How Amplitude computes conversions through funnels

This article will help you:

• Familiarize yourself with conversion computations using funnels
• Identify key differences between Funnel and Event Segmentation charts

When calculating conversion for a funnel in which users can complete the steps more than once, Amplitude will bucket each user based on the values tied to the first occurrence of each event.

How this first touch is determined in your analysis will depend on whether you opt to hold a property constant or break down by specific properties.

Understanding these implications is vital in drawing accurate conclusions from your analyses.

First-touch attribution scenarios

Let's assume we have a funnel to track registrations (complete registration) broken down by the landing page each user sees (view landing page). If we hold constant by session_id, users must complete both steps of the conversion process in the same session in order for Amplitude to count them as converted.

Scenario 1: Funnels using both Hold property constant and group conversions by 

When a funnel analysis uses both the hold constant and broken down by functions, Amplitude bases conversion on the earliest entry within the session.

Consider the following events and conversion results as examples:

NOTE: When holding constant by session ID, your chart will display the number of user sessions that included a conversion, and not the number of users to convert.

Scenario 2: Funnels using group conversions by, but not Hold property constant

In cases where only the broken down by function is used, Amplitude bases conversion on the earliest entry within the lookback window.

Users are grouped by the first landing page they saw within the lookback window (how they entered the funnel). They'll be considered converted if they trigger the final event within the duration of the conversion window.

For example, let's consider the following events and related conversions:

The logic of unique user counts in funnel analyses

When counting by unique users, the baseline conditions for conversion are:

• A user must be eligible for inclusion into the funnel: The user cannot be filtered out from consideration via the user segmentation panel. Any filters set in the Segment by module only apply at the time that the user triggered the first funnel event.
• A user must enter the funnel and complete all steps in the conversion window: The user must enter the funnel, and complete all steps of the funnel within the stated conversion window to count as converted in the final funnel step. Otherwise, the user will be counted based on how far they progressed through the funnel.

Further, when Amplitude is counting by uniques, it will only count the earliest and longest conversion for each unique user:

• Longest: In this context, longest means the most complete conversion; i.e., the completion of the most required steps within the funnel.

If Amplitude finds multiple conversions meeting the longest definition, it selects the first one and counts that as when conversion occurred.

• Earliest: Amplitude measures earliest using the first converting sequence chronologically, if there is more than one.

When using the broken down by function, Amplitude continues to use the longest/earliest logic to bucket users in accordance with the property that was present at their point of funnel entry.

If you're using Hold constant by in your analysis, Amplitude will look for the longest/earliest converting sequence within the same user session. When doing so, the unit of measurement changes to unique user and session ID pairings.

However, when counting by event totals, the earliest/longest logic does not apply. Instead, Amplitude considers all conversion paths taken or attempted, rather than just the earliest/longest path per user. The paths are then attributed to the property for the event in the step it was broken down by.

Funnels versus event segmentation

The Funnel and Event Segmentation charts provide different types of analyses, and as such, could display differing results. The following table highlights some of those differences: