— September 20, 2019
The way we do retargeting is restrictive.
Creating a tailored, personalized campaign is often done with micro-triggers: Did the user spend more than X minutes on the site? Did they view more than Y pages? Did they add to cart? Are they visiting on mobile?
All of these data points personalize the messaging for retargeting campaigns.
But even a talented campaign manager can juggle only so many variables. Eventually, you end up with tons of segments that are difficult and time-consuming to manage.
Platforms with built-in machine learning capabilities can make this a problem of the past.
Using (not just talking about) machine learning
Machine learning is a promising technology. But, until recently, many of the purported benefits—especially in paid acquisition—have been discussed but remain largely unrealized.
New tools, however, offer marketers more possibilities without the pain of managing those complicated systems. Google’s machine learning–based metrics, for example, give you access to insights and audience segments you can put to work immediately in your campaigns.
From Google’s Session Quality and Conversion Probability reports to related smart retargeting platforms, machine-learning segments users like no PPC manager can. These tools are available with a price tag that ranges from free to highly affordable.
Why machine learning for retargeting?
Audiences that have shown explicit intent—such as users who added to cart—make up a tiny portion of traffic. Identifying other indicators for intent will help you reach more ready-to-buy people with a CPA that’s similar to your best-performing audiences.
The question should always be, “How do I find the people who intend on buying but didn’t explicitly show intent?” This is where this technology shines.
Machines don’t mind handling a ton of variables. A computer can easily point out (after a short learning period) that, on your website, users who visit between 8 and 10 p.m. on a mobile phone and visit at least 3 pages per session are more likely to convert.
The process of predicting future behavior based on a variety of given variables is called propensity modeling. The outcome of a propensity model is a propensity score, which rates the likelihood of a conversion.
Different tools use different propensity models to predict the likelihood of conversion. The tools mentioned in this article have unique approaches to predict future outcomes given a set of user behaviors.
The best way to see whether (and how) these models actually work is to try them out yourself.
Google Session Quality (works only on Google platforms)
Last year, Google quietly released two interesting Analytics features—Session Quality and Conversion Probability. Both give marketers access to machine learning–based analysis and retargeting on Google’s stack.
Session Quality and Conversion Probability split your traffic into segments on a score from 1 to 100 based on the likelihood that they’ll convert. They provide a Dimension as well as an average Session Quality/Conversion Probability metric, which makes them versatile.
The features are similar. (Google even uses pretty much the same explanation for both.) Session Quality has been around longer, and I’ve tested its performance many times. In my experience, Session Quality provides more consistent results, so we’ll focus on it for the rest of the post, but you should try both to see which works best for you.
Google uses machine learning to compute Session Quality based on many (undisclosed) factors. One thing is clear by looking at the report: The score correlates with the likelihood of purchasing.
You can export these segments into Google Ads for remarketing campaigns. And, since Session Quality is also a Dimension, you can view different metrics and split them by Session Quality to find correlations and analyze your activities further.
Let’s take a look at a test our agency ran to see this metric’s power in action:
In the above screenshot, you can see the summary data for a campaign that ran for four months. The top section shows the results for campaigns that ran with a Session Quality–based audience. The lower table includes a summary of all other campaigns.
Over the four-month period, Session Quality campaigns outperformed other retargeting campaigns, generating a Return on Ad Spend (ROAS) three times higher, as well as CPAs 65% lower.
But Google being Google, you can use these features only under specific conditions.
Limitations to Google’s Session Quality/Conversion Probability
First off, you can’t even view these dimensions/metrics unless you have more than 1,000 transactions per month. These are ecommerce only, transaction-based features.
If your business doesn’t generate at least 1,000 transactions per month or doesn’t have Enhanced Ecommerce in place, Google’s features aren’t the right solution.
Moreover, you can use Google’s segments only on Google’s platforms (e.g., Google Display Network, YouTube)—no smart Facebook/Linkedin/Twitter ads.
So what are your other options?
Smart retargeting platforms
I’ve recently tested Fixel, a platform that segments your audience similarly to Google’s Session Quality, with some minor differences. Fixel segments your audience into three categories:
The platform excludes the low-quality audience (similar to Google’s “1” score), so even “Basic” is pretty damn effective.
Fixel generates events, meaning you’re able to segment each audience on Facebook (by targeting users who’ve fired Fixel’s High/Medium/Basic events), Analytics (same idea), Linkedin, and even Native platforms like Taboola and Outbrain.
- The algorithm. How each of these platforms segments users and determines their quality.
- Pricing. The more features a tool has, the more it costs, and pricing models vary. Fixel has a set price based on website traffic; Criteo has a CPA model where you set a target CPA, and they “earn” the margin.
