A very good starting point to get a feeling about the interesting customer attributes is to look into the brand traffic.
- What is the distribution of age/genders?
- What is the seasonality within a day, week or over the year?
- What is the device impression share?
It depends a lot on your business how the derived user attributes look like, in the end you have a set of attributes with a value that is typically for your existing customer base. You can also cluster them using e.g. kmeans and build a group of the most important buyer personas. The result is a fingerprint of our target users.
Too abstract? I will show you an easy example: Let’s say we run a B2B business. We realized in the brand traffic that there is a huge difference in weekday vs. weekend impressions. Not a big surprise…
One new feature could be the ratio of (Impressions Tuesday + Impressions Wednesday) / (Impressions Saturday + Impressions Sunday). Let’s name it business demand. The higher the better for our B2B environment.
Search for the brand fingerprint in your non-Brand data
Now the interesting part: How does your non brand traffic look like? Normally different because in most cases the used keywords are also searched by customers that are not in a b2b context. To find similar demand fingerprints we do the same business demand calculation for different account levels we can bid/target on. Here are two examples:
- Score your keywords with the business demand. A huge advantage is that only a small amount of impressions are necessary to score each keyword. By looking at conversions you need hundreds or thousands of clicks – especially in the longtail it’s difficult to get enough samples. If you round the calculated business demand to e.g. 1 digit you get pretty large groups with hopefully a big gap in performance. Bingo!
We are spending a lot of time in feature engineering like that to build custom bidding models that deal very well with longtail keywords.
- Let’s do the same for location targeting now. Normally your sample size is to low to even adjust bidding on citylevel when looking on conversion data only. By using your classification on impression level we can now go down even to postal level. The process is the same: for each postal level we get a score – we can group conversiondata by the calculated score. A straight forward action could be:
Bid Adjustment on Postal = ValuePerClick(Postal group)/ValuePerClick(Total)
Remember: In those two examples we only build a fingerprint with one feature. Depending on your business you will find a lot more. What we observed in a lot of accounts is that the same keyword can perform completely different when you start mixing them with segments based on user demand.
By identifying those interesting user micro segments you can start building some completely new campaign types that are bidding super aggressive for those users. In our opinion this is a very smart approach for new customer acquisition.