With roughly two billion users, Facebook is among the most powerful Internet services on the planet. That it collects a significant amount of information about these users should be news to no-one, especially not marketers. According to a recent Washington Post article, the social media giant looks at 98 personal data points to target ads on the platform. It does this not only by tracking your activity while you’re on the site - ad clicks, device and location settings, phone brand - but also every site you visit while you’re logged in. Even when you’re logged out, it is alerted every time you load any page that has a ’like’ or ‘share’ button, effectively allowing it to collect almost your entire browser history. If this wasn’t enough, Facebook has partnered with Experian, Acxiom, and Epsilon, who hold even more consumer data from government and public records, loyalty card purchase histories, and so forth.
Essentially, this helps to create the nearest thing to a complete consumer profile in history, and it's one a canny marketer has to be able to exploit if they are to succeed. Those who fail to do so will potentially be reaching out to people who are extremely unlikely to purchase their product or use their service and waste both time and money.
At the recent Social Media Analytics Summit, Dominic Williamson, Lead, Marketing Science at Facebook ran through the most effective kinds of targeting that marketers could use based on the data available.
The first layer of targeting is things like demographics, location, interests, and so forth. Some of these are extremely clever noted Williamson, for example, partner data. But they are all things that are yes or no flags. This level of binary signal is really useful in certain circumstances, but you have to go beyond such deterministic signals and look at probabilistic points if you want to truly target the right people.
The second stage is lookalike targeting - finding people who share traits and interests with your customers. In order to do this, you need to find a seed of people that you know very well, such as a list of very qualified users. Facebook then uses this to build a model in a black box somewhere alongside all the factors in the first layer. It then scores everyone in the population and takes the people most similar to your initial seed group. So, rather than just having basic binary flags such as yes/no, lives in New York, most of the signals are far more subtle and nuanced. This allows you to target similar people outside a certain area.
The next stage of targeting with Facebook advertising is finding the people most likely influenced by your advertising. Williamson uses an example from the New England Journal of Medicine of a chart that seems to suggest a country’s level of chocolate consumption is strongly correlated to the number of Nobel Prize winners it produces. There are three possible explanations for such findings. One option is that it could be pure coincidence, another that chocolate could genuinely aid cognition somehow and help people become a Nobel Prize winner, or finally, there could be a confounding variable. To test this, you would need to either build a really big model that takes into account every single factor and teases out all the possible causes, or you run a test. You hypothesize that eating chocolate helps you win a prize, and split the population into two - half of which you allow to eat chocolate, the other half you stop from doing so. Facebook goes with the second option. It sees relationships between response and clicks, but acknowledges that it is hard to say whether the relationship is causal. In order to get around this, Facebook has a test control framework. A marketer has to identify their business objective and the campaign they're looking to measure. Facebook then randomizes and splits the audience into test and control groups. It then runs the campaign to the target audience, with ads getting delivered to the test group but not delivered to the control group. Facebook calculates lift by comparing conversions into the test group to conversions in the control group.
This leads to stage four, in which everyone is divided into four groups showing the probability to convert: a sure thing, a lost cause, do not disturb, and persuadable. A sure thing is going to buy regardless, so you don’t need to waste money on them. Equally, a lost cause is not buying regardless of any ad, so don’t waste money on them. Those flagged as do not disturb will actually even be turned off by an ad. It is the category, ‘persuadable’ that you should be focused on.
This is an extremely powerful tool. If we see the internet giant’s use of our data as valuable in so much that it shows us ads relevant to us, this is a good thing. If you see it as creepy, it is not a good thing. For marketers, however, it is a necessary thing.
VIEW DOMINIC'S FULL PRESENTATION BELOW: