The Secret to Accurate Affinity Data

Every moment we’re online, we make little signals of “affinity data”. Millions of affinity data points are created each day as we click, don't click, follow, unfollow, watch, stop watching, buy + bail on websites and shopping carts. 

Affinity data is the summation of the “crumbs” of data we leave along the way as we move through the digital world. It indicates whether we’re interested in a particular brand, topic or idea. It’s the “secret sauce” behind digital ad targeting, suggested content and news feed algorithms.

As marketers, affinity data is music to our ears. After all, we can leverage it to target ads, emails and events based on very precise interests. However, affinity data can quickly become overwhelming, outdated and misleading. Millions of affinity data points are created each day as we click, don't click, scroll, stop scrolling, watch, stop watching, buy and bail on e-commerce websites. This includes when a friend uses your laptop and your kids steal your Spotify (I see you). Layering affinity on top of first-party fan profiles that can be uniquely identified (i.e. a Facebook account versus a pixel-based browser) is the best way to know we’re associating these affinity signals with the right person.

Affinity data can be powerful but problematic to maintain. Activating data is about so much more than just storing hordes of data. Fan interests are also constantly changing on a daily basis, so datasets need to be constantly cleansed and updated to remain useful for targeting. In fact, out of date affinity data can be detrimental to an event’s ROI and fan experience.

So, how do we keep up? How do we know which data points are relevant? How do we prevent overwhelm and data bloat?

Just like how you can season a delicious meal with only salt, pepper and cayenne, it only takes a few ingredients to make a winning data strategy. The secret sauce for accurate affinity data has three ingredients: recency, reliability, and relevance.

Recency

Hardly anyone (my co-founder is a rare exception) listens to the same artists year over year. Most music lovers are constantly discovering new artists, growing tired of old favorites, and developing affinity for songs they hated only days previous (Old Town Road, anyone?). Who you “Liked” on Facebook ten years ago is not the best measure for who you’re jamming out to today.

We recently worked with an event that featured a fan-favorite headliner that had just released a new album after a long hiatus. This seemed like a perfect opportunity to engage their existing fan base, targeting ads at people who had followed them on Facebook or purchased tickets to their concerts years ago. However, what we didn’t predict, was the unfortunate fact their previous fan base hated their new album. Suddenly, what we thought would be our “best audiences” were flopping. We had fallen into the trap of forgetting to consider Recency.

By shifting our strategy to target recent listeners on Spotify or new fans since the latest album release, we were able to turn the results around. Understanding that just because someone has liked an artist for a decade, doesn’t mean they’ll love their new music, was a crucial example of why affinity data must be recent and dynamic.

One issue with relying on static metrics for affinity (tags, follows, likes, historical purchases) is that they only positively indicate affinity. Music lovers often “like” or “follow” an artist, but they rarely go out of their way to go back and “unlike”. They also obviously cannot “un-attend” or “un-buy” a previous concert ticket. Making sure to layer real-time indicators of affinity on top of historical data is how to keep audiences fresh because we’re considering their past and present states of affinity.

Shifting the affinity data approach to methods that involve negative affinity is another way to keep datasets dynamic. I know I rarely sit through an entire song I hate. So, when we pay attention to listening or browsing behavior, we find a much more dynamic and recent way of discerning a fan’s current affinity.

Managing and understanding the constantly changing state of affinity is much easier today than it was 20 years ago. When we have the right model to leverage machine learning, we can surface the information we need much faster by crunching through this dynamic, real-time dataset. Staring at a giant list of a thousand “interests” isn’t going to help us make data-backed decisions. Although machine learning can never replace human insight, it can help you test your models faster and more effectively. Spotify’s “Discover Weekly” is a fantastic example of this technology at its prime. That said, my kids stealing my account and throwing off the algorithm is a perfect example of why we need to correlate listening data with other indicators of affinity like purchases, and associate it with a truly unique fan profile. 

