With almost a year now working from home, every now and then I like to stay at a nice hotel over the weekend just to change the scenery. Usually after I did some browsing in my favorite hotel booking app, I would get Instagram ads from that app showing hotels that are very relevant to my search history. This is usually called “retargeting” in digital marketing terms; the goal is to push or nudge me to make the purchase (as I mentioned previously, it is questionable how effective this kind of advertising is, but that is another matter entirely). But then I started to get similar ads from another hotel booking app, the one that I never use and is the competitor of the app that I did use. Could it be that by trying to nudge their users to buy through Facebook, my hotel booking app inadvertently gives up access to their customers for their competitors? The answer is “sort of, yes, but they will still have to do it regardless”.
First the basic: how do Facebook know what I do in other apps?
The Magic of Advertising ID
If you go to https://www.facebook.com/privacy, and then go to ad preferences, and click “See your ad settings”, you will see this section:
Click on the first section to enable Facebook to personalize ads shown to you based on your activity in other apps (that are partnered with Facebook).
Advertisers usually send user events behavior data to Facebook so that they can retarget their users, or optimize their campaigns toward certain goals (e.g. optimize to purchase). So if you enable the setting above, Facebook will use your activity data from other apps to decide the best ads for you. For example, if you often browse for laptops in e-commerce apps, you will be more likely to see ads about laptops or other gadgets related to laptops.
How can Facebook link its user data with other companies’ data? The answer is a common identifier called “advertising ID”. Advertising ID is a random ID number that is set by your mobile devices. The function of this ID is, funnily enough, to anonymously identify users. For example, if my advertising ID was “XYZ123”, then when I browse for a hotel in a travel app, the app would send to Facebook that a user with ads id “XYZ123” is looking for a hotel to book. And since I use Instagram and Whatsapp, Facebook also has access to my ads id, and thus knows that ads id “XYZ123” is me. Then it can decide what ads to show to me based on all of the data it has on me from Facebook apps and its partners.
The Problem of an Unusually Opportunist Machine
Next question is how do they decide what ads to be shown to each user? Facebook uses various data points (or features) in their machine learning algorithm to make this prediction: given a user, which ads will have the highest probability of eliciting response from the user. This paper deep dived on the technicality of the methodology. And though it didn’t disclose the exact features that Facebook used, they are some clues that we can discern:
“.. the historical features depend on previous interaction for the ad or user, for example the click through rate of the ad in last week, or the average click through rate of the user”
“… historical features provide considerably more explanatory power than contextual features. The top 10 features ordered by importance are all historical features”.
What it means is users’ history toward the advertisers’ ads would play a big part in the algorithm. For example, if I am a loyal customer of app A, I should have a lot of interaction with ads from app A relative to ads from app A’s competitor, say app B. And hence, app A ads should appear more frequently in my IG feed. But since both apps are competitors to each other, most likely if I was a customer of app A, I was also a potential customer for app B. So when app B says to FB “show my ads to someone who like to travel”, FB algorithm would think (so I imagine) “Angga here like to clicks and sometimes purchase in app A. App A is a travel app. He never interact with App B, though, hmm. Ah, I know what I have to do! Occasionally I will show App B’s ads to Angga, and then I would charge a little higher to App B for that impressions! PROFIT $$$!!”.
So that is my very accurate rendition of the AI thought process. You can’t prove me wrong because all of these are hidden deep in double black boxes that are hard to penetrate: the first black box is that the data points used for the algorithm are being kept a secret, and the second black box is the machine learning algorithms themselves. Even with full knowledge of the data points, it is often still hard to interpret the decisions made by machine learning algorithms.
So does it worth it for advertisers to send users’ behavior data to ad networks like Facebook? In other words, will the benefit gained by sending the data outweigh the risk of their competitor able to reach their customers? I have no idea! And that’s exactly the problem. With 1.8 billions pairs of eyeballs browsing Facebook each day, any big advertisers will have to use any means necessary to gain an edge in competing to win these attention, even at the risk of giving their competitors access to their customers. The problem is there is just not enough transparency on the scale of that risk for companies to make more informed decisions.
