machine learning, ivan jerkic, the marketing guy, splitx, one ivan

5 ways to leverage machine learning in Facebook campaigns

Reading Time: 8 minutes

Table of Contents

Ivan Jerkić

Ivan Jerkić

I help businesses with my knowledge and experience. I might know a thing or two about technology, marketing, strategy, branding, and growth.

Long gone are the days when we used to adjust bids manually, and when we took care of each keyword, line of text or image individually. Machine learning is already there for several years now. However, a lot of people still don’t want or better said don’t have the guts to fully utilize it, so machine learning remains a bit of a mystery.

Let’s be honest for a second – machine learning in performance marketing will replace humans in operational tasks, it’s not even a doubt. It is just plain more effective and efficient in doing what it does best – optimizing our ad spend to achieve the best possible results within the strategy and budget constraints. We should embrace it because it gives us, marketers, more time to channel our energy on things that matter:

  • Campaign strategy development
  • Campaign planning and testing
  • Creative thinking and creative optimization
  • Analyzing performance on an aggregated level

Simply put, machine learning enables us to be more focused, strategic, and creative. I am writing this post to enable you to embrace machine learning and give yourself more time to work on things that matter more.

Here are 5 ways I leverage machine learning in Facebook campaigns.

Automatic placements

Automation between placement ensures high budget liquidity, and therefore the highest possible efficiency considering your goal and budget. There are a few key points to understand here:

Machine learning optimizes delivery in real-time, which means that it aims to get you the lowest possible cost per result across the whole Facebook ecosystem at that particular moment. This means that it treats the whole ecosystem (Facebook, Instagram, Audience Network) as one unique platform, and we know that more people we find online in that given moment, there is a better chance to encounter the ideal candidate for our ad impression.

By enabling automatic placements, you enable machine learning to effectively create ideal conversion paths for each prospective customer. This enables the possibility to show an ad to one person multiple times, across multiple platforms, and nurture them to conversion. For example, the machine learning algorithm distinguishes that “Susan” uses Instagram in the evening to browse for ideas on where to travel next and that she mainly uses Facebook in the morning when she is more likely to click on an ad and make a conversion. In this case, the algorithm will show her our ad for destination booking first on Instagram and then on Facebook the following day, therefore making it “more probable” that she will buy our travel package.

automatic placements facebook, machine learning

You’re probably thinking now “But, Facebook and Instagram are different platforms that we use in a different way – Why should we run the same ad on both platforms?” No one said the ad needs to be the same. You can easily customize any ad to look and feel different on each platform and placement while still being the same “ad unit” in Facebook Ads Manager. I will write more about it a bit below in a section “best practices on ad level”.

So, to conclude this part, the only valid reason to not use automatic placements is if you run platform-sensitive campaigns. Campaigns with goals like engagement (page likes or post engagement) or traffic in some specific cases are doing better, considering the target outcome, if you run them on a single platform. For example, when you want to drive people to your IG profile in order to boost the followers’ count – you will use Instagram only as the platform in order to make sure that each user who sees your ad has an Instagram account.

Remember: automatic placements = more opportunities to convert = more results for the same amount of money = more profit, so turn on the option of that automatic placement and have no worries about it!

Audience liquidity

I’ve mentioned earlier how automation between ad placements ensures high budget liquidity. Generally speaking, liquidity in performance marketing is a condition in which every penny can flow to the most valuable impression, wherever it may be.

Therefore, to achieve audience liquidity you should remove the constraints of demographic targeting such as gender and age and expand detailed behavioural and interest categories. In other words, you should use as broad as possible audiences in order for machine learning to work properly. This is easily done by targeting all age and gender groups (unless your product is specifically non-suitable for some of them), and by ticking the expand detailed targeting box in order to allow machine learning to test the broader audience outside your detailed targeting constraints.

CBO (campaign budget optimization) scaling also tackles the same issue of liquidity, read more about it a bit later.

Of course, take note that optimizing for liquidity requires you to set sound campaign goals. These goals are the signal that systems follow when making calculations and should align closely with real-world business outcomes. Also, be aware that in order to fully utilize machine learning capabilities for conversion objectives your Pixel needs to be “fired up” – i.e. you need to have at least 50 conversion events for which you optimize for per week.

CBO and scaling

CBO (campaign budget optimization) ensures liquidity on an ad set level. With CBO turned on, machine learning systems dynamically manage ad spend across ad sets in order to get the overall best results within the campaign.

For example, let’s say that you run a campaign with two ad sets. Normally, you would spend the same amount on each ad set (50-50), but with CBO on – you may spend 90%, or even more of your total budget on a single ad set if that brings the best possible overall results.

facebook cbo, machine learning

So, how does CBO setup help us scale our ad spend without losing money? Well, there are two methods to successfully scale up Facebook campaigns that use CBO optimization:

  • Combination of best-performing ad sets
  • Guided scaling: best-performing ad set + broad audience set

The first one is pretty self-explanatory. You combine the best performers across your ad account into a single CBO and increase the budget. Machine learning will automatically adjust ad spend among already proven ad sets to get the best overall results. However, I advise you to do this only if you have already tested ad sets and you’re absolutely sure they perform well on a regular basis.

