GA4 brings machine learning to the forefront

As we covered here, Google Analytics 4 has a few new features and structures that set it apart from Universal Analytics. One of the most important structural aspects is GA4’s flat data model, which makes it perfect for machine learning.

As part of our ongoing video collaboration with Romina Henrizti of Bluebird Media, we wanted to get the low down on some of the most impactful ways that machine learning in Google Analytics 4 can help marketers. 

New sophisticated insights

While you were able to take advantage of insights already in Universal Analytics, the options for custom insights and automated insights available to you in Google Analytics 4 are much more sophisticated and prominent. They have the power to highlight any web performance statistic that you may be interested in.

For example, the system can spotlight if a page is loading unusually slowly, or you experience a significant uptick in traffic from a lead source. Basically, the system loves to point out anomalies in your website’s performance. 

Now, you may be saying, “Hey, wait a second. I already had alerts in UA.”

You’re not wrong. You could set up all sorts of custom alerts like when conversion dropped by 20 percent, or if more than 50 percent of users came from direct traffic. 

The difference in GA4 is that the underlying AI automatically detects these anomalies, using predictive metrics to recommend actions based on insights, without the need for you to set them up. It’s almost like your secret assistant constantly analyzing web performance and notifying you when anything looks outside of the norm. 

How to handle anomalies

In addition to simply highlighting these anomalies, the algorithm can also provide you with recommendations of how to respond to these insights. To do this, the system runs your data through a few potential scenarios and recommends the course of action with the greatest positive result. 

For instance, the algorithm can make recommendations on Google Ad spend to optimize your return on investment. 

Pretty cool, right? 

Predictive metrics

So, if Google Analytics 4 can analyze your data and recommend a course of action, can it sort of see into the future?

Well, there aren’t any crystal balls built into the system (that we know of). However, Google Analytics 4 does provide you with three types of predictive metrics, which can achieve almost the same effect of foresight. 

These predictive metrics (churn probability, purchase probability, and predicted revenue) are powerful because they identify users likely to take specific actions. The real magic lies in using these predictive metrics to build predictive audiences that are relevant for your business. These audiences group users based on their predicted behavior, allowing you to tailor your marketing messages for maximum impact.

Churn probability

As the name suggests, this predictive metric analyzes and defines how likely it is that a user will disengage with your site. Specifically, this predicts whether a user who has been active in the last seven days will cease to be active in the upcoming seven days. This can later influence retargeting efforts and other retention marketing efforts. 

Purchase probability

No beating around the bush with this one. This measure looks at users who were active in the last 28 days and determines how likely they are to make a purchase on your site in the next seven days. This can be invaluable for e-commerce sites looking to forecast short-term sales. 

Predicted revenue

So, you may have an idea of how many sales will occur in the next week, but are those small- or high-value sales? That’s where predictive revenue comes in. 

The real power underlying each of these three predictive metrics is their use of predictive audiences. Some of these come pre-packaged in Google Analytics including likely seven-day churning purchasers, likely first time seven-day purchasers, and more. You’re also able to build your own predictive audiences that can be tailored a bit more tightly to your own business. 

Powerful Google Analytics 4 audiences

These audiences that enable you to predict user behavior don’t end at the borders of GA4. They can also be shared with other Google products like Google Ads or video 360. 

By sharing the audiences across these products, you can begin to supercharge your marketing strategies. Let’s explore an example together. 

Imagine your popular streetwear clothing brand is looking to run a promotion. You can build a custom audience that includes likely churning users that have been visiting pages that relate to your brand’s apparel. This audience can be found using predictive metrics. Then we can set this audience aside for targeting in Google Ads with the promotion, since they likely won’t return to your website directly. 

You can also create a custom audience that excludes users. For instance, you may want to exclude an audience that is already likely to purchase something on your website without seeing the promotion. This means you won’t pay for those impressions or clicks, and you may likely get those users to pay full price - assuming they don’t jump into the sale section on your website, of course. 

The audiences can get pretty sophisticated, which gives digital marketers loads of options for performance improvement. However, there are some minimum requirements that you’ll need to meet to access this powerful tool. 

A high bar to meet

You need at least 1,000 positive and negative users for each metric over seven days. This means that you need at least 1,000 purchases and 1,000 users that didn’t buy anything in a week. And, this only works with the purchase event at the moment. 

This all means that Google is currently reserving this powerful tool for online retailers that see a fair amount of volume. 

Data-driven attribution

We previously touched on Google Analytics 4’s data-driven attribution techniques here. After digging into it in detail herself, Romina wasn’t too impressed by the difference. 

In her own words, it was “less spectacular than I thought.”

In her analysis, she found that the attribution keyed in on paid shopping ads as the clear winner with attribution rising from 10 to 13 percent. Perhaps it’s not so surprising that another Google advertising product was favored in the new GA4 attribution model. 

However, the attribution model was too blatantly skewed as even the YouTube and display ads channels for Google were on the losing side of the attribution equation. 

This outcome appeared across the board for Bluebird’s client base regardless of industry type, size, media mix, etc. In Romina’s opinion, things got a bit more interesting when analyzing clients with a wider media mix, but the Google Ad channels still won the lion’s share of attribution. 

