Proving the machine’s superiority

The problem we were solving

To prove this rapidly emerging hypothesis, in 2021 we have tested a brand new approach to solving one of the above-mentioned seller problems: the management of Ad Rates. Under a normal (manual) scenario a seller would manually attach either recommended or preferred ad rates to each of the listings they’re promoting. The problem sets in almost immediately, as ad rates begin to “spoil”. An ad rate that was once competitive at the time of launch (maximum impressions without overspend) loses its efficiency very rapidly as competitors’ ad rates begin to skew the competitive balance. As a result, static ad rates begin to either over-compete (driving up the Ad Spend) or under-compete (driving down impressions and resulting sales).

The lack of time required for the active ad rates management had created an epidemic of “evergreen” campaigns with progressively deteriorating performance. On average, a seller would have around 22 CPA campaigns with around 2,000+ listings in each. It simply becomes cost-prohibitive to micro-manage and groom this volume of promoted inventory on a daily basis. Even in the best possible scenario, it would simply be above a human capacity to track and analyze the search results in real time, making non-stop hairline adjustments to competitiveness of each individual ad rate. What’s worse, eBay is legally bound to keep ad rates confidential, leaving the seller competing against an unknown variable. All of this had resulted in the “set it and forget it” approach presenting one of the key problems in eBay’s advertising backlog.

To address the root cause of the problem we had launched a close Beta to test a new approach in a live environment: The Ad Rate Optimizer.

The test

We build the test with a sole goal in mind: can we create a backend system that would automatically groom and manage ad rates without seller’s participation? To get to the reliable insights, we created the equalized control and the test groups from sellers with nearly identical set of performance and selling velocity characteristics.

Both groups would create their campaigns manually using recommended ad rates (optimal rates at the time of launch). The difference would be in the way the ad rates would be managed during from the point of launch along the entire 30-day duration.

Active management of ad rates

The ad rates in the test group would undergo continuous cross-checking against the competition and are throttled up or down within predefined range. These adjustments would effectively “refresh” the ad rates, keeping them relevant and competitive without overspending.

The results

Ad Rate Optimizer test rates have led to significant performance uplift in the test group. It has since remained as a core functionality for creating a CPA campaign at eBay. These results have also created energy and the momentum for the team to continue to explore machine-assisted advertising solutions in the years since.