The second one is a bit more tricky to pull off successfully, but can be more rewarding at the same time. You’ll need to have a decent enough budget because you absolutely need to have more than 50 conversions per week on a campaign level in order to make this work. The idea is to use the most successful ad set as a machine training sample for a broader audience without any restricting parameters such as interests or too-narrow demographic parameters.

Basically, in this setup we use Facebook Pixel to do dirty work of finding best prospects in the broad audience, based on the data it gets from a smaller/targeted, but high-performing audience.

Best practices on an ad level

To maximize the results and efficiency of the ad delivery, you should treat Facebook ads like Google’s counterpart – as a group of ad assets. Your Facebook ads should essentially be dynamic and responsive whenever possible, as it will enable machine learning to maximize the results and delivery by combining different assets into a large number of ad variations. In order to achieve this, you should:

  • Adjust aspect ratio for placement groups
    • This one is pretty straightforward. By adjusting the creative for each placement group (feeds, stories, etc.) you will ensure that your ad doesn’t look out of place on different placements, while still enabling machine learning to maximize its delivery since you’re not creating separate ads.
  • Use multiple ad copies: primaries, headlines, descriptions
    • Different users react differently to ads. By using different ad copies and ideally combining longer and shorter primary texts (captions), emojis, headlines, and what not – your ad will have thousands of variations, and machine learning will be able to find the best ones for each user in your target audience.
  • Use dynamic variables
    • If you have a product catalog, you can use dynamic variables in your ads such as dynamic ad copies that use product names, price, etc., dynamic carousel cards with product images, collection showcase, and so on. My suggestion is to use this option as much as possible because it will personalize your ads based on customer behavior, especially in remarketing campaigns.
  • Don’t create too many ads
    • Since you can run multi-asset ads, there is no need to have more than 3 ads per ad group, especially if you don’t have large enough budgets to exit the learning phase fast enough. Simply put: more ads = more ad spend needed to optimize the results.

Make these tips a standard practice and you will be able to run ads much longer before you get ad fatigue issues. Also, it will enable machine learning to optimize the look and feel of each ad to a specific user, therefore maximizing the results and delivery. 

Measurement and “The breakdown effect”

“The breakdown effect” refers to the common point of confusion when evaluating Facebook ads reporting between ad sets, placements, and ads. The confusion arises as it may appear that the system shifts impressions into underperforming ad sets, placements, or ads. 

In order to understand “The breakdown effect”, I need to explain how Facebook’s ad delivery system works. The delivery system uses both bidding and pacing to determine how to deliver your ads. Pacing is the system that allows the budget to last the entirety of the schedule. Discount pacing lowers your bid when appropriate to help get the lowest-cost results available, balanced with ensuring that you spend the full budget by the end of the campaign.

discount pacing facebook, machine learning

As your budget increases, or as you exhaust all low-cost possibilities, Facebook’s ad delivery system starts to search for more opportunities which usually increases your CPA – but, it’s still the lowest-cost CPA at the given time.

So, how is this connected to “The breakdown effect”? By breaking down campaign metrics without taking this into account, it may lead you to a wrong conclusion. For example, let me qute the situation explained in the official Facebook documentation:

Let’s say you choose to run a campaign utilizing the conversions objective. You choose two placements to deliver your one creative asset to – Facebook Stories and Instagram Stories. The total budget is $500.00 USD for a single ad set using ad set level budget.

When the campaign begins, our system starts to deliver ads to both placements to see which will drive the most efficient results for your target audience — this is called the learning phase.

Facebook Stories starts out driving cheaper acquisitions, but then our system identified an inflection point at which the cost per acquisition (CPA) on Facebook Stories begins to exceed cost per acquisition on Instagram Stories.

The cost per acquisition on Facebook Stories is $0.35 on the first day, compared to $0.72 on Instagram Stories. However, as the campaign continued Instagram Stories received more budget, even though it still has a higher cost per acquisition. At the end of the campaign, Instagram Stories delivered significantly more budget compared to Facebook Stories, even though Facebook Stories originally had a lower cost per acquisition.

the breakdown effect facebook, machine learning

So, to avoid “the breakdown effect” you should always:

  • When using campaign budget optimization, evaluate your results at the campaign level.
  • When using automatic placements (without campaign budget optimization), evaluate your results at the ad set level.
  • When running multiple ads in 1 ad set, evaluate your results at the ad set level.


Final words

I will be short and concise: let the machines do what they’re supposed to do and your campaigns will shine, and your clients will be grateful. You should concentrate on strategy and creativity – we, humans, are thankfully still better than machines at it!

As always, feel free to reach out and I will be more than happy to reply back or connect you with my team at SplitX MAD in the case you’re interested in our performance marketing services.By the way, as a recognition for our smart (not hard) work, SplitX MAD became the official Facebook Marketing Partner last month!

Latest posts

can AI be creative?, dall e image
Ivan Jerkić

Can AI be creative?

Reading Time: 3 minutes So here it comes – the overarching question many people in tech debated over the last few months. Can AI be creative? In this post, I will explain my stance and see how some of the most prominent AI systems respond when prompted about their creativity.

Read more
deadlines, oneivan, ivan jerkic
Ivan Jerkić

Set realistic deadlines

Reading Time: 2 minutes The word “deadline” is overused in modern jargon. I’ve been hearing it everywhere, all the time. Yet, we are so bad at meeting deadlines, especially self-imposed ones. Why do we tend to set unrealistic deadlines? How to make them more accurate? Do we even need them?

Read more