“It’s not that surprising if you think about it,” said Romina. “[Those channels] usually have a bigger revenue share to begin with, and they have a lot of touch points on the user’s journey.”

Google Ads may also receive a boost in the attribution model, since it can share the underlying audience and user data from GA4. 

“That’s why our tip is to provide Google Analytics 4 with the essential marketing performance data like impressions, clicks, and costs from all of your non-Google channels,” said Romina. That way, GA4 can get a more holistic view alongside the predictive metrics it generates.

What’s so great about behavioral modeling?

In addition to the sophisticated new insights, predictive metrics, and audiences, Google Analytics 4 has also introduced behavioral marketing tools. This functionality is particularly useful for users in markets with strict privacy laws - like Europe, Japan, and some parts of the US. 

With privacy laws in place, digital marketers in these regions must gain consent before they can start tracking user behavior. And when users opt out of tracking entirely, that can make things a bit more ambiguous for marketers who are used to having a plethora of user behavior data at their fingertips. 

To better help marketers and web analysts in this age of data privacy, Google has launched a new technique called “consent mode.” If a user does not give consent for tracking, Google will send anonymous pings to a third server. These pings are similar to the basic information that every web server collects about users. 

Google then combines this anonymous, basic data with that of the users who did provide consent in a machine learning algorithm. The AI then models the anonymous behavior based on the known, consented user data. 

Essentially, it learns all about the behaviors of the consenting users. Then, it applies its learnings to the anonymous users and makes informed assumptions about how they likely viewed and used your site. 

More requirements for such powerful tools

Much like the audiences feature, you’ll need to meet some minimum requirements in order to use the behavioral modeling tools. You need at least 1,000 non-consent events per day, along with at least 1,000 users that gave consent. 

Even if you meet these requirements, Google points out that the volume may still not be large enough for the behavioral modeling to be as accurate as possible. Remember, machine learning likes large (make that HUGE) sets of data to learn from. 

Google Analytics 4 machine learning at a glance

Google Analytics 4 machine learning has a high barrier to entry for certain features due to data volume requirements, but the ability to leverage predictive metrics is a game-changer for marketers with sufficient data. If you have the volume, these tools can almost help you see the future. 

Well, you can make informed predictions, at least. The more data the better. 

When applied to marketing initiatives, we should be able to expect improved return on ad spends and better conversion rates overall. After all, machine learning will help us better understand past user behaviors, as well as how they will respond in the near term. 

It’s also important to remember that Google Analytics 4 is constantly being improved. This is just the ground floor, and we should see more advanced predictive metrics and even greater machine learning capabilities as Google refines and evolves the platform. 

FAQ

What’s the difference between predictive metrics and predictive analytics? 

Predictive metrics and predictive analytics are closely related in the world of marketing and data analysis, but there is a subtle difference between them. Here's a breakdown:

Predictive metrics:

• Specific data points that forecast a user's future behavior.
• Think of them as pre-built predictions based on machine learning algorithms.
• Examples in GA4 include churn probability, purchase probability, and predicted revenue.
• Essentially, they answer the question: What is likely to happen to this user?

Predictive analytics:

• The broader process of using machine learning and statistical techniques to uncover patterns and predict future trends.
• Involves analyzing large datasets of user behavior and website interactions.
• Generates the insights that feed into the creation of predictive metrics.
  It's the engine that powers the predictions, while predictive metrics are the specific outputs.
  
  

So, to put it another way: imagine you're a baker…

Predictive metrics would be like pre-made packets of cookie dough that tell you how many cookies you'll get when it’s mixed up (IE: makes 12 chocolate chip cookies).

Predictive analytics would be the entire baking process: mixing flour, sugar, eggs, and using your knowledge of baking to predict how many cookies you'll get based on the recipe and ingredients.

How do I use purchase probability in my marketing strategy?

There are a few ways you can leverage purchase probability:

• Targeted advertising: Use GA4 to create custom audiences based on purchase probability. You can then target these audiences with specific ad campaigns tailored to their buying intent. For example, you might target users with a high purchase probability with retargeting ads featuring special offers or discounts.
• Improved product recommendations: Leverage purchase probability to recommend products to users who are most likely to be interested in them. This can be done through on-site product recommendations or personalized emails.
• Inventory management: Use purchase probability data to anticipate demand for specific products and optimize your inventory levels. This can help you avoid stockouts and ensure you have enough of the right products in stock to meet customer needs.

How can I use churn probability in my marketing strategy?

Churn probability can be used in a few useful ways:

• Win-back campaigns: You can identify users with a high churn probability and target them with special offers or incentives to keep them engaged. This could involve personalized email campaigns, retargeting ads, or loyalty program promotions.
• Improved user experience: Try using churn probability data to understand why users might be disengaging and identify areas for improvement on your website. This could involve optimizing the user experience, addressing usability issues, or providing better customer support.
• Proactive engagement: Reach out to users with a high churn probability through personalized messages or in-app notifications. You can offer them helpful resources, address potential concerns, or remind them of the value your product or service offers.