How to Maximize Ecommerce Revenue Growth Using

Olcan Sercinoglu

Oct 03, 2018


In previous posts, we introduced Autonomous Optimization (AO) and described how to design and implement effective AO projects. In this post, we dig into a specific use case where we seek to maximize ecommerce revenue growth.

As a reminder, a well-designed AO project should be able to achieve its optimal metrics (in this case revenue) within 2-4 weeks and can be expected to deliver a significant improvement (+10-20% or more) during this period.

Introduction: A Simple Online Store

Our simple online store consists of our home page, product pages, cart page, checkout page, and thank you page. We can illustrate this as follows, with the edges indicating the ideal progression of a visitor during a visit.


Designing Metrics

In this use case, our goal is to maximize revenue per visit. To achieve this, we need each session to represent a visit and contain the key events that we need to calculate revenue (and possibly other metrics) for that visit.


An important design detail is how much time is allowed between events that are considered part of the same visit. In this use case we follow a common standard, which is to allow up to 30 minutes from the last event (e.g. between page views) and up to 24 hours from the first event (AmpSession) in the session.

Designing Fast Metrics

The first design consideration is to have fast metrics that are determined faster/earlier. This is not a concern in this design since the 24-hour limit on session length (mentioned above) ensures that all metrics are determined within 24 hours.

Designing High-Signal Metrics

The second design consideration is to have high-signal metrics that vary more across sessions and thus provide more “signal” (high or low metrics) for faster learning and optimization. For this purpose, it is helpful to write down a representative set of sessions and their metrics.


We see clearly that revenue alone is a low-signal metric, because most visits would generate zero revenue and look exactly the same to the AO system, even though we know intuitively that those sessions are not the same. Now consider instead the following set of metrics that are clearly more informative across sessions.


Designing Optimization Priorities

Optimization priorities must naturally reflect our intuition and domain knowledge about the ability of each metric to drive our key metric revenue, which itself should always have the highest priority.

For non-revenue metrics, the key consideration is this: if the AO system is unable to optimize directly for revenue, which other metric should it optimize for instead? In other words, which other metric would best capture an increase in buy intent?

Although the answer to this question can depend on the store and the available variants (see next section), we recommend that metric priorities generally follow (inversely) the depth of the relevant events in the purchase funnel.


Designing Variants

In this use case, we consider four variants (including baseline) of our home page. We denote these variants using four colors: Gray (Baseline), Red, Blue, and Green.


These variants could represent any changes to our home page, but as a concrete example, we consider four variants of the hero section (image and text) that would focus on different products or features and thus appeal to people who would care most about those products or features.

Designing High-Impact Variants

The first design consideration is ensuring high impact relative to the baseline variant. Even if a variant has low overall impact, can extract higher impact from it by applying it only on a smaller segment where it has higher impact.


For example, if we design our home page variants to emphasize products or features that each appeal strongly to 20% of our visitors and make them 50% more likely to buy, then (assuming same base revenue as the remaining 40%) we could achieve +30% overall revenue impact, even though each individual variant may have ~zero or even negative overall impact.


In general, we seek to design variants that have high impact on segments that are not-too-small and not-too-overlapping. We illustrate these criteria below.


The challenge in designing high-impact variants is that we usually do not know the size or overlap of the corresponding high-impact segments. Indeed, in most cases the best segments can only be determined from data. Still, designing based on hypotheses that incorporate our domain knowledge, intuition, and learnings (including from past AO projects) is well worth the effort as it will lead to much faster growth over time.

Designing Decision Point and Context for High-Impact Segments

The second design consideration is to ensure that the context available at the decision point helps discern the high-impact segments for our variants, enabling to zero in on these segments to maximize overall impact.


With the above design, the hero section content is decided on the home page, just before it is used, and is likely to impact the session (by getting noticed by the customer). This helps avoid unnecessary/unused decisions that would otherwise need to be filtered out to speed up learning and optimization.

The context available at the decision point depends on which events are observed before the decision. Additional relevant context can be introduced simply by observing a new event such as MyContext.


Below are some example context variables that are often (depending on the variant) helpful in discerning high-impact segments:

Implementing using

Implementation is easy and consists of the following steps:


In future posts we will look at price optimization, offer optimization, push-button integration with major ecommerce platforms such as Shopify, and several other ecommerce related topics. We will also continue to explore many other use cases across the customer funnel and software stack.