Impacts of Learning Mode in Amazon’s A9 Algorithm

Few search engine algorithms have as much influence in determining business success as Amazon’s A9 system. With a US eComm market share estimated between 40% and 50% (Walmart in second with only 7%), it is possible (though not always recommended) to build a successful business that exists solely on Amazon. Like many search algorithms, A9 works to match users to products that they are most likely to purchase. There could be a whole book describing how A9 works, but this post is going to focus on one critical feature that can go a long way towards determining advertising performance: Learning Mode in advertising campaigns.

We have already seen the impacts of an advanced machine-learning based system in Google Smart Shopping, and outside of advertising performance, it is evident that these systems do tend to provide a better user experience. As such, A9 has been transitioning to such a system over the past year without making it as clear as Google that they are doing so – matching their mission of customer (not advertiser) obsession.

What is Learning Mode?

In Amazon’s A9 machine-learning algorithm, the system tracks how certain user types interact with specific SKUs/ASINs (Amazon Standard Identification Number), attaching these learnings to that ASIN over time. With this information on how different user signal types interact with a specific ASIN, the system determines the future users that are most likely to be interested in the product in question. The product’s ranking in the search results is a direct reflection of this likelihood.

In Amazon campaigns, these learnings are attached to both the product and the campaign. This is especially true in Sponsored Brand placements, which are not accessible in a non-paid format. Any time a campaign is completely paused and then reactivated, the system enters a “learning period” where it needs to figure out which types of users best match the campaign settings and advertised ASINs. The result of this is decreased efficiency in ad spend until the algorithm settles in. While Google’s Smart Shopping indicates that this is happening via a “Learning Mode” tag attached to the Campaign Status column, A9 does not notify you of being in learning mode.

What does Learning Mode Look Like?

Let’s take a look at an example. Above is an image of an advertiser that uses a weekly budget from Monday to Sunday – this budget tends to cap out and shut off many campaigns each Sunday night. These results are for May 2020, with the blue line representing advertising cost of sale (ACOS) and the orange line showing spend by day. Campaign budgets were refreshed this May on Mondays 5/4, 5/11, 5/18, and Tuesday 5/26 after Memorial Day weekend. As indicated on the chart, when campaigns are reactivated each week, we typically see a higher ACOS compared to the previous active day of post-learning performance (averaging +7.38% in May). To further underscore this: in February 2020, prior to instituting weekly budgets, the same account averaged a return on ad spend (ROAS) of 4.50 on Mondays and a 3.96 on Tuesdays, with a monthly return of 3.98. In May 2020 (with weekly pacing and Learning Modes) ROAS for the same days were 3.60 and 3.89 respectively, with a monthly return of 4.37.

In short, triggering a Learning Mode each week has resulted in two higher performing days becoming the two least efficient days each week. Daily Return Charts illustrate this well:

February – Pre Weekly Campaign Caps + Learning Mode

May – Weekly Campaign Caps + Learning Mode

Key Takeaways

Though not formally acknowledged in the advertising platform, learning periods in ad campaigns are real and have a measurable impact. As Amazon continues to update A9 to incorporate more machine algorithm features, accounting for Learning Mode will become even more important. Plan with these in mind and you will be able to deliver even more efficient performance for all Amazon campaigns! Contact us if you’re interested in learning more about Amazon advertising.