Metrics Glossary
Definitions
Attributables
Rec Click to Purchase Attribution Metrics
Rec Click to Purchase (RCP) attribution refers to a sequence of events where a single customer is shown a product in a recommendation strategy placement, Clicks on that product in that placement and Purchases that product all within a time period of 14 days or less (depending on site configuration.)
Base Definitions (names come from sales report)
Attributable Sales: Total sales of RCP attributable products
Attributable-Order: An order that contains at least one RCP attributable product
Orders from Recs: A count of attributable-orders
Attributable-Order-Items: The number of items (including non-attributable) in an attributable-order
Items from Recs: The count of RCP attributable products, including multiple instances in one Order
Derived Metrics
Rec Sales %: (attributable sales) / sales
Avg. Order Size from Recs ("size" here means "sales"): (attributable sales) / (orders from recs)
Units per Order from Recs: (attributable-order-items) / (orders from recs)
Split Purchase Attribution Metrics among attributable clicks on it
When a purchase is attributed to clicks of different recommendation strategy placements with different context dimensions, for example, strategy, placement, channel, rule etc., each dimension combination gets a partial attribution of purchase metrics at the ratio of each combination to total combinations among these attributable clicks.
For example, on a purchase with $100 in sales, there with four rec clicks that are attributable to it within a time period of 14 days or less as follows:
click1 on placement1, rule1
click2 on placement1, rule1
click3 on placement2, rule2
click4 on placement3, rule2
Then for (placement1, rule1) the attributable sales = $50, (placement2, rule2) the attributable sales = $25, (placement3, rule2) the attributable sales = $25
If we only look for rule associated attributable sales, for rule1 the attributable sales = $50, rule2 the attributable sales = $50 as well.
Metrics Terminology
Base Definitions
User: A user ID. Some user's events may be found by correlation of events with sessions ID and User ID
Visit: A User's events split by gaps of no activity for more than 30 minutes. This is a better entity than Session because it is not device specific: a visit can cross devices.
View: A page view event
Click: A click on a placement event. There can be more than one click per placement or sub-placement entity
Click-in-a-visit: A Click in a Visit that also has its related View
Order: A purchase-view event also known as Purchase
Visit-with-purchase: A Visit that includes a Purchase event
Sales: Aggregated Order prices (see Outlier below)
Units: The count of products, including multiple instances in the same Order
Derived Metrics
CTR (click-through-rate)
Clicks-in-a-Visit/Views: The number of Clicks-in-a-Visit divided by the number of Views in the same visit.
RPV (revenue-per-visit)
Sales/Visits: The value of Sales divided by the number of Visits (see Outlier below)
AOV (average-order-value)
Sales/Orders: The value of Sales divided by the number of Orders (see Outlier below)
Conversion
Visits-with-Purchase/Visits: The number of visits with purchase divided by the total number of visits.
Other
Outlier Threshold: After log conversion, 3 standard deviations above the mean Visit Sales for a site over a 28-day window & will scale down for new sites with less than 28 days of data.
Outlier: A Visit with total Order Sales above the Outlier Threshold. Outlier Visits will have their Sales capped at the Outlier Threshold and each Order in the Visit will have its sale scaled by the same factor that the overall VisitSales were reduced.
The Outlier Threshold window can be a day behind the current day
Currency: A visit is associated with only one currency by looking for it in the visit in the following order:
-
first looking for the currency in the latest purchase view
-
no purchases, then looking for the currency in the latest view
-
no views, then looking for the currency in the latest click
Confidence: Statistical Confidence is calculated as 2-sided test without Bonferroni correction.