Skip to main content

Understanding Experimentation Metrics

Experimentation metrics can be described using many attributes, often combining those attributes together. In this article we try to explain the most important attributes and what they mean in the context of experimentation.

Role

In ABsmartly and many other experimentation platforms, metrics are often described as primary, secondary, guardrail, exploratory, those attributes describe the role that the metric plays in the experiment.

Primary metric

In an experiment the Primary metric is the single most important measure used to determine whether the tested change achieves its desired outcome and whether or not the hypothesis is validated or rejected. It reflects the experiment's primary objective and directly aligns with the business’s strategic goals. The primary metric is the metric used to inform the experiment design regarding defining the minimum detectable effect (MDE) and the sample size (to ensure sufficient power to detect a meaningful change).

Examples:

  • revenue_per_visitor
  • conversion_rate
  • retention_rate

Secondary metrics

Secondary metrics, while not the main decision-making criteria, play a big role in ensuring a comprehensive understanding of the experiment’s impact. They provide additional context and insights beyond the primary metric and can help detect unintended side effects.

Examples:

  • items_added_to_cart
  • product_page_view
  • banner_interaction

Guardrail metrics

Guardrail metrics are safeguards used to monitor and ensure the health, stability, and overall integrity of the system during an experiment. They do not measure the success of the primary business objectives but are critical for detecting unintended negative impacts on the business, user experience and/or operational performance. Guardrail metrics act as early warning systems, identifying potential risks such as degraded performance, increased errors, or adverse user behavior before they escalate into larger problems.

Examples:

  • errors
  • app_crashes
  • page_load_time
  • support_tickets

Exploratory metrics

In ABsmartly, exploratory metrics refers to metrics of interest not used in decision-making. Exploratory metrics are often used in post-analysis and are a great source of insights on top of which new hypotheses can be built. Exploratory metrics should not be used to evaluate the experiment.

Purpose

A metrics can be described as a business metric, a behavioural metric or an operational metrics. Those attributes describe the purpose of the metric, what it is measuring.=

Business

In experimentation, business metrics refers to metrics measuring the impact of a change on a business KPI. Business metrics are often used as primary and/or guardrail metrics.

Examples:

  • revenue_per_visitor
  • conversion_rate
  • retention_rate
  • calls_to_customer_support

Behavioural

Behavioural metrics are metrics measuring the impact of a change on the visitor's behaviour. Behavioural metrics are usually measuring the direct impact of a change and as such have high sensitivity. Behavioural metrics are often used as secondary metrics.

Examples:

  • items_added_to_wishlist
  • clicks_on_banner
  • product_page_views

Operational

Operational metrics, also known as technical metrics, measure the impact of a change on system performance. Operational metrics can be used as guardrail metrics but also possibly as primary or secondary metrics depending on the goal of the experiment.

Examples:

  • page_load_time
  • app_crashes
  • error_rate

Data structure

All metrics are either binomial or continuous, this is a reference to how the underlying data is structured and measured.

Binomial

Binomial metrics represent a binary outcome for each visitor in the experiment, where each instance falls into one of two categories (e.g., success/failure, yes/no, 0/1). They are typically represented as a percentage (ie: 10% conversion rate), binomial metrics follow a normal distribution. Binomial metrics are easier to interpret and communicate.

Examples:

  • conversion_rate
  • click_through_rate (ie: the percentage of users clicking on a link)
  • churn_rate - email_open_rate

Continuous

Continuous metrics on the other hand can take on a wide range of values (either measured or counted). Continuous metrics often represent quantities or durations. Their underlying distribution varies depending on the data. Continuous metrics are more sensitive (they capture a wider range of data) and offer more insights but they can be heavily influenced by outliers and are harder to interpret.

Examples:

  • time_on_page
  • time_to_first_booking
  • number_of_items_in_cart
  • revenue_per_visitor

Time horizon

Another aspect of experimentation metrics is their time horizon, typically metrics can be referred to as short-term or long-term.

Short-term

Short-term metrics refer to metrics that measure immediate or near-term outcomes, typically during or shortly after the experiment. They can typically be measured accurately in the experiment’s runtime and provide quick feedback on the effects of changes.

Examples:

  • real_time_conversion_rate (during the test)
  • time_spent_on_page
  • click_through_rate

Long-term

On the other hand, long-term metrics measure delayed outcomes which make it hard to measure during the runtime of an experiment. Typically long-term metrics represent the strategic goals and align with the desired business outcomes. Using such a metric for decision making requires adapting the experiment design so it captures this long term impact.

Examples:

  • true_conversion_rate (after cancellation and returns have been processed)
  • customer_lifetime_value
  • long_term_revenue
  • retention_rate (over 6 months or more)

Functionality

Finally metrics can also be described by how they operate in the context of the experiment.

Proxy

Proxy metrics are indirect measures used to evaluate an outcome that cannot be measured directly (see for the example of long-term metrics above). In experimentation proxy metrics can be used as a replacement for the actual desired goal. There should be a strong correlation between the proxy and the actual goal and this should be validated frequently.

Examples:

  • time_on_site as a proxy for engagement
  • click_on_buy_button as a proxy for conversion

Composite

Composite metrics combine multiple individual metrics into one measure to capture a nuanced view of success. They are often used strategically but can dilute sensitivity. Examples:

  • Overall Evaluation Criterion (OEC) as a weighted combinations of metrics like engagement, revenue, and satisfaction