Mixpanel Product Analytics Certification cheat sheet

Event = a tracked action with a name, a user identifier, a timestamp, and optional properties; every analysis (funnel, retention, segmentation) is a query over the event stream joined to user profiles.

Funnel conversion = (users who completed step N) / (users who completed step 1); step order matters, and an unordered funnel counts steps regardless of sequence.

Retention on day N = distinct users active on day N who were first active on day 0, divided by the day-0 cohort size; retention is cohort-based, not period-based.

DAU/MAU stickiness ratio measures how often monthly users return daily; values near 0.5 indicate habit-forming products, values near 0.1 indicate occasional use.

Segmentation splits any metric by an event property or user profile property; a property must be sent with the event or stored on the profile to be usable as a breakdown.

Identity resolution: anonymous_id is assigned pre-login, distinct_id is the canonical user key, and identify() merges the anonymous history into the known user so pre-signup events are not lost.

North Star Metric should be a leading indicator of retained value (e.g., weekly active teams completing a core action), not a lagging revenue figure.

Cohort analysis groups users by a shared first-event date or property; comparing cohorts over equal elapsed time isolates product changes from growth-mix effects.

Pageview tracking fires an event on every page load; single-page apps require an explicit track call on route change because no full page reload occurs.

UTM parameters (utm_source, utm_medium, utm_campaign, utm_content, utm_term) are captured as properties on the first event and used to attribute acquisition channel.

First-touch attribution credits the earliest campaign that brought the user; last-touch credits the most recent campaign before conversion - the same data yields different channel rankings.

Bounce rate = sessions with a single event / total sessions; a high bounce rate can be a tracking artifact when the landing page fires only one event type.

Referrer is the URL of the page that linked to yours; direct traffic means no referrer and no UTM, and is the default bucket for bookmarks, email clients, and untagged links.

Session definition is configurable (timeout window, typically 30 minutes); session-scoped metrics such as sessions per user depend on that window.

Cross-domain tracking requires the same project token and identity passing across domains, otherwise one user appears as two distinct users with split histories.

Heatmaps and click maps aggregate click coordinates over a page URL; they show where attention goes but not why a click did not convert.

Mobile SDKs queue events locally and flush in batches; a session that ends before flush can lose events unless the app calls flush on background.

App lifecycle events (install, update, open, background) are auto-tracked by the SDK and form the base of activation and retention cohorts.

Push notification events distinguish sent, delivered, opened, and bounced; open rate = opened / delivered, not opened / sent.

Device properties (OS version, app version, device model, carrier) are auto-collected and are the primary breakdowns for release-impact analysis.

Deep links and deferred deep links attribute installs to campaigns; a deferred link resolves after first launch, so install attribution depends on the link payload surviving install.

Mobile attribution differs from web: the install is the conversion, and the store (not the referrer) sits between the ad click and the app open.

App version breakdown is the fastest way to detect a regression: compare a metric across app_version before and after a release date.

IDFA/GAID availability is limited by user opt-in, so device-level attribution is probabilistic and should be reported with its match-rate caveat.

A/B test compares a control and a variant assigned by random bucketing; the unit of randomization (user, device, session) must match the unit of analysis.

Statistical significance is a p-value threshold, conventionally 0.05; p < 0.05 means the observed difference is unlikely under the null, not that the effect is large.

Statistical power (typically 80%) is the probability of detecting a real effect; underpowered tests produce false negatives and inflated effect sizes when stopped early.

Minimum sample size per variant grows with the square of the effect size denominator: halving the detectable lift roughly quadruples the required sample.

Peeking at results and stopping at the first significant reading inflates false-positive rate; fix the sample size or use a sequential method designed for continuous monitoring.

Guardrail metrics (crash rate, latency, revenue per user) must be checked alongside the primary metric so a win on one axis is not a hidden loss on another.

Novelty and primacy effects distort early readings; run at least one full business cycle and compare week-one versus week-two effects.

Exposure logging must record the variant at the moment the user is bucketed, otherwise users who never reach the feature dilute the treatment group.

Session replay reconstructs a user session from DOM mutation snapshots plus input events, not from video; the recording is a compressed event log replayed in a browser.

Masking rules (mask all text, mask inputs, block selectors) are applied at capture time; data not masked before transmission cannot be unmasked later for privacy compliance.

Replay sampling rate controls what fraction of sessions is recorded; 100% capture is rarely needed and increases storage and SDK overhead.

Replay links to the event stream by session and distinct_id, so a funnel drop-off step can be opened directly in the replay at that timestamp.

Console logs, network requests, and JavaScript errors can be captured alongside the replay to explain a rage-click or a failed form submission.

Rage clicks are repeated rapid clicks in a small area; dead clicks are clicks on non-interactive elements - both are computed from the replay event stream.

Replay is a qualitative tool: it explains why a metric moved but does not replace the quantitative funnel or retention analysis that detected the movement.

Retention of replay data is shorter than event data by default; confirm the project's replay retention window before relying on old recordings for an investigation.