Advanced digital marketing forces us to go beyond what everyone else is doing and approach it from new angles. One of the ways to stand out in your SEM analysis and performance is to use advanced techniques like regression analysis. Regression is actually a basic form of machine learning (ML) and a relatively simple mathematical application. This type of analysis can help you make better predictions from your data beyond educated guesses.
Regression might sound scary, but it’s not that advanced in the world of math. For anyone who passed math in 10th grade, you’ve probably worked with the regression formula before. We’ll look at using regression in your Google Ads to predict how much conversion you can get by adjusting campaign spend. Building the model and applying it is much easier than you think!
What is Regression
A regression model is an algorithm that attempts to best fit the data presented. In essence, it’s a line of best fit. It can be linear, like a straight line passing through the data, or non-linear, like an exponential curve, which curves upward. By fitting a curve to the data, you can then make predictions to explain the relationship between a dependent variable and one or more independent variables.
The graph below shows a simple linear regression South Africa Phone Number between an independent variable “cost” (daily spend on Google Ads) on the x-axis and a dependent variable “conversions” (daily conversion volume on Google Ads) on the y axis. We fitted a linear regression line (blue). We can now say that at $3,000 on the axis, that point on the regression line would correspond to 35 conversions. So, based on the data-fitted regression model, if we spend $3,000, we expect to receive 35 conversions.
Headstart on feature selection
I’ve used several of these regression models and will share what I found to be true, which will give you a head start on where to start looking.
Multiple regression involves using certain independent variables (rather than just one, as in the example above), to predict a dependent variable. With Google Ads, I’ve found that there’s always an independent variable that’s the best predictor of conversions. You probably could have guessed which one it is already.
When running ML models on daily labeled training data to predict whether certain features would result in a conversion, we continually found that all things being equal, campaign spend is the strongest predictor of conversion volume.
The following table shows the “root mean squared error” (RMSE) for different ML models.
RMSE is an error measure, it shows how far the fitted model is from the training data. The lower the error, the better – it means the model fits the data more accurately. (2) All features include: day of the week, keyword, CTR, CPC, device, final URL (landing page), ad position and cost.
We ran five different machine learning algorithms: Decision Tree, K Nearest Neighbors, Linear Regression, Random Forest, and Support Vector Regression. In most cases, removing “cost” as a feature in the dataset increased the error value more than removing any other feature. This means that the model has become less accurate in predicting the correct outcome.
We can also analyze the importance of the features used by the random forest (the best model). It is clear that cost is the key characteristic that the algorithm uses to determine its results:
It shouldn’t surprise you: the more you spend, the more likely you are to receive sales. Using cost as a predictor of sales is a great starting point for your regression analysis.