This is the first in a series of posts about algo­rith­mic audi­ence mod­el­ing offered by AudienceManager.

Look-alike mod­el­ing arrives
I’d like to take a moment to announce the release of look-alike (algo­rith­mic) mod­el­ing in Adobe Audi­ence­M­an­ager. Look-alike mod­el­ing has long been a buzz­word, and a set of desired fea­tures, within the online adver­tis­ing com­mu­nity. Indeed, indus­try pro­fes­sion­als have been talk­ing about auto­mated audi­ence dis­cov­ery since the late 1990s. How­ever, no one really has been able to deliver on the promise of algo­rith­mic data analy­sis, dis­cov­ery, and mak­ing this infor­ma­tion easy to use. True, some prod­ucts in the ecosys­tem say they offer look-alike mod­el­ing, but most sim­ply offer black-box solu­tions and don’t pro­vide trans­parency into their sys­tems. As a result, you may never be sure that data was sub­jected to a true, algo­rith­mic analy­sis. How­ever, for those of us work­ing on Audi­ence­M­an­ager, we really wanted to take the mys­tery out of the mod­el­ing process, describe how our model works, and deliver results that mat­ter to our cus­tomers. Audi­ence­M­an­ager mod­el­ing is the result. This is a stan­dard fea­ture avail­able to all Audi­ence­M­an­ager users.

Let’s review how our look-alike model works and the poten­tial ben­e­fits you might get from this new feature

Find new users with TraitWeight
Audi­ence­M­an­ager uses a pro­pri­etary algo­rithm called TraitWeight to dis­cover new, unique audi­ence mem­bers. This process starts after you select a trait or seg­ment, a time inter­val, and first or third-party data sources for analy­sis. As it runs, TraitWeight looks for users in the data sources that are iden­ti­cal to qual­i­fied users in your selected trait or seg­ment. Next, TraitWeight weighs the results. Weight numer­i­cally ranks newly dis­cov­ered traits in order of influ­ence or desir­abil­ity. The scale runs from 0 to 1. Traits weighted closer to 1 means they’re more like the audi­ence in your base­line pop­u­la­tion. Also, heav­ily weighted traits are valu­able because they rep­re­sent new, unique users who may behave sim­i­larly to your estab­lished audi­ence. In the final step, Audi­ence­M­an­ager dis­plays the weighted results in Trait Builder, where you can cre­ate traits based on the weighted score gen­er­ated by the algo­rithm. You can use these results to build accu­rate traits, or trade some accu­racy for reach to help expand audi­ence size.

Advan­tages of look-alike modeling

Some ben­e­fits to work­ing with algo­rith­mic mod­el­ing include:

  • Data accu­racy: The mod­el­ing processes run at reg­u­lar inter­vals, which keeps results cur­rent and rel­e­vant. The Audi­ence­M­an­ager mod­el­ing sys­tem is always work­ing for you to find and extract new value from all your data.
  • Automa­tion: The model will find new users for you. As a result, you don’t have to man­age a large set of sta­tic rules for audi­ence dis­cover and creation.
  • Save time and reduce effort: With our mod­el­ing process, you don’t have to guess at the traits or seg­ments that may work or spend scarce resources (time, effort) on cam­paigns to dis­cover new audi­ences. The model does that for you.
  • Reli­a­bil­ity: Mod­el­ing works with server-side dis­cov­ery and qual­i­fi­ca­tion processes that eval­u­ate your own data and the selected third-party data that you have access to. This means you don’t have to see vis­i­tors on your site to qual­ify them for a trait.

Buy-side use case: Extend­ing the util­ity of retargeting

Now that you have a gen­eral under­stand­ing of how the model works and its ben­e­fits, let’s look at how look-alike mod­el­ing might be applied to help an adver­tiser with audi­ence retargeting.

Let’s not mince words here: adver­tis­ers love retar­get­ing. What isn’t to like? It deliv­ers on dig­i­tal advertising’s promise of find­ing a very spe­cific set of users. And, the per­for­mance for these users is usu­ally off the charts. How­ever, the prob­lem with retar­get­ing is that it mainly reaches exist­ing cus­tomers rather than new cus­tomers. Also, retar­get­ing typ­i­cally reaches a rel­a­tively small pool of users. So, basi­cally, retar­get­ing does not address the need to find new cus­tomers that have never inter­acted with a par­tic­u­lar brand online. AudienceManager’s look-alike mod­el­ing helps solve these prob­lems by find­ing new users who may be inter­ested in a prod­uct (accu­racy) or by help­ing expand your poten­tial qual­i­fied cus­tomer audi­ence (reach).

Finally, let me point out that the real value in retar­get­ing is more about giv­ing mar­keters insight into how their cus­tomers behave rather than the tar­get­ing itself. Pre­vi­ously, we’ve had to guess at the answer to the ques­tion “What is spe­cial about these users that make them my cus­tomers and how do I find more of them?” Look-alike mod­el­ing can help answer that ques­tion for us. In this case, behav­ior is key.

Sell-side use case: Deliv­er­ing unique value to the buyer

As much as the buy-side loves retar­get­ing, it is safe to say the sell-side loathes it. Fre­quently, the first phrase I hear from many pub­lish­ers (when the topic comes up, espe­cially in rela­tion to RTB), is ‘cherry-picking.’ His­tor­i­cally, adver­tis­ers were forced to come to pub­lish­ers and buy pages that con­tex­tu­ally matched the audi­ence they were look­ing for. How­ever, more recently, many large pub­lish­ers have moved away from sell­ing con­tent and towards offer­ing audi­ence ad prod­ucts based on their users’ online behav­ior. With both the buy-side and sell-side look­ing to iden­tify a spe­cific audi­ence, it is often dif­fi­cult to match the users that an adver­tiser is look­ing for with the audi­ences the pub­lish­ers have built using their own first part data sources. This typ­i­cally causes the buy-side to revert to basic re-targeting, so they can feel com­fort­able they are buy­ing the right audi­ence. Look-alike mod­el­ing offers a solu­tion to this prob­lem that can work for both sides.

In the end, both the sell-side and the buy-side ben­e­fit from look-alike mod­el­ing. Pub­lish­ers can build mod­els and cre­ate new audi­ences from their own data. They no longer have to sell con­text spe­cific pages or pre-packaged audi­ences to buy­ers. This helps adver­tis­ers find net new users on a publisher’s site and allows busi­ness part­ners to iden­tify the audi­ences that per­form well. Basi­cally, pub­lish­ers ben­e­fit by expand­ing reach vs tra­di­tional retar­get­ing cam­paigns and by cre­at­ing strate­gic adver­tis­ing pro­grams for the buyer, a rare sit­u­a­tion in the tra­di­tion­ally (some­what) antag­o­nis­tic rela­tion­ship between the buy and sell side.

Fol­low up post: Imple­men­ta­tion steps

Now that we’ve reviewed how look-alike mod­el­ing works and how it can help buy­ers and sell­ers, we’ll fol­low up with this post with a step-by-step walk­through of the model cre­ation process.

All of us on the Audi­ence­M­an­ager team are very excited about TraitWeight and look-alike mod­el­ing. How­ever, we think this just the begin­ning, a scratch on the sur­face of what this and other Adobe pre­dic­tive ana­lyt­ics solu­tions can pro­vide. In the com­ing months we hope to expand these capa­bil­i­ties by improv­ing TraitWeight and adding new algo­rithms for improved modeling.