Product analytics

Product Analytics

You cannot improve what you do not measure, but most teams measure the wrong things badly. A messy event taxonomy is technical debt that misleads every decision built on it. This skill sets up analytics that actually inform product work.

Start With Questions, Not Events

Before instrumenting anything, write the decisions analytics must inform:

• Where do users drop off before first value?
• Which features correlate with retention?
• What does an activated user do that a churned one did not?

Instrument backward from these questions. Tracking everything "just in case" produces noise nobody queries.

Event Taxonomy

A disciplined naming convention is the foundation. Get it wrong and you spend a year cleaning data.

• Use a consistent structure: object-action (e.g. project_created, invite_sent). Pick one convention and enforce it.
• Define properties deliberately: each event carries the context you will segment by (plan, source, role).
• Maintain a tracking plan - a single source of truth listing every event, its properties, and its owner. This is the contract between product, eng, and data.
• Size it: an early-stage tracking plan should be roughly 20-50 events, not hundreds. If the plan exceeds ~100 events before product-market fit, it is tracking implementation details, not decisions.
• Govern changes: new events go through review so the taxonomy does not rot.

The North Star and Inputs

• Pick one north-star metric that captures delivered value (e.g. weekly active teams, queries run, documents shipped) - not a vanity count.
• Decompose it into input metrics you can actually move (acquisition, activation, engagement, retention).

Core Analyses

Funnels

Map the path to first value and measure conversion at each step. The biggest drop-off is your highest-leverage fix. Watch time-to-convert, not just rate: for a self-serve product, most activation that will ever happen happens in the first session or first day - a median time-to-first-value beyond a day is itself the finding.

Cohort Retention

Group users by signup period and track retention over time. A healthy product shows a retention curve that flattens (a "smile" for the best products); a curve decaying to zero means no product-market fit yet.

How to read the curve:

• Judge where it flattens, not the month-1 number - the asymptote is the product's real retained base.
• Consumer apps: roughly 40% day-1 / 20% day-7 / 10% day-30 is a good curve; well under that (say ~25/10/5) means the leak comes before the habit forms. A consumer curve flattening above ~20% is a strong signal.
• B2B SaaS: retention is measured in weeks/months, not days; expect the curve to flatten within the first 4-8 weeks, and monthly logo retention below ~95% (churn above ~5%/month) is a red flag for anything sold to businesses.
• Compare newer cohorts against older ones - cohorts flattening higher over time is the clearest evidence the product is improving.

Activation

Define the activation event - the action that predicts long-term retention (the "aha moment"). Find it by comparing what retained users did early versus churned users: take users still active at week 4+, contrast their first 1-7 days of behavior against churned users' first 1-7 days, and look for the early action with the biggest retention gap. Then optimize the funnel to it.

Feature Adoption

Track breadth (how many use a feature) and depth (how often). Correlate adoption with retention to prioritize the roadmap.

Avoid Vanity Metrics

• Total signups, total pageviews, cumulative anything - they only go up and inform nothing.
• Prefer rates and cohorts over totals; prefer leading indicators over lagging ones.

Tooling and Hygiene

• Choose one analytics tool as the source of truth; pipe to a warehouse for deep analysis.
• Validate instrumentation regularly - silent tracking bugs corrupt every downstream decision.
• Document metric definitions so "active user" means the same thing to everyone.

Quality bar

• Every event in the tracking plan traces back to a named decision question; none exist "just in case".
• The taxonomy follows one object-action convention with zero exceptions, and the plan lists properties and an owner per event.
• The north star measures delivered value (a rate or active count, never cumulative) and decomposes into movable input metrics.
• The retention read states where the curve flattens and against which benchmark, not just a month-1 percentage.
• The activation event is backed by a retained-vs-churned comparison, not picked by intuition.

Deliverable

Produce an analytics plan: the decision questions, a north-star metric with input-metric tree, an event tracking plan with naming convention, and the priority analyses (activation funnel, cohort retention) to build first.