Watch the video below or click here before reading this article, otherwise nothing will make sense:
When you throw a bunch things at the wall to see what sticks, is often taught as a bad way of doing business, but try throwing a handful of coins in a box all the way across the room, a lot of them will miss, some of them will land, the more you try; the more of them will land, because you will optimize your hands’ rotation, speed, movement pattern, gradual deceleration and acceleration to land more of them and develop a combined skill and muscle memory precision.
Ever wondered how Facebook optimizes your ads?
We as humans; are equipped with logic, we’re able to resonate and conclude the obvious; obvious things such as: “makeup products are mostly for women, neckties are mostly worn by men”.
A computer; does not understand these apparent concepts that we naturally inherit from our families or from interacting with society, and there are thousands if not millions of these concepts all over the world, therefore; a computer has to throw a lot of stuff at the wall to see what sticks.
Trending AI Articles:
1. Making a Simple Neural Network
2. From Perceptron to Deep Neural Nets
3. Neural networks for solving differential equations
4. Turn your Raspberry Pi into homemade Google Home
The human that created the machine understands that if the wrong thing sticks, it would be a very bad idea to keep throwing the same wrong thing over and over again, yet Facebook keeps throwing our ads at the wrong people, over and over again, just like most of the cars in the video kept crashing, while very few were succeeding, so the computer knows how to succeed yet it allowed a good portion of the cars to fail over and over again.
I often see a budget optimized campaign with 2 adsets:
1st adset spent $1372.50 with 0.8 ROAS — this is terrible
2nd adset spent $73 with 4.7 ROAS — this is amazing
“Why are you doing this to me Facebook?” — said every marketer out there.
Enter machine learning:
At first; the blue picture looks like a crazy wall made by some obsessed detective.
Except it’s a machine’s wall, however; it’s not a crazy wall, it’s a very logical one.
Machine learning, if boiled down to its fundamental principle; is basically a machine doing a bunch of wrong things until it gets enough right things, just like the car video you just watched, the cars made a lot of random moves and decisions, that is the same thing every time your ads gets shown to people; as it becomes easier for Facebook’s algorithm to know what to do, it has to deviate from perfection and do random things all the time because it needs to collect more data of the things not to do as well, the smarter it gets the less it deviates.
Now we know that machines learn by doing, we also know that Facebook doesn’t know what we want, and finally we know for certain that Facebook knows what people are interested in but not what they want.
To summarize; Facebook doesn’t know what people want; whether they’re users or advertisers, unlike Google; Google knows exactly what users want, because a user is searching for a cat toy, the user wants a cat toy, the advertiser is selling a cat toy, but on Facebook; people are only interested in cats/toys, and advertisers are targeting cats/toys.
There are instances where Facebook may predict what people want, like when someone is searching groups for a product or a new apartment, someone have been shopping for a new shirt 3 days ago, if you’re selling shirts; that person is your lookalike audience, Facebook can also predict wants and needs by comparing people’s behavior to similar people that already completed an action like buying a shirt, let’s say Sam wakes up every morning at 6am; likes certain brands, watches certain shows and follows certain personalities, Sam was recently geo-tagged at Times Square wearing a blue horizontal striped shirt; we’re not here to judge, Sam recently purchased something and triggered a pixel for a shirt brand, that shirt brand created a lookalike audience of their buyers and now you have a lot of Sams, ready to buy a bunch of shirts that shall not be judged in this article.
Facebook is your matchmaker, you’ve got to go on a few bad dates until you figure out your type, the sooner you know your type the sooner you find your perfect match.
Now we know that it is inevitable that we’ll have bad dates and we also understand that we must determine our type.
The best thing about Facebook is that it allows us to try thousands of combinations and search for potential perfect matches in a blink of an eye, we have dozens of metrics to help us understand people and determine if they’re interested, things like link-click through rate (CTR), cost per click(CPC), relevance score, but these things don’t matter if you’re losing money.
I’ve had ads with insanely high CTR and 10/10 relevance score, cheap link clicks but most of them failed to generate positive ROAS
However, few of them were able to generate positive ROAS
These are your blue swans (as Sam Ovens likes to call them)
Blue swans are the perfect combinations that trigger people’s buying impulses, in the picture above, people who are interested in Cats, watched our Creatives A, B, C, and read our Copy A and were positive-ROAS buyers.
The red swans are the ads that either produced purchases but weren’t profitable, or didn’t produce purchases at all.
Now we know that Cats is a blue mother swan, this interest is a winning interest and is interacting very well with all of our creatives.
Pets is a red mother swan.
Dogs is a sick swan, only 1 ad is generating purchases, we can change the text and video creatives, or simply stop selling a cat toy to people interested in Dogs.
A lot of people target small interests and even narrow them, which is a mistake, if enough people are interested in cats and also interested in toys, Facebook will narrow it for you automatically, if you restrict it, you will restrict its narrowing options, you want it to actually optimize for you, it’s smarter than all of its advertisers combined.
Same thing goes for variants, you need multiple videos, even if the difference is just the first 5 seconds, and multiple ad copies, some people might respond to emojis, some people hate them, some people like to read long copies, some people like to click a link after a short copy, test all of these combinations until you find enough profitable blue swans to scale with.
You need to start advertising with pictures, Facebook has something called inventory, a dedicated balance between image, video and other different types of ads like carousel, these different slots are reserved for different types of ads, that means there is a good chunk of people that mostly interact with just image ads, if you’re using just video; your reach will be limited to people who are most likely to convert with video, that is why you need to test pictures and different ad types alongside videos.
Finally we understand that for testing; Facebook needs to have everything under its control, all placements, all ages, all auto, we understand that it needs few interests as a guide/hint, we understand that Facebook needs more variants, videos, images, copies so it can actually optimize for the perfect combination, because there’s nothing to optimize for if you specify that you want to target Cats narrowed down to Toys, Mobile timeline, female age 25–64, with 1 copy, ad video, there is nothing to optimize for, you think you’ve already determined your winning combination, your car will keep hitting the same wall over and over again because the directions are restricted.
Now go back to the blue crazy wall picture and look at it again, it will all make sense.
Eslam Omar was here — Feb 2019