was a new usability challenge. The first mockup (left), showing hypothetical pickup points on a zoomed-in map, failed in usability tests when users assumed they needed to select a preferred pickup point before booking. The second mockup (right) zoomed out too far and left riders unable to judge whether the walk would be reasonable. Both failed because they gave riders less clarity and control, not more. After several iterations, the team landed on a design with clear pickup zones and transparent information on price, wait time, and walking distance, giving riders exactly enough context to decide whether ExpressPool suited their trip.
iii. Pilot launch
The pilot rollout in Boston and San Francisco in November 2017 quickly showed that riders cancelled less, complained less, and used ExpressPool more than expected. Within six months:
ExpressPool cancellations fell 40%.
Customer support requests dropped 17%.
Usage exceeded projections by 4.6%.
ExpressPool expanded to 6 more US locations in February 2018, followed by a full rollout to cities across four continents in summer 2018. The data gave the team real conviction to push ahead with the expansion and plan a big media launch. ExpressPool's launch tagline, "walk a little, save a lot", shows how central the research team's trade-off insights were to the product's strategy and positioning.
Product research done right
This project shows what good product research can do: close the gap between a company's goals and its customers' needs, and reach a solution that works for both. The impact rippled across Uber. The conjoint insights were presented at offsites, strategy sessions, and vision workshops throughout 2017 and 2018, reinforcing a simple truth: an efficient rider experience paired with the market's lowest prices could unlock big growth in the carpooling category.
This article is only possible thanks to the researchers who led the project, Jenny Lo and Steve Morseman, who published their work in a 29-page paper back in 2018. It's rare that researchers get to share detailed case studies like this with real-world competitive insights. Jenny and Steve not only achieved their mission of making carpooling cheaper and more reliable, they also inspired countless researchers to try survey methods like conjoint and maxdiff analysis. My hope is that this breakdown shows these methods aren't as intimidating as they might seem: any product manager or researcher can use them with confidence.
The Full-Stack Researcher newsletter exists to help more people confidently master quantitative research methods. It's also why we built OpinionX, a survey platform that makes it far easier and cheaper to use methods like conjoint and maxdiff analysis in your own product research. If you enjoyed this, subscribe for more case studies and research playbooks straight to your inbox.
Until next time, Daniel
Frequently asked questions
What research methods did Uber use for UberPool? Four in sequence: user interviews to identify the trade-offs riders face, MaxDiff analysis to measure which factors mattered most, conjoint analysis to work out how preferences shifted as those factors changed (like walking distance and price), and usability testing plus a pilot launch to implement the findings.
Why did Uber use MaxDiff instead of a rating scale? Because rating scales (1 to 10) and full rankings tend to produce flat, clustered results where everything looks similarly important. MaxDiff forces a choice between the most and least important factor in each set, which separates them cleanly.
What did the conjoint analysis actually tell Uber? That extra discounts had a far greater positive effect on rider preference than the negative effect of inconveniences like walking or waiting. That insight justified pushing for the lowest possible price, which became ExpressPool.
Can I run the same methods Uber used? Yes. MaxDiff and conjoint analysis both run on OpinionX, including the scenario simulator Uber used to put a price on each inconvenience. The playbook in this article is deliberately reusable, which is why Uber's researchers published it.
Who published the original Uber research? Uber researchers Jenny Lo and Steve Morseman, in a 29-page paper in 2018, specifically so other teams could copy the approach.
The lesson from Uber isn't the finding so much as the order they found it in: each method answered a question the last one raised, and handed a sharper question to the next. That sequencing is the part any product team can copy, whatever they're building.
Want more like this? Over 42,000 researchers and product people get The Full-Stack Researcher in their inbox, including case studies and playbooks like this one. Subscribe for the next one.
You can run every method in Uber's playbook on OpinionX for free: MaxDiff, conjoint analysis and the scenario simulator, at $0 for up to 25 participants per survey, then $900 a year for unlimited participants (full pricing). See how another team did it in the Airbnb MaxDiff case study.