What fields belong in an experiment tracking event?

WHAT THIS TESTS: Whether you understand that reliable experiment analysis depends on disciplined instrumentation and a single source of truth, not just capturing clicks and hoping the data is useful. Interviewers want to see that you think in terms of schema governance, developer alignment, environment hygiene, and long term maintainability from the moment an event leaves the client. They are listening for whether you treat the payload as a structured contract rather than an unstructured log line.

A GOOD ANSWER COVERS: First, a tracking plan that defines what events and properties to track, what they mean, and where they are tracked, because this becomes the contract between product, engineering, and analytics and prevents drift over time. Second, event properties and user properties that are explicitly scoped to the experiment so the payload is minimal and purposeful and every field has a documented reason for existing. Third, an environment identifier to separate dev and production data, since mixing them corrupts experiment metrics and can lead to false positives in conversion rates. Fourth, consistency in naming conventions and data types across the codebase so the schema stays simple and the tracking plan remains the single source of truth that analysts can trust. Fifth, only the fields required to answer the specific experiment question, because extraneous properties create noise, increase payload size, and create maintenance debt when schemas evolve.

COMMON WRONG ANSWERS: Dumping every available attribute into the payload without a schema or tracking plan, which turns the data stream into a swamp that analysts must clean later. Treating analytics as an afterthought rather than a planned instrumentation task that happens at the start of a feature. Failing to separate dev and production environments, which invalidates traffic counts, pollutes funnel metrics, and can break statistical significance. Using inconsistent naming conventions or casing that break the single source of truth and force analysts to write fragile transformation rules retroactively.

LIKELY FOLLOW-UPS: How do you validate that a new event matches the tracking plan before it ships to production and reaches experiment pipelines? How do you handle identity stitching when a user moves from an anonymous state to an authenticated state without double counting? What is your process for deprecating an event or property without breaking downstream dashboards, experiment results, and stakeholder reports?

ONE CONCRETE EXAMPLE: Suppose you are tracking a new onboarding button for an A/B test. The tracking plan specifies the event name, the goals and metrics that define success, the user properties needed for segmentation, and the exact location in the code base where the event must fire. The client payload includes only those defined event and user properties, plus an environment flag to keep dev traffic out of the experiment, ensuring the schema stays clean and the results are trustworthy. When the experiment concludes, the tracking plan makes it easy to know which properties can be retired and which must be preserved for future analysis.