Metric Examples and Use Cases

Examples of different metrics you can define within GrowthBook

Fact Tables in GrowthBook are a powerful way to quickly build out a library of metrics.

Below are some common types of tables in your data warehouse and how to map them to Fact Tables with example SQL and metric definitions.

The Add to GrowthBook buttons will open GrowthBook and let you customize the metric before adding it to your account.

Raw Events Table

The most common form of analytics data is a raw event stream. Every time a user does something, you append a new row to an events table in your warehouse. All types of events from all users are intermixed in this one table.

Fact Tables in GrowthBook work really well with raw events. When defining your Fact Table in GrowthBook, there are 4 important types of columns you will want to select:

1. Timestamp - When the event occurred
2. Ids - Who triggered the event (user id, device id, etc.)
3. Event Type - What action was performed (page view, start checkout, add friend, etc.)
4. Context - Any helpful user or app context (geo-location, current url, account plan, etc.)

When creating metrics, you will almost always want to filter on the Event Type. To make this easier, you can edit the column representing Event Type and check the “Prompt all metrics” box (after creating your Fact Table). This helps GrowthBook better streamline the UI when creating metrics.

Example SQL: Google Analytics 4 (GA4) Events

When you configure Google Analytics 4 to sync events into BigQuery, they get written to a raw events table (actually multiple tables partitioned by day, but that’s more of an implementation detail)

Here’s example SQL for a fact table. Comments added to highlight the 5 column types above

SELECT
  -- Timestamp
  TIMESTAMP_MICROS(event_timestamp) as timestamp,
  -- Ids
  user_id,
  user_pseudo_id as anonymous_id,
  -- Event type
  event_name,
  -- Context
  event_value, -- e.g. the page view path
  geo.country,
  device.category as device_category
FROM
  `events_*`

For faster and cheaper queries, we can take advantage of the date partitioning table scheme. GrowthBook lets you use date template variables that are evaluated right before executing the SQL. By doing this, we can restrict our query to just the relevant dates for the experiment we are analyzing.

Change the end of the query and add the following WHERE clause:

FROM
  `events_*`
WHERE
  -- Daily tables
  ((_TABLE_SUFFIX BETWEEN '{{date startDateISO "yyyyMMdd"}}' AND '{{date endDateISO "yyyyMMdd"}}')
  -- Realtime streaming intraday table
  OR (_TABLE_SUFFIX BETWEEN 'intraday_{{date startDateISO "yyyyMMdd"}}' AND 'intraday_{{date endDateISO "yyyyMMdd"}}'))

Example Metrics

Percent of users who submit a form

1. Select Proportion as the metric type
2. Choose form_submit as the event type

Percent of users who visit the checkout page

1. Select Proportion as the metric type
2. Choose page_view as the event type
3. Add a “Row Filter” and select page_path = '/checkout'

Page Views per User

1. Select Mean as the metric type
2. Choose page_view as the event type
3. Select Count of Rows as the value

Unique Page Views per User

1. Select Mean as the metric type
2. Choose page_view as the event type
3. Select event_value as the column
4. Select COUNT DISTINCT as the aggregation

Active Days per User

1. Select Mean as the metric type
2. Choose Distinct Dates as the column

Checkout completion rate

1. Select Ratio as the metric type
2. For the numerator (confirmation)
   1. Choose page_view as the event type
   2. Add a Row Filter for page_path = '/confirmation'
   3. Select Unique Users as the value
3. For the denominator (checkout)
   1. Choose page_view as the event_type
   2. Add a Row Filter for page_path = '/checkout'
   3. Select Unique Users as the value

Modeled Tables

Another common way to store data in a warehouse is with modelling. This usually involves a tool like dbt that transforms raw event logs into multiple well-structured tables. There are also some event tracking tools like Segment which do this automatically.

For our purposes, we are only considering modeled tables that do not include a GROUP BY in the data pipeline - in other words, the rows are still at the raw event-level. Check out the Pre-aggregated Tables section below if this is not the case.

In GrowthBook, we recommend creating a separate Fact Table for each one of your modeled tables. At the very least, you will need a Timestamp and IDs of who performed the action. Any additional columns you add are optional and can be used to build out metrics.