Reliability

The second crucial ingredient for accurate affinity data is reliability. Is the method for indicating affinity actually controlled by the fan, or is it assumed? As much as clicking, scrolling and browsing behavior may suggest affinity, the most reliable way to understand a fan’s affinity is to look at what they actually do and say.

Rather than relying on “assumed” affinity, look at things like recent purchases, listens, and positive mentions on social media. It's safe to say that if a fan is listening to an entire song or album, they're actively interested in that genre, artist, mood or activity. Affinity data derived from listens is therefore extremely reliable because it's direct - both from an acceptance and rejection standpoint.

Same goes for a purchase - when was the last time you bought something you didn’t like? Hardly anyone will put up their own precious cash to go to a concert or event they genuinely dislike. They may get roped in by their friends who buy them tickets, but they definitely won’t be the one navigating through a ticketing platform’s interface at crunch time when tickets go on sale.

These “signals” are behavior-based, first-party and directly controlled by the fan. Their transparency allows us to trust the conclusions. Also, make sure to own your affinity data. Avoid relying on someone else’s anonymized dataset or “patent pending” black box product, given you have no idea how it was modeled. Third-party data and pixel-based targeting are the fastest way to bleed budget and insights, given you’re advertising into an abyss and “just trusting” that the models are reliable. Even if we’re laying third-party data on top of our first-party fan profiles, understand that the assumptions that went into that dataset can still taint the results and think critically about the insights. 

Thinking back to recency, pixel-based affinity models also face tremendous challenges with updates like Apple’s expedited cookie decay. Layering third-party data onto first-party data inherently takes more time than going directly to the source - so affinity data may disappear before the platform even has the chance to link it. Collecting owned first-party affinity data allows us to see clearly whether or not our data and its derivative insights are recent and reliable.

Relevance

Affinity data points are not like Pokemon - we don’t want to catch’em all. When we collect and store every single affinity data point our fans signal, we end up surfacing only the most popular brands and artists in the world. Although it’s always nice to confirm our suspicion that everyone loves Harry Potter - that is hardly helpful when we go to sell an obscure indie artist’s concert or an EDM music festival. Also - just like Pokemon Go - even the world’s most popular fads have an expiry date.

Rather than hoarding everything, decide before collecting data, what we care about. This could be a lineup, a list of potential partners, our team, our rivals… the list goes on. But think before you leap and decide what brands, artists and teams you care about before you get lost in a sea of ever-changing affinity. Then, match current fans’ behavior to the list to reveal who is actually interested in the things (we call these “Idols”) we care about. This becomes far more helpful for targeting and prevents analysis paralysis.

Deciding our Idols before collecting affinity also protects us from “Interest Red Herrings”. Just because all our fans like Harry Potter doesn’t mean that all Harry Potter fans like our product. Many brands waste thousands of ad dollars targeting massively popular fan bases with huge competition without ever making a cent back, all because they fell into the “relevance” trap of affinity data.

You might be thinking - if I'm not a musician or music promoter, how is Spotify data relevant to my business? What people listen to indicates more than just their musical tastes. As a result, the interests derived from fans' listens can be relevant for marketers looking to tap into user's current emotional state - an important driver for conversions. In addition to signifying favorite artists and genres, Spotify listening data can indicate a fan's mood and favorite activities. This is decided through Spotify's categorized playlists like workout, party, relaxation and travel mixes. Spotify users choose these playlists to amplify their current state of mind, thus directly indicating their present interest in that field.

Conclusion

Let’s leave the thousand-ingredient recipes to the master chefs (mmm… mole), and keep our recipe for accurate affinity simple. When we start with recency, reliability and relevance, we know the results of our affinity modeling will be delicious. 

But remember, just like how a pressure cooker can’t fix a bad chef - machine learning only makes smart marketing strategies, smarter with the right models. Machine learning does not replace human judgment, it just allows us to test our models and theories more thoroughly and efficiently. When we approach affinity data and analytics with critical thinking, we reveal insights that can take our event ROI and fan experience to new heights.

Looking for advice on how to make affinity work for your business? Let’s chat over lunch (...all these culinary metaphors are making me hungry).