- Ease of implementation. Some tools require a complete setup to define page types and click events. Others are more “plug and play.”
- Features. Unlike Fixel and Google’s Session Quality metric, platforms like Criteo offer other features, like cross-platform dynamic retargeting capabilities, which, of course, affect pricing.
What’s interesting about both Google’s Session Quality and Fixel is that they’re unobtrusive—all they do is fire an event during sessions, giving users a score.
Other retargeting tools offer an all-in-one solution, which requires you to move your workflow, as well as assets, to a new platform. That means they’re harder to test head-to-head against the original campaign.
In this campaign, we created seven retargeting audiences. Four were Fixel audiences. Comparing Fixel’s 14 day + High audience with a generic retargeting + 14 days audience, Fixel’s CPA was 70% lower. It was also 56.8% lower than the average CPA for all ad sets.
Find top-performing channels based on quality scores
Machine-learning segmentation can also help you optimize your marketing strategy. Instead of using these tools strictly to improve conversions, consider using them to find out which channels bring the best traffic to your site.
You can use the score as an indicator that predicts the success of each channel and identifies the best audiences. This is especially valuable if your user journey is long and you can’t gather much insight from analyzing purchases/conversions, or if you want to learn more about the performance of top-of-funnel efforts.
This is even more relevant to content marketing, where achieving the end goal takes time, and you’re paying to bring back users who may or may not make a purchase.
Here’s how to do it:
- Create a Custom Report in Google Analytics that splits your traffic by source.
- Add Average Session Quality/Fixel High event as the metric.
- Compare channels to see which generates the most high-quality traffic.
For the report above, compare Paid Search and Display. Both have a similar ecommerce conversion rate. But Display’s average Session Quality is way higher than Search, even though Search resulted in more transactions. Session Quality—as a predictive tool—seems way off.
But take a look at the Revenue column. Display generated approximately 30 times the revenue of Search. If you’re optimizing solely for conversion rate—or any single metric—you may miss what machine learning’s more complex calculation takes into account. That can be the difference between hitting vanity metrics versus revenue goals.
Smart retargeting campaigns vs. behavioral segmentation
Our agency tested machine learning–driven retargeting tools on about 30 clients, and the following statements apply to at least 85% of tests:
- These platforms generate CPAs that more closely resemble the results of “Add to Cart” than any other retargeting tactic.
- By simply excluding users with Session Quality <2, or targeting Fixel Basic (which excludes retargeting audiences that Fixel did not “catch”), you’ll narrow the retargeting audience by approximately 60% and reduce CPA by at least 30%. As shown above, you may see up to 60–70%).
- Users who scored low rarely convert.
To test how these metrics can expand your high-quality audience, simply create an A/B/C test of your audience with the following setup:
- Separate the machine learning–derived audience to its own campaign on your ad platform of choice, and exclude people who have shown explicit intent (added items to cart, for example) from that audience.
- Create another campaign and target the rest of your site traffic, excluding the machine-learning audience and cart adds.
- Create a third campaign targeting people who have performed the explicit-intent action (e.g., added to cart) exclusively.
By running these 3 campaigns—with the same budget—you’ll be able to compare the CPAs of these segments.
- How does the machine-learning audience compare to your (presumably) best-performing segment?
- How much better does it do than the “generic” audience?
In the example below, the machine-learning audience was—remarkably—nearly as efficient as the “Add to Cart” audience.
You can now decide if you wish to exclude the generic audience or lower its budget. You can also run additional tests against your top hand-crafted segments to see how the machine-learning audiences perform against your own work.
Machine learning to automate campaign optimization
While Fixel, Google, Criteo, and the like use machine learning to segment your visitors based on engagement, Kenshoo, Acquisio, and similar tools use AI to optimize campaign performance based on data from the advertising platforms themselves.
In other words, these platforms use machine learning to analyze advertising platform data, rather than behavioral user data. As a result, their main focus is on campaign management automation—removing the need for manual optimization.
The effectiveness of these platforms is harder to test since, to use them, you have to transfer your workflow to a new system. The number of variables (including time saved/wasted for the platform’s learning curve) make it hard to measure their value in the same way we’ve measured smart retargeting tools.
At the same time, paid campaigns are a lot of work. And, regardless of the tool, the idea is the same—you’re leaving audience segmentation to the algorithm to focus more on strategy and acquiring high-quality users.
Marketing is fertile ground for AI and machine learning, and the use cases already exist. By implementing one of the above solutions, you’ll be able to harvest AI benefits today.
- Google’s Session Quality or Fixel’s scoring help you identify audiences with high intent even when users’ on-site actions don’t reveal it.
- Testing machine-learning audiences against explicit-intent users, your best hand-picked audiences, and your general audience shows the potential value of machine learning for your campaigns.