Example SQL: E-Commerce Orders Table

SELECT
  -- Timestamp
  received_at as timestamp,
  -- IDs
  user_id,
  -- Other custom columns
  num_items,
  grand_total,
  shipping_type
FROM orders

Because this is a modeled table just for orders, we can select event-specific columns like num_items, grand_total, and shipping_type. These columns will help us build out a rich set of metrics. This is one of the big benefits over using a raw events table directly where the columns you select need to be more generic and work across a wide range of different events.

Similar to raw event tables, you can optionally add a WHERE clause on dates for faster/cheaper queries:

FROM orders
WHERE
  timestamp BETWEEN '{{startDate}}' AND '{{endDate}}'

Example Metrics

Users who Purchased

1. Select Proportion as the metric type and that’s it!

Users who Purchased Multiple Times

1. Select Proportion as the metric type
2. Add a “User Filter” based on Count of Rows
3. Enter >= 2 as the User Filter value

Orders per User

1. Select Mean as the metric type
2. Select Count of Rows as the metric value

Revenue per User

1. Select Mean as the metric type
2. Select SUM(grand_total) as the metric value

Average Items per Order

1. Select Ratio as the metric type
2. For the numerator, select SUM(num_items)
3. For the denominator, select Count of Rows

Users Without Orders

1. Select Proportion as the metric type
2. Add a “User Filter” for Count of Rows = 0
3. Under Advanced Settings -> Display Settings, change the goal to Decrease the Metric Value

Percent of Orders with Free Shipping

1. Select Ratio as the metric type
2. For the numerator, add a row filter for shipping_type = 'free'
3. Select Count of Rows for both the numerator and denominator

Pre-Aggregated Tables

Pre-aggregated tables include a GROUP BY in the data pipeline to compress raw event-level data down to fewer rows. This is usually done when querying raw data directly is prohibitively expensive.

GrowthBook supports pre-aggregated tables as long as they satisfy two requirements:

• Must be grouped by both user and date
• Pre-aggregated columns can only be basic sums or counts. No averages, percentiles, count distinct, or complex derived formulas that break statistical assumptions.

Using pre-aggregated tables for experimentation comes with additional complexities and downsides, so we highly recommend sticking with event-level data whenever possible.

Example SQL: Social Media Interactions (User/Date Rollup)

If you have a social media app, the number of raw interactions might be huge. A user/date rollup table can drastically reduce the size of the data set. Here is an example Fact Table SQL definition in GrowthBook that references a user/date rollup table.

SELECT
  date as timestamp,
  user_id,
  -- Aggregated columns (sums/counts only)
  interactions,
  likes,
  comments
FROM user_date_rollup

Pre-aggregated tables have a problem. The timestamp column isn’t very accurate. All events for a day are recorded as happening at the beginning of the day (usually at midnight). We only consider Fact Table rows with a timestamp after a user first enters an experiment. That basically means a user’s first day of data will be excluded from the analysis. This is a big deal because those first few hours after entering a test are often the most important.

We can fix this by incrementing all timestamps by 1 day in our Fact Table definition (specific SQL varies by db engine):

SELECT
  DATE_ADD(date, INTERVAL '1 day') as timestamp,
  user_id
  ...

This introduces a new problem however - we will now potentially include actions from before a user saw an experiment, which can introduce noise to the results. In our experience, this is usually the lesser of two evils and thus, this is the approach we recommend.

Of course, you can solve both of these problems by not pre-aggregating in the first place, but we understand sometimes that’s just not an option.

Example Metrics

Percent of Users with High Engagement

1. Select Proportion as the metric type
2. Add a “User Filter” and select SUM(interactions) as the value to filter on
3. Add some threshold as the User Filter value. Something like > 100

Interactions per User

1. Select Mean as the metric type
2. Select SUM(interactions) as the value

P90 Likes per User

1. Select Quantile as the metric type
2. Check the “Group by user” box
3. Select SUM(likes) as the per-user value
4. Select P90 as the quantile level

Comment-to-Like Ratio

1. Select Ratio as the metric type
2. Select SUM(comments) as the numerator value
3. Select SUM(likes) as the denominator value