With so many factors in play, it was difficult to predict what was going to be the best solution. We decided as a team that the best way forward would be to start simple, create hypotheses to test, test those, adapt.
The first step was developing a tool that exported job locations in a longitude and latitude format. This was a first step to gather data with.
Immediately, the first limitation was noticed.
The volume of locations exported - many employers exceeded the 500 proximity target limit as set by Google per campaign.
This brought up a lot of questions on how locations had to be grouped (more on that later).
But, once we had the first exports we got started with our tests and iterations:
➡️ First iteration - Moving from Google Ads defined locations (i.e. zip codes and cities) to proximity targeting (longitude and latitude coordinates).
➡️ Simplistic and quickest approach - i.e. export of coordinates from location feed.
➡️ Relatively small radii - no differences between different locations.
➡️ Lots of overlapping radii
➡️ Poor total coverage of total jobs as result
➡️ No prioritisation for most profitable campaigns
Iteration 1 gave us a starting point, the easiest, simplest first move. From here we could have a benchmark and test other ideas against.
One of the very first issues we noticed was that with so many radii overlapping we were wasting precious space on our campaign location targets (as mentioned previously, we are limited to 500 per campaign). We didn’t want to create an overly complex account structure, so with the first test we wanted to see if we could group locations together and see how that performed against iteration 1.
➡️ Developed from Iteration 1
➡️ Wanted to test a geo-location method where we could overcome the 500 proximity target limit (and try to avoid creating an overly complex campaign structure).
➡️ Developed a grid system over the US that created blocks and we would then group each target location within that block to create the smallest circle which would incorporate all radii.
➡️ Radii are now larger, embarking on bigger areas and thus more jobs.
➡️ Limited radii overlap.
➡️ However, larger radii might open us up to inefficiencies as we target areas that might not have any roles.
➡️ Still no prioritisation for most profitable campaigns.
In order to group locations, a system was developed to create a grid system over the US using ‘blocks’, then create the smallest circle in that block that covered all the radii within it. Below is a screenshot of how that looked.
Within Google ads that transformed into the targeting shown below:
Test 1 Results (Iteration 1 vs Iteration 2)
We ran an experiment where we tested Iteration 1 vs. 2. We wanted to understand if smaller more precise radii could bring more revenue & ROAS than larger radii.
Iteration 2 did a better job on ensuring we had more ‘job coverage’, but there were still obvious issues with it. The issues included being too broad and potentially acquiring irrelevant traffic.
We ran the test over the course of 40 days and once we reached statistical significance we had our answer. On every metric that matters, CPI, ROAS, CTR, CVR, and Revenue Iteration 1 won.
It was good to sense check and confirm our initial hypothesis around how accurate our targeting should be, we were now ready to move on.
Iteration 2 was a deadend, moving onto Iteration 3….
➡️ We wanted to limit radii overlapping & use predictive value weighting for locations (to ensure we had the best job coverage in the location targets).
➡️ Experimented with using Hexagonal Plotting (originally made famous by how Uber mapped the world) to overcome grouping and potential value weighting.
➡️ These gave us some substantial benefits which included no overlapping, and almost 100% job coverage.
➡️ Prioritises grouping locations based on a blended calculation between historical ROAS performance and job volume instead of purely looking at job volume.
Using Hexagonal Plotting to group locations across the US:
The output in Google Ads:
Test 2 Results (Iteration 1 vs Iteration 3)
Our second test focused on eliminating some location overlap problems with the previous versions & we wanted a way to weight locations that we knew had good historical performance.
Iteration 3 used Hexagonal Plotting in order to approach location grouping in a different way - profitability vs. volume of jobs.
While Iteration 1 and Iteration 2 would ensure that areas with the most openings were covered, Iteration 3 would look at a blended number between the volume of jobs and forecasted profitability per location.
We ran the test for just over a month, and once we reached statistical significance we found Iteration 3 (Hexagonal Plotting) Drove higher CPI, CVR, ROAS, and revenue vs. Iteration 1.
At this stage we were happy with geo-location targeting methodology and decided to move on to building the automation systems and monitoring systems.