While most dig­i­tal mar­keters are inti­mately famil­iar with using rec­om­men­da­tions to pro­mote prod­ucts and other con­tent, it’s impor­tant to look at what makes for effective—or ineffective—recommendations strategies.

Suc­cess­ful rec­om­men­da­tions that deliver real, mea­sured returns are built on under­stand­ing how prod­ucts are related to each other in var­i­ous ways, ver­sus just look­ing for prod­ucts that are cat­e­gor­i­cally sim­i­lar. This means mak­ing rec­om­men­da­tions based on many con­tex­tual fac­tors, from a person’s past buy­ing his­tory to where they are from to what items they are view­ing in their cur­rent session.

Mar­keters need a sys­tem that can make pow­er­ful rec­om­men­da­tions based on: 

Con­tex­tual sim­i­lar­ity.

When solv­ing con­tex­tual sim­i­lar­ity prob­lems, a rec­om­men­da­tions engine should take less obvi­ous rela­tion­ships into account: Do the two prod­ucts belong to the same brand? Are they in the same price range? Do they have sim­i­lar rat­ings from sim­i­lar users? These types of deci­sions are best made by a search engine that can quickly deter­mine con­tex­tual sim­i­lar­ity between semi-structured data objects.

Affin­ity between items.

Prod­ucts can also be “linked” by users’ brows­ing and pur­chase behav­ior. Con­sider the clas­sic association-rule exam­ple of beer and dia­pers: dur­ing cer­tain times of the day, male shop­pers tend to buy them together, even though on the sur­face, the items are com­pletely unre­lated. Valu­able insights can be gained from uncov­er­ing these unex­pected asso­ci­a­tions. Under­stand­ing the affin­ity of one item to another based on observed user behav­ior can be a sig­nif­i­cant ana­lyt­ics challenge—especially when you’re deal­ing with mas­sive scale—and requires the use of mul­ti­ple algorithms.

Rich data.

Your rec­om­men­da­tions will be much more pow­er­ful if the cor­re­spond­ing user data is rich and detailed. The more demo­graphic and behav­ioral infor­ma­tion that can be asso­ci­ated with a user pro­file, the bet­ter, since the rec­om­men­da­tions engine will be bet­ter equipped to draw the right con­clu­sion about what prod­uct sug­ges­tions a user is most likely to respond to.

In part II of our dis­cus­sion on Get­ting the Most from Rec­om­men­da­tions, we’ll take a deeper dive into lever­ag­ing machine learn­ing in rec­om­men­da­tions to increase con­ver­sion and boost revenue.