By Samantha Kwok

Those keeping close tabs on the latest Google Analytics updates will have no doubt seen a new report appear in the past few months – Conversion Probability. So, what does it tell us and how should we use it?

Back to blog home

Tell me more about this new report..

The Conversion Probability report uses machine learning techniques to determine a user’s likelihood to convert in the next 30 days. Whilst it’s perfectly possible to model GA / CRM data to create your own propensity score using machine learning, Google’s done most of the legwork with this addition.

Here, 'conversion' specifically means completing an ecommerce transaction, so unfortunately if you don’t have ecommerce tracking switched on, you won’t be able to see this report and it’s not available for non-transaction conversions either. For those lucky ones with ecommerce tracking, you can head to Audiences > Behaviour > Conversion Probability to find the report.

What’s the maths behind this?

As with all things Google, this new report is part of their machine learning ‘black box’ so we don’t actually know the exact methodology used - something worth keeping in mind when using the report.

Put simply, the model takes the last 30 days of GA data and evaluates transactions for each user, resulting in a score that ranges from 1-100. This score indicates the probability of conversion for that user (1 being extremely unlikely to convert and 100 extremely likely to convert), which are categorised into ranges, shown as a new dimension: % Conversion Probability. A new metric is also available in this report – Average Conversion Probability – which is calculated across all users in relation to the date range and dimensions(s) used.

ga-conversion-probability

So… what can I do with the data from this report?

The report enables segments to be easily built using the % Conversion Probability dimension – just two clicks and it’s created a segment based on the conversion probability range. These segments can be applied to reports throughout GA and used in creating audiences too. 

Alternatively, users can create their own segments using the % Conversion Probability dimension along with any dimensions. Again, these can be applied to reports through GA which can be quite insightful, as it’ll help you understand a range of things such as:

  • How many users fall into this segment and how does this compare to the overall user base?
  • Which channels and campaigns drive users who are highly likely to convert (and vice versa)?
  • What are the key characteristics of users who are highly likely to convert (and vice versa)?
  • Is there specific content or pages that users highly likely to convert are consuming?
  • What products are those highly likely to convert purchasing / browsing?

From a marketing and product perspective, there are some clear actions that can be taken:

Marketing

Product

Create remarketing audiences based on users highly likely to convert > activate using platforms such as Google Ads / DBM

Create audiences to send to Optimize to understand where site changes can be made to optimise for conversion based on audiences likely to convert vs. audiences unlikely to convert

Use data from report to optimise channels/campaigns based on where highly likely to convert users are landing on site from

 

How accurate are the predictions?

It’s still early days so the effectiveness of using the data in remarketing is still up for debate. We’ve run a few spot checks across some of our clients to understand the differences between actual conversion rate and predicted conversion rate (i.e. Avg Conversion Probability).

The key thing we’ve noticed? There tends to be less of a difference between predicted and actual on sites with large volumes of transactions. However, we’ve also noticed that some regions, such as Germany, show a bigger difference even if they have a large volume of transactions due to the nature of ad blocking in that particular market. It’s possible that due to a higher prevalence of ad blockers and cookie deletion behaviours in some regions, this results in the inability to track a user across the full 30 days and therefore results in a less accurate predicted value.

The best way to know how effective this data is, is to use it in media campaigns – we recently ran a test for an ecommerce client and saw that over the course of a month, campaigns targeting audiences with a conversion probability of greater than 50% saw an extremely successful ROAS of 126, compared to the average ROAS of 1.2. Additionally, revenue generated during that month saw an increase of +47% compared to the previous month. 

You can also consider testing the effectiveness of the model by using experiments in Google Ads or A/B testing in Display & Video 360 (DoubleClick) to identify if there’s any incremental value driven by audience lists defined by the conversion probability data.

As with all Google products, this report will no doubt continue to update and improve overtime, so watch this space.

Anything else I need to know?

Yep, there’s a few things you should be aware of:

  • The report’s still in beta, so it’s slowly rolling out across all accounts which meet the prerequisites; it might not have appeared just yet, so keep an eye out.
  • For the report to appear, you’ll need to have ecommerce tracking implemented and a minimum of 1000 transactions per month with at least 30 days of data.
  • If Google deem that the machine learning model behind the report doesn’t meet their accuracy target, the report won’t appear until it does.

If you’d like to find out more about the implementation and strategy behind propensity scoring, get in touch with the Audiences team at Merkle | Periscopix.

Share this article