Why You Need to Understand Marketing Machine Learning Models

Warning: this content is older than 365 days. It may be out of date and no longer relevant.

Why You Need to Understand Marketing Machine Learning Models

One of the technical marketing hurdles I hear marketers struggling with on a regular basis is the idea of an algorithm. Marketers talk about Google’s algorithm, Facebook’s algorithm, Instagram’s algorithm, and this bit of language matters a great deal in our understanding of what’s going on behind the scenes with big tech and marketing.

To clarify, an algorithm is a process with a predictable outcome. Any time you pull out a cookbook, follow the instructions for a recipe, and cook the dish more or less as it’s described and depicted, you’ve used an algorithm.

That is not what Facebook et. al. use when they serve us content and ads. It’s not a single monolithic process, but a complex mixture of processes and data to create their desired outcome (which is ad revenue). When we talk about machine learning and AI in this context, these companies don’t have algorithms. They have models.

Machine Learning Models Explained

A machine learning model – from the most basic linear regression to the most complex multi-task unified model – is essentially a piece of software. The difference between regular software and machine learning software is mainly in who wrote it – machine learning software is written in part or in whole by machines. Google’s search AI? That’s a model (it’s actually a collection of models, but that’s a story for another time). With Instagram’s slightly more transparent explanation of how its feed works, we see that it too is comprised of a sophisticated model with many different pieces. Here’s what head of Instagram Adam Mosseri had to say recently on a now-deleted blog post:

We start by defining the set of things we plan to rank in the first place. With Feed and with Stories this is relatively simple; it’s all the recent posts shared by the people you follow. There are a few exceptions, like ads, but the vast majority of what you see is shared by those you follow.

Next we take all the information we have about what was posted, the people who made those posts, and your preferences. We call these “signals”, and there are thousands of them. They include everything from what time a post was shared to whether you’re using a phone or the web to how often you like videos. The most important signals across Feed and Stories, roughly in order of importance, are:

Information about the post. These are signals both about how popular a post is – think how many people have liked it – and more mundane information about the content itself, like when it was posted, how long it is if it’s a video, and what location, if any, was attached to it.

Information about the person who posted. This helps us get a sense for how interesting the person might be to you, and includes signals like how many times people have interacted with that person in the past few weeks.

Your activity. This helps us understand what you might be interested in and includes signals such as how many posts you’ve liked.

Your history of interacting with someone. This gives us a sense of how interested you are generally in seeing posts from a particular person. An example is whether or not you comment on each other’s posts.

From there we make a set of predictions. These are educated guesses at how likely you are to interact with a post in different ways. There are roughly a dozen of these. In Feed, the five interactions we look at most closely are how likely you are to spend a few seconds on a post, comment on it, like it, save it, and tap on the profile photo. The more likely you are to take an action, and the more heavily we weigh that action, the higher up you’ll see the post. We add and remove signals and predictions over time, working to get better at surfacing what you’re interested in.

In his language, he clearly describes the basics of the machine learning models that power Instagram, the inputs to those models, and the expected outcomes. That’s essentially an explainability model for Instagram.

Why Understanding Machine Learning Models Matter to Marketers

So what does this all mean? Why does this matter? When we think about machine learning models, we recognize that they are essentially opaque pieces of machinery. We, as marketers, have little to no control or even oversight into what’s inside the models or how they work. Frankly, neither do the companies who make them; they control the means by which the models are assembled, but they’re so complex now that no one person understands exactly what’s inside the box.

To put this in a more understandable context, what do all the pieces inside your blender do? We know the basics – electricity activates magnets which turn gears which make the blender go – but beyond that, if someone put a pile of modern blender parts in front of us, the chances of any of us reassembling it correctly are pretty much zero.

But we don’t need to, right? We need to know what it does, and then the important parts are what we put in the blender, and what comes out of it. If we put in sand and random plant leaves, we’re not going to have a particularly tasty outcome.

Machine learning models are just like that: what we put into them dictates what comes out of them. In Mosseri’s post above, he calls the inputs signals – essentially, data that goes into Instagram’s model, with the outcome being a feed that keeps people engaged more (and thus showing them more ads).

Which means that the only thing we have control over as marketers in this scenario is what goes into our audience’s machine learning models. We can do this by one of three ways:

  1. Create such amazingly great content that people desperately want to see everything we share. They mark us as Close Friends in Instagram, or See This Person First in Facebook, or hit the notifications bell on YouTube, etc.
  2. Buy ads to show our stuff to our audience more frequently. This is what the tech companies are aiming to optimize for.
  3. Divert attention through external means to our content on the algorithm we want to influence most.

Point 1 is table stakes. If your content isn’t good, none of the rest of this matters. Get that right first.

The real question comes down to 2 and 3; I lean towards 3 because it tends to cost less money. By using external platforms to influence what ingredients go into the various machine learning models’ inputs, I can change what comes out the other side.

If I put even one strawberry in a blender with other ingredients, everything will come out with at least a bit of strawberry flavor. If I can get my audience to at least one piece of content that’s seen by machine learning models, then I change the signals that model receives, and in turn I influence that model to show more of my stuff to my audience.

How do you do that? Here’s an actual example. I featured a video recently in my newsletters, which many of you watched:

“>Example video in newsletter

What does that do to YouTube’s recommendation engine? It looks at watch history, watch time, etc. and then recommends things you might also like that are in a similar vein. This in turn means that other videos on the channel get recommended more often to people who have watched the one I shared. What does that look like?

Video views history

At point 1, we see the baseline of all video views on the channel before I started these tests.

At point 2, we see the video I published and promoted heavily in newsletters.

At point 3, we see a new baseline established for all video views.

By using an external mechanism to promote the video, I changed – briefly – the inputs into YouTube’s recommendation engine for all the people who watched the video. If I sustain this process, I should see the channel’s videos do better and better over time, including videos I haven’t shared or promoted.

That’s how we change the inputs to machine learning models, by using external promotion mechanisms. We can of course do this with advertising as well, but if we have the assets and capabilities to promote using lower cost methods, we should do those first.

Where should you do this? On any channel where you care about the performance. I don’t do this on Facebook, for example, because I don’t particularly care about the channel and engagement there is so low for unpaid social media content that it’s a waste of attention to send people there. YouTube’s performance for me has been substantially better over last year or so, so I direct attention there. Decide which channels matter most to your marketing, and use this technique to alter what the recommendation engines show your audience.

You might also enjoy:

Want to read more like this from Christopher Penn? Get updates here:

subscribe to my newsletter here

AI for Marketers Book
Take my Generative AI for Marketers course!

Analytics for Marketers Discussion Group
Join my Analytics for Marketers Slack Group!


Leave a Reply

Your email address will not be published. Required fields are marked *

Pin It on Pinterest

Share This