Recommended Tweets
I keep my Twitter account pretty small. It's set to private. I only follow 91 people, and my followers don't amount to a much higher number. Of those 91 accounts I follow, 17 of them, my data report informs me, I have muted, meaning that I still follow them, but their content doesn't show up on my feed. That means there's only 74 accounts whose posts are guaranteed to make it to my feed, and even then, many of those accounts haven't been active in years or only post rarely. In summary, I have less than 74 accounts whose content I regularly see, and yet I manage to spend more than an hour a day on Twitter, the majority of it spent on my feed and not on the explore tab. On their own, the people I follow are not tweeting often enough to fill the feed with over an hour's worth of content, so how does Twitter keep me scrolling? There is about as much, if not more, recommended content on my feed than there is content from the accounts I follow.
​
Of the first 8 tweets on my feed on the evening of November 19, 2020, only one is a tweet from an account I follow. One is a retweet from an account I follow. The rest are recommended content.
Tweet from an account I follow. Not a retweet or a quote tweet.
Recommended based on who I follow.
Retweet from an account I follow.
Recommended tweet because someone I follow liked it.
An ad tweeted by Bloomberg promoted by American Express.
From this small sample of my feed, it is clear how much recommendations can affect a user's experience using Twitter. The algorithmic decision-making behind what content shows up on my feed is a more complex calculation than who I follow or the tweets people I follow like. That is evident in my data report.
​
Twitter has all the staples of personal data reports in a user-friendly html file similar to Facebook's.
​
Targeted Ads and Data Dopplegangers
Besides a log of my tweets, likes, and direct messages. Twitter also includes a list of every ad that's been placed on my feed and how long I interacted with it. Twitter provides more detail than the other platforms about why they showed me certain ads.
The reasons behind an ad for an article about the 2020 presidential election from the Wall Street Journal are straightforward. The election was coming up at the time of the ad and my data says I'm an American of voting age.
Sometime in June 2020, I visited the Wall Street Journal's website, putting me on a list that WSJ then used to include me in the audience for this ad.
Despite a similar display, it's not nearly as clear how this ad from the Microsoft store came to be on my feed. The answer is coded in indecipherable terminology followed by an equally indecipherable string of numbers. Here, Twitter can only be as transparent as the ad owner allows them to be.
Follower look-alikes are another way Twitter curates recommended content for their users. The advertiser identifies accounts whose audience they want to target, and then Twitter provides accounts that look similar the followers of the designated account.
​
Here, Kraft selected a long list of popular accounts to find look-alikes. Then Twitter determined that my activity and data are similar to the followers of these accounts, which made me part of the target audience for this ad.
​
While Kraft has chosen some of the most followed accounts on Twitter to identify follower look-alikes (i.e. Beyonce, Chrissy Teigen, Justin Bieber), Twitter actually advises against the practice of using extremely popular accounts to find an ad audience. Instead, Twitter Business suggests, it's more effective to "target the @handles that are most closely tied to your business. While very famous individuals or companies may have millions of followers, those millions are not necessarily aligned around a topic that’s relevant to your business."4
Unlike Kraft, the pharmaceutical company Novartis seems to have followed Twitter's advice. For this ad about the "synergy between art and science," they only used one account to find look-alikes. I learn from his twitter bio that Atul Gawande is a surgeon, writer, and researcher. He is not followed by anyone I follow on Twitter, and I do not follow him myself. Yet, for reasons not apparent in my data download, I resemble his followers enough for the Norvartis ad to appear on my feed.
Nodes in a Web
My Twitter data, particularly the concept of follower look-alikes, is the best example so far of our how data identities do not exist independently. A chain of connections is what leads suggested content to my feed.
Tweets tagged with "[user] liked" or "[user] and 8 others follow" are simple instances of this. There is an established connection between my account and the accounts that I follow. When one of the accounts I follow likes a tweet, it establishes another connection between them and the tweet they liked. To determine what suggested tweets should show up on my feed, Twitter's recommendation algorithm looks at the established connections between users and content (i.e. a like) and then at the established connections those users have with other accounts, how often those accounts like similar content, and from that information and other factors like it, they identify the content I have the highest probability of engaging with and place it on my feed.
​
Identifying look-alikes is a similar process on a larger scale. The starting point is the account identified by the advertiser. The established connection is between that account and its followers. Look-alikes are found through the existing connections between the followers and the other content they like. By comparing those connections to the connections of other users who don't follow the starting point account, Twitter is able to find look-alikes. Our data identities are all nodes in an always-expanding interconnected web. The meanings and inferences made about our data is dependent on how each of our data identities relate to each other.
