Introduction
Today, Peloton boasts over 6 million active members who can explore a vast range of workouts across 16 fitness modalities - available on Bike, Tread, Rower, and the Peloton App. Members can choose their favorite instructors, music, and non-instructed content for a fully personalized experience. With the recent launch of Peloton IQ cross training features this Fall, we have expanded our member Personalization features beyond only content recommendations, enabling our members to get a holistic view and guidance into their fitness journey using cutting-edge AI/ML technologies. One of the key features, Insights and Analysis, turns all member activities, both on and off-platform, into actionable insights for a more personalized and motivating journey by surfacing meaningful trends in performance, recovery, and consistency that align with their fitness goals. The feature also guides them on what to do next. Figure 1 shows current choices of fitness goals and example insights on the Member Profile and Homescreen pages.
Insights on fitness platforms
The fitness technology space is already replete with various AI generated “Insights” features that leverage data (primarily health metrics from wearables) to provide users with great summaries. For instance, Strava's Athlete Intelligence provides users with performance trends, milestone celebration and feedback. Similarly, Apple Health aggregates data from various sources and applications to provide a comprehensive view of a user's health metrics and trends, while Garmin Connect+ provides motivating personalized insights and suggestions throughout the day based on activity data. All in all, modern fitness platforms underscore a broader industry trend toward turning raw activity data into actionable and personalized guidance.
Insights on Peloton IQ
Peloton is in a unique position to not only provide these data-driven insights, but also supplement them with concrete action suggestions, for instance, taking classes that align with their fitness goals and preferences (class recommendations), or making adjustments to their Weekly Plans, and many more. Figure 2 shows another view of Insights with class recommendations (horizontally scrollable on Bike Homescreen).
These action suggestions aren’t just tied to the Peloton platform - they can include data from third parties (e.g., Apple Health and Garmin, upon member consent) and suggest off-platform activities, such as increasing distance on running or various recovery activities.
“Consistent rest is powerful - schedule a recovery day each week with gentle stretching or foam rolling to aid muscle repair and boost gains.”
Moreover, Peloton IQ has introduced Strength workout tracking features with new Bike+, Tread+ and Row+ hardware that come equipped with an enhanced camera and Computer Vision technology. Telemetry data from the strength workouts such as Rep Counting and Form Feedback add a new dimension and lay the foundation for true cross training guidance for Peloton members. Leveraging these metrics, insights can guide the members by reminding them when to level up (reps and weights) or stay consistent. Here’s an example insight that leverages strength data from telemetry.
“You've been chest pressing 25 lb dumbbells recently. Try adding 2 reps or increasing weight by 5 lbs next week to tap into progressive overload.”
Generating Insights and Actions
With Insights, Peloton aims to bridge the gap between simple activity tracking and true behavioral change, driving deeper engagement across all our fitness modalities. To drive this vision, we’ve built an Insights Generator Engine that leverages a solid data infrastructure to glean member contexts, and makes use of AI in order to generate Peloton Insights.
At the data infrastructure layer, the system ingests and processes a wide array of data streams, including all on-platform activities, off-platform (3P) metrics, and the novel telemetry data from strength tracking features. Example Peloton workout metrics include but are not limited to, distance, output, power, heart rate zones and power zones. It also includes modality specific metrics, such as, cadence and resistance for cycling, and incline, pace, speed for running, and many more. Other information, such as the user’s weekly plan, activity targets and their fitness goal are also fed as the contexts to the Insights Generator engine.
Once Insights Generator receives all contexts, it arranges the history, metrics and other metadata along with a long set of instructions for a LLM (Large Language Model) to generate various types of Insights (Tips, Performance, Suggested Targets, Your Plan). For generating Action Suggestions, the LLM returns specific class Filters and Modification Operation, which are further translated into concrete actions by leveraging other services, such as the Class Recommender and the Weekly Plan Generator system, as shown in Figure 3.
Action Suggestions
Unlike celebratory and milestone insights which focus on past achievements, Action Suggestions from IQ Insights are forward-looking. By synthesizing a comprehensive picture of a member's fitness journey, Peloton insights move beyond simple summaries to generate highly personalized actions that align with individual fitness goals and promote progression and consistent recovery. The key categories of action suggestions currently powered by Peloton IQ are the following.
Browse Classes and Class Recommendations
These suggestions provide members with specific Peloton classes that align with their preferences and activity history. Consistent with the Insights text (e.g., "increase cardio endurance" or "build lower body strength"), the LLM translates the desired action into specific class Filters (e.g. fitness modality, instructor, class duration). The Insights Generator Engine then creates a candidate set of classes using these filters, and feeds them into the Class Recommender service, which returns highly relevant, personalized class suggestions.
Activity Targets and Plan Modifications
Insights shine in helping Peloton members to track progress toward their Weekly Activity Target (Active Days, Active Time and Workouts) and Weekly Workout Plan, promoting consistency and progressive training. For example, the insight might suggest - "You’ve consistently hit 3 workouts a week, let’s modify the target to 4 days.", with an in-place “Update Target” button.
Similarly, if a member is enrolled in a Personalized Plan, the insight may suggest - "Swap Thursday’s recovery day for a 20-min Yoga Flow to bring mobility and balance to your plan." , as well as returning Insights Generator engine the associated Modification Operation to apply to the current member Plan. Once the member chooses to tap on the “Update Plan” button, the Operation would be executed by the Weekly Plan Generator system, updating the member's personalized plan to reflect the new, challenging, yet attainable target for the following week.
Education and Contextual Tips
Action suggestions often come bundled with educational context (More Info) to help members understand the why behind the suggestion, promoting fitness literacy and long-term behavioral change. For example, a tip might explain the principle of "progressive overload" when suggesting an increase in weight, or the importance of "active recovery" when recommending a gentle stretching class.
Results
Before we launched this feature, we conducted multiple A/B tests over several months with a small group of field testers. During the tests, the insights were generated using multiple LLMs and the testers rated them on criteria, such as Personalization, Motivation and Relevance. A majority of the respondents gave positive ratings across most of the criteria. Some negative feedback was primarily from members who were either skeptical of AI or did not realize that they could connect their third-party workout data to give the system a full picture of their workout routine. We made our final choice of the model that strikes the right balance between cost and quality.
Additionally, we established an offline evaluation framework utilizing an LLM-as-a-judge approach to assess the quality of the insights. Insights were scored from 0-1.0 based on the following key criteria: Factuality, Personalization, Actionability, Insightfulness, Safety, Tone, and Toxicity. Figure 5 shows an illustration of LLM judging on an insight. To eliminate any bias, the model used for judging is different from the model used for generating insights. We are working toward incorporating both human feedback and LLM scores to further improve the quality of Insights.
After the launch, the Insights and Analysis feature was very well received by the Peloton members. Profile Insights, which are deep in nature, have seen a whopping 90% positive rating! Since the launch of this feature, we also have observed around a 10% increase in profile page views (month over month) among our members across platforms.
Conclusion
The introduction of Peloton IQ's Insights and Analysis marks a significant leap in our Personalization mission to move beyond content recommendations toward providing holistic, personalized fitness guidance. By leveraging cutting-edge AI/ML, rich member data, and a robust Insights Generator Engine, we are successfully translating member activity into actionable insights that promote progressive overload, consistent recovery, and meaningful engagement across all fitness modalities.
Future work
Based on member feedback, we are working toward enhancing the instructions and contexts that support our existing Insights, as well as expanding to new types of insights. We want to ensure that the insights and action suggestions lead directly to a tangible and quantifiable progress toward members’ fitness goals. We are also working toward incorporating new information into insights and a continuous quality improvement.
Our current system Insights generation is a combination of batch and real-time processes. The batch generation of Insights, a decision influenced by high LLM inference latency, limits our ability to modify the Insights immediately after a workout is completed (e.g., providing post class insights). We are keen on bridging the gap in order to support real time generation of Insights.
Acknowledgement
We have been fortunate to build this feature together with a very talented team at Peloton. Our Product, Design, Research and Engineering teams have put a great effort into bringing this to life! We specially thank Emily Kane, Tanya Lerma, Crystal Qin, Aditya Batheja, Mohit Jeste, Alberto Villafane, Santhosh Thammana, Greg Johnsen and Saif Khan among many others who navigated through various design and engineering challenges and launched this Peloton IQ feature.
