Amplitude Certification cheat sheet

Event segmentation counts or filters events per user; a chart's metric (event totals, uniques, properties) determines whether results are event-level or user-level.

Active/New/Returning users are defined by first-seen date: New = first event in range, Returning = seen before but not in range, Active = any event in range.

Retention is measured as a cohort of users who performed a start event in period 0 returning to a return event in period N; N-day and unbounded windows differ.

Funnel conversion requires ordered steps within a conversion window; users who skip a middle step do not convert unless the funnel is set to unordered.

A user property is set once per user (or overwritten); an event property is scoped to a single event occurrence and can differ each time.

Formula metrics combine events with arithmetic (A/B, A/B*100); divide-by-zero yields null rather than zero.

Group-by on a property splits a chart into series; group-by user property assigns the user's current value, not the value at event time.

Percentile and distribution charts require numeric event properties; median is more robust than mean for skewed revenue data.

The Amplitude ID is the internal user identifier; User ID and Device ID are the identity keys used to stitch events to a single Amplitude ID.

Identity merging occurs when a User ID is set on a device previously tracked by Device ID; events before the merge are attributed to the merged profile.

Event and property volume is governed by the plan's event quota; exceeding it can drop or block ingestion depending on contract terms.

Taxonomy (Data) governance uses tracking plans to flag unexpected event types and property values via governed schemas.

Cohorts can be synced to destinations (e.g., ad platforms, warehouses) and are refreshed on a schedule; membership is not real-time by default.

Event streaming exports raw events to a warehouse or S3 destination; the export schema includes event_type, user_id, and server_upload_time.

PII should be handled via property blocking or deletion requests; Amplitude supports user deletion APIs for GDPR/CCPA erasure.

Source-level settings (e.g., SDK, HTTP API, warehouse import) determine ingestion path and can affect latency and dedup behavior.

A flag is evaluated per user via targeting rules; the first matching rule wins, and a default/fallback variant applies when no rule matches.

Bucketing is deterministic on a hashing key (often user ID or device ID); the same user receives the same variant unless the key or allocation changes.

An experiment requires a primary metric, exposure event, and variants; statistical significance depends on sample size and effect size.

Exposure (assignment) events must fire only when the user actually sees the treatment, otherwise intent-to-treat vs. treatment effects diverge.

Mutually exclusive groups prevent a user from being in two conflicting experiments simultaneously.

Holdouts withhold treatment from a slice of users to measure cumulative long-term impact across experiments.

A feature flag rollout percentage is independent of experiment allocation; changing rollout can break experiment validity.

Sequential testing and fixed-horizon p-values answer different questions; peeking at fixed-horizon results inflates false positives.

A session is a sequence of events with gaps under the session timeout (default 30 minutes of inactivity); replay is scoped to a session.

Replay captures DOM mutations and user interactions; it does not record keystrokes in masked fields when masking is enabled.

Masking options include masking all text, masking specific elements by CSS selector, and blocking elements entirely from capture.

Replays link to the user timeline, so you can jump from an event (e.g., error) to the replay moment for that user.

Replay availability depends on sampling and plan quotas; not every session is recorded.

Console logs and network requests can be captured alongside replay for debugging, subject to privacy settings.

Replay search filters by properties like country, device, or funnel step to find sessions matching a behavior.

Cross-domain replay requires the SDK to be initialized on each domain with consistent identity to stitch sessions.

Heatmaps aggregate click, tap, and scroll activity over a page's rendered layout, not raw coordinates in isolation.

Click maps show where users click; rage clicks are repeated rapid clicks in a small area, indicating frustration.

Scroll maps show how far down the page users reach; the depth metric is the maximum scroll position per session.

Heatmaps require page/screen URL matching; dynamic content or A/B layouts can misalign aggregated data.

Element-level heatmaps attribute activity to a DOM element, while coordinate heatmaps show raw position on the page.

Dead clicks are clicks on non-interactive elements; they surface UX issues where users expect interactivity.

Heatmap data is sampled and aggregated; small samples produce noisy maps and unreliable conclusions.

Mobile heatmaps map to screen names rather than URLs, since native apps lack a URL concept.

A guide is an in-product message (tooltip, modal, banner) triggered by a targeting rule; a survey is a structured question set.

Guide targeting uses user/event properties and audience cohorts; the trigger event determines when the guide appears.

Survey question types include NPS, rating, multiple choice, and open text; NPS is scored 0-10 with promoters 9-10 and detractors 0-6.

Throttling and frequency caps limit how often a guide or survey is shown to the same user.

Guide eligibility requires the user to match the audience and not have already seen/dismissed the guide per its display rules.

Survey responses are stored as events/properties that can be analyzed in charts and joined with behavioral data.

A/B testing a guide compares variants by a chosen goal metric (e.g., conversion or completion rate).

Guides render via the SDK on web and mobile; a guide won't display if the SDK isn't initialized or the page doesn't match targeting.