PPC Machine Learning

Adwords management requires making decisions with limited data. A keyword might have 100 visitors and conversion rate of 20%. The campaign as a whole might have an overall conversion rate of 10%. This means that in the next 12 months, the conversion rate of the keyword will probably be between 10 and 20 percent.

Correctly predicting the conversion rate is important. We use machine learning to predict the future conversion rate for this keyword, and set the keyword bid to maximize revenue and minimize loss as defined by a Mean square error (MSE) loss calculation.

Machine Learning and Polynomial Regression
The fundamental challenge we are solving using ML is how much “trust” to give to datasets based on the type of data, and sample size. The best way to do this is to create a polynomial that maps levels of “trust” with the size of the sample set. Essentially, the larger the sample set, the more we trust it.

A system to minimize loss is not linear. As such, we use a regression model for to find a best fit line thru the data points. This best fit line minimizes the loss associated with either trusting data of an arbitrary sample size either too much, or too little. Once we have use machine learning to solve for the best fit line, we can use it to calculate how much to trust a specific dataset for any of our client based on the size of the dataset. More information about polynomial regression is available here.

The actual bid/bid adjustment computation turns into simple high school algebra after we have calculated how much to trust each dataset used in our keyword bid and bid adjustment calculations.

We aggregate the data sets of all our customers. We use this aggregate data as we calculate the Tensors we use in our ML calculations. This approach allows for the best possible predictions for clients with less the $100k per month on Adwords. For clients with over $100k per month in spending, we create custom ML calculations, and use these custom solutions in our Adwords management.

As a simple example, clients with a strong focus on a specific geographical location would “trust” datasets that show conversion rate data for a specific location more then companies with a nation-wide focus. This is primarily because any geo related anomalies in the latter company are most likely just that – statistical anomalies which should not be heavily trusted. Conversely, if a company is located in a specific geographical location, then we need to carefully look at, and trust data that shows location specific conversion rate data.

How much does it cost?
Because of our automation, the system is suprisingly afordable. The setup cost is $500. Monthly packages start at just $250/month.

Reporting
We provide monthly reports that show all changes we made to your account over the last month. Our reports also show cost/conversion trends over the last 12 months.

Machine learning (ML) algorithms build a mathematical model based on training data. The model allows the system to make predictions without being explicitly programmed to perform the task. Machine learning measures the error in any predictions as loss, and determines how to minimize loss.

In Adwords, “loss” can be defined as any keyword bid, or keyword bid adjustment that is either too high, or too low. If a keyword bid is too high, the loss is “wasted money”. If the keyword bid is too low, then the loss is “lost conversions”. The challenge is getting the keyword bids and bid adjustments just right. This is where machine learning comes in. You can learn more about ML training and loss here.

Why is this technology useful?
This system is useful for two main reasons.
1. We can make better decisions. We can figure out the correct bid for each and every keyword.
2. We can make faster decisions. We can calculate the correct bid adjustment for every city in the entire world, and we can update the bid adjustment every day.

This technology is a game changer.