by Tobias Peggs
Mark Suster has a great post outlining several reasons why he’s doubling down on Twitter. One such reason is the power of implicit data that exists within the Twitter ecosystem.
Targeting ads to “Sports Guys” using implicit Twitter data.
As Mark writes: “Do you follow Fox News, Rush Limbaugh, Sean Hannity and Glenn Beck? Well if you also follow Keith Olberman and Rachel Maddow then I can interpret something about you. If you follow the former and not the latter I can interpret something else. Companies [can] run correlation analyses to determine probabilities that you are a, b or c. Even if they’re not using this to target you with information sent via Twitter, they might use it in aggregate to determine things like the likely electoral votes in a region that will swing for a candidate. Or the probability of Southern Democrats to buy cable versus satellite or an Android versus an iPhone.”
At OneRiot, we use similar analysis to determine what users are interested in right now, and then target ads using that data.
You can “see” how it works using this tool. Simply type in a Twitter ID and, based on implicit data, you can “see” what that person is interested in right now. This is broken out by IAB interest categories (e.g. “Sports” is a Tier 1 interest, with “Sports->Cycling” being Tier 2). The tool also shows a sample advertisement that OneRiot might display to that user.
Sonos ad targeted to a male user with interests in Music and Technology
This simple tool sits on top of some pretty sophisticated Big Data Analysis, Machine Learning and Natural Language Processing. We’ll cover the technical detail in a later post, but some important considerations include:
1 – Who a user follows. As Mark points out, who you follow can be used to help interpret what you are interested in. To use a simple example: I might never tweet… but if I follow a number of influential tech bloggers, I’m clearly using Twitter to consume tech content, and I’m interested in “Technology”. To get a more accurate read of interests, it’s also important to understand how influential those tech bloggers are in certain content categories. To enable this, we’ve developed a granular Klout-like score for category influence. For example, @parislemon might be a tech influencer who OneRiot scores highly on content related to “Apple” but low on “Microsoft”. If I chose to follow him over another tech influencer with the opposite scores, it might indicate that I’m more interested in products coming from Cupertino than Redmond (which in turn influences what ads OneRiot will display for me).
2 - Content in Tweets tweeted. This seems like the most obvious signal, but we use it in a non-obvious way. If a user tweets “Watching the movie Horrible Bosses #LMAOOO”, you could imply that the user likes comedy movies. However, more interesting to us is the word-level unigram, “#LMAOOO”. By studying n-grams in hundreds of millions of tweets, from millions of users, we’ve built a statistical language model that can determine user attributes such as age, gender, ethnicity and even income level. As an illustrative example, it maybe that 18-24 year old Hispanic females are more likely to emphasize the “O” in “#LMAOOO” than other age groups or other ethnicities. Hundreds of thousands of similar indicators compete inside our language model. The end result is that it’s possible to target ads to specific demographic groups with incredible accuracy.
3 - Content in Tweets consumed in the stream. Twitter is a global conversation, and the realtime firehose of Tweets reflects the changing interest of its users. Accordingly, OneRiot processes the content in that firehose to identify temporal interests that a user might have right now. To illustrate, let’s assume that “User A” is not typically a movie fan, and follows no defined “movie influencers”. We would not normally target that user with a movie advertisement. Meanwhile, elsewhere in the Twitterverse, a cluster of known movie influencers suddenly start Tweeting a lot about “Super 8”. Our system will automatically connect the dots and understand that “Super 8” is hot in the movie category. To continue with the illustration, let’s now assume that because “Super 8” is hot, a lot of people that “User A” does follow also start Tweeting about it, and “User A” consumes those tweets. OneRiot would now automatically categorize “User A” as being temporarily interested in that movie. There’s a window of opportunity to effectively target “User A” with an appropriate movie ad. When we use targeting techniques like these, the performance is through-the-roof fantastic. (Our theory is: from User A’s perspective, the ad content is out of their norm, so it stands out. But it’s also “socially relevant right now”, which is intriguing. The one-two combination results in a killer click-through rate).
Now, we combine all this magic (and more) for targeting ads specifically on mobile. Why? Well, targeting for mobile ads is hard. Generally speaking, cookie-based technologies which work well on the web for audience targeting do not work on mobile. So in the “cookie-less world” of mobile, media buyers need other signals to identify a target audience. OneRiot’s primary targeting signal is “social”, including a heavy dose of Twitter. Using techniques like those covered above, OneRiot analyzes social media activity published and consumed by mobile users to implicitly determine audience characteristics – and we target ads based on that data. This approach capitalizes on the fact that mobile is inherently social, meaning social signals are by far the strongest way to determine what content the mobile audience will engage with at any point in time.
I hope this provides some insight into the “how” of OneRiot’s socially targeted mobile ads, and how we use implicit data in the Twitter ecosystem to help. For the “why”, read here, here and here.