...

• Source → Results tell you what is the result, but not why you have it and what to do to improve it
  For example: My return on ad spend is very law low for one of my key brands and therefore the Google algorithm doesn’t allocate much budget to it. What should I do?
  

• Behaviors → Results are hard to put into context and it is quickly hard to know if you are addressing a situation which can be improved
  For example: My mobile conversion rate is much lower than my desktop conversion rate and I invested months of work to try to improve it until realizing this is the case for most e-shops

...

Step 3: The Data-Driven Hypothesis: Collect & Conclude

Here the idea is to share the analytics internally and to collect a poll / form to answer two types of questions:

...

what look like the biggest issues

...

If everything went according to plan, as a result of Step 2, you have a lot of information about not only the performance of your activities, but about key aspects of the Customer Journey which correlates to good or bad results.

However, this is usually not enough to act which we will describe on step 4.

There is a crucial step before which, as much as possible, we recommend to address in a collective way.

Looking at the reports will hopefully give many ideas to each key person of your organization about what needs to be done, but not everyone will necessarily conclude the same things or at least the same order of priorities.

We therefore recommend to share these reports within your team (possibly after an initial presentation) and to collect their feed-back in a structured way.

You can do it in a simple manner at first (create a simple poll to decide on a 1 or 2 first changes), but we recommend to do it in a complete and structured way over time, as described in Step 5.

However, this step is important, because it will ensure that a collection of relevant optimization candidates are well defined, which will avoid the typical issue of identifying only 1 idea and to decide immediately to act on it without considering at least a list of 3-5 (or more).

Use Case: Google Shopping Optimization

As a result a Step 2, a collection of candidates can be defined, some of which might be in the following list, but don’t take the list as a collection of hot-tip, because the entire goal of the present methodology is to avoid taking hot-tips that your “gut feeling” tells you ought to be true, but instead to detect them in a data-driven way:

1. Structure your campaigns based not directly on the margin of the advertised products (proxy products) but on the revealed margin generated by visitors clicking on these products
   

2. Structure your campaigns based products which are good at acquiring new customers or at reacquiring elapsed customers (or generating higher lifetime values)
   

3. Structure your campaigns based products which do not frequently cause a multitude of back and fourth between your e-shop and Google Shopping during the same visits (and for which you pay for the ad click every time)
   

4. Structure your campaigns based on products which have a stock level higher than their daily purchases to avoid visitors seeing long deliveries often
   

5. Structure your campaigns based on products which have a discount % level which is in the sweet-spot of your conversion rate
   

6. Structure your campaigns based on products which have currently a positive marker on a tracked event on the PDP (good ratings, multiple pictures often all viewed, high add to wish-list rate, interest shown in reading their description, related content often clicked, …)
   

7. And of course (as we will discuss more in Step 4): A/B Testing of a visual changes on the PDP itself (based on the findings of any of the tracked event)

Step 4: Targeted Testing: Data, Process & Visual

Here we speak of the set-up of a targeted testing a change which could be of the following types

...

As a result of Step 3, most of your optimization ideas are likely to belong in one of 3 groups:

• Data
  These are the best kind (at least for data scientists). Basically it means that the change you want to test requires no change in your processes (IT or otherwise) and can be done behind the scene.
  This is the case for example of any algorithmic or rule-based change in the sorting / selection of products in product listing or product recommendations. The change simply modifies what product will appear in which context and to whom.
  The typical Data change consists of an data processing in BigQuery which results in pushing data to the Boxalino Lab (which are then automatically uploaded in Boxalino Real-Time Platform) and then to configure the Boxalino Admin to make usage of these new data, typically in an A/B Test.
  

• Process
  we are changing/improving the management of our e-shop based on new analytics
  example: we are improving our stock management process based on the information of the most viewed product with non optimal delivery times
  

• Visual
  we are chaning changing what the user sees on the web-site
  example: We are making a visual change on the page to show similar recommendations higher on the page
  

• Hybrid
  This is a very common case typically when both Data and Process or Data and Visual are both needed together in a change, or even all 3 together
  example: create margin groups to change the campaigns of Google Ads and change the source
  A data change is required, but also a structing of your Campaigns which is a change in your processes
  

About Targeted Testing:

• Testing
  If possible, we do the change as a test (if possible an A/B Test) to have a direct causal understanding of the effect of the change
  

• Targeted
  we are doing the testing in a targeted way if possible, which might mean “personalized” either individually or in a customer segment but can also mean we are implementing the change a segment of the product sortiment

...