Our new ads assistant – SOLD! uses reinforcement learning techniques to optimize Google Ads budget and improve marketing results.
Selecting where to spend money on your Google Ads campaign is possibly as important as choosing what to market. The manual process of managing Google Ads budget and choosing where to best allocate the resources without overspending can be a daily headache for account managers.
Our partner JOT Internet Media is a digital media agency which specializes in Google Ads management. They were looking for a data-driven, machine learning approach to managing their clients’ Google Ads budgets.
Amplify Analytix, a proud winner of the European Data Incubator (EDI) challenge, together with JOT Internet Media developed the AI ads assistant – SOLD!.
Our solution was to optimize Google Ads expenditure by using machine learning algorithms in order to increase CTR and ROI returns and reduce manual data processing time.
Firstly, we collected data on the daily steps an account manager has to take in order to determine the best course of action for spending the Google Ads budget. We then created an algorithm which mimics human learning by utilizing 1000+ daily iterations to produce the best bid options per ad group and campaign. This process is equivalent to the collective learning of 100 account managers working as a team and it improves exponentially over time due to the nature of the algorithm we are using.
We combined ideas and methodologies from the family of reinforcement learning, namely Contextual Multiarmed Bandit to solve for the customer’s need. The results are more sophisticated than the A/B testing mechanism because in CMAB, options are explored and exploited at the same time.
By deploying this algorithm on a historical dataset, the solution goes through a vast number of options as if these were real time decisions. The resulting optimal allocation of investments per option is then cast forward as the best possible investment split for the current budget. The model retrains every day, learning from the decisions and actions it took yesterday and continuously improving forward.
We did an A/B test to confirm the value that the model can bring. The test was running across 2 markets with a total of 4 groups. In two of them, we implemented the bids from the CMAB algorithm and in the other two, the account manager was making decisions on the bids. We used 50/50 random traffic allocation by Google to ensure maximum fairness in the groups.
The chart below shows the difference between the CTRs of the groups (blue line). If the number is positive, it means the algorithm outperforms the account manager, and vice versa. The trendline (orange line) is positive, which means that the algorithm performs better and better over time with a bigger difference between the groups’ CTR. This shows the nature of reinforcement learning very well, as it learns by doing and becomes more accurate with time. In other words, the longer it learns from its actions, the better it will compare against manual decision-making. The best part is, we proved it can replicate the decision making of senior account managers very well!
The result of a month and a half A/B test was approximately 8% higher ROI and 15% CTR uplift within a year. Data crunching was about 38% of an account manager’s time daily, and on-boarding a new member on the knowledge of budget allocation was estimated to be one year.
With manual processing time reduced to a minimum, account managers are now able to focus on the creative and important aspects of the job, reduce onboarding time for new team members while still achieving great returns.