If we were to sum up the digital marketing buzzwords for 2019, there is no doubt that machine learning (ML) would be placed in the top three, and for good reason; machine learning can be used to solve problems more efficiently and more effectively than humans alone, which means increased performance and reduced costs for businesses.

While the technology is still in its infancy, there are already ways we can use ML to improve digital performance. In this blog, we’ll run through some of the ways brands are using machine learning to drive results, and three algorithms we have built to provide solutions for our clients.

What is machine learning?

Machine learning and artificial intelligence (AI) are often used interchangeably, but they are not the same thing. AI is a broad concept referring to machines which uses algorithms to complete a task in an ‘intelligent’ way.

Machine learning is a subset of AI – so while all machine learning is artificially intelligent, not all AI is machine learning. Machine learning involves supplying a machine with data which it then uses to teach itself an algorithm by comparing the data to find patterns and trends. The machine continually learns, and as it receives new data it updates the algorithm to reflect changing trends.

Why is ML being hailed as the next big thing?

While artificial intelligence has become a hot topic in mainstream media, the machine learning element has become more enticing for marketers, and for good reason. We can use ML to build bespoke marketing solutions more quickly and more efficiently than ever.

Machines can analyze huge data sets and identify patterns much quicker than any human could. This means we can solve problems more efficiently and implement solutions quicker, for better results. By eliminating the need for data analysis to be conducted by marketers, our time is freed up to be spent on more strategic activity.

How is machine learning being used in digital marketing?

Many brands are already using machine learning in one way or another to improve their digital efforts. For example, Netflix use machine learning to personalize each individual account – not only customizing what programs to watch using your past activity but going as far as personalizing the images used to increase the likelihood you decide to watch. Spotify, Pinterest and App Store and Play Store also all use machine learning to make better recommendations.

Chat bots are another way that brands are using machine learning to deliver better customer experiences. Machine learning chatbots collect data from each conversation had and use this to provide more relevant and accurate responses to customer queries than if they were to only use pre-programmed auto responses.

Smart bidding

Smart bidding is a type of automated bidding strategy which uses machine learning to improve the performance of paid media campaigns. The machine uses historical campaign data to see how a wide range of signals impact performance and adjusts bids at each auction based on these signals. Google Ads, Search Ads 360 and Display & Video 360 all allow for smart bidding, and these strategies have been shown to improve performance over manual bidding.

Three ways we have used machine learning so far

While machine learning solutions can be built for almost any problem, there is one prerequisite – we need an abundance of high-quality data in order for the machines to build algorithms which are reliable. This often means brands need to go through the process of setting up data capture, cleansing existing data, and bringing multiple data sources into one central infrastructure, before it is feasible to build a machine learning tool.

As well as helping our clients go through this data cleansing process, we have built several ML tools already. We will run through three of them in this blog.

Content Relevancy Checker

We built a new ML tool which uses Google’s Universal Sentence Encoder to measure how relevant webpage content is to the key search terms being targeted.

While there are many content checking tools out there, none take into consideration the visual cues used by humans and search engines to understand web content. We trained a TensorFlow deep learning model to predict the likelihood of a given text-block forming part of the main page content, factoring in multiple features about the text block (such as size, page location, text size and text density) and the webpage layout itself. We created a huge labeled data set of text blocks and their corresponding page images by getting the Search Laboratory team to look over page images and classify whether a text block formed the main webpage content or not.

You can read more about how we built the Content Relevancy Checker here, as well as try out the tool on your own webpages for free.

Using ML to predict the value of an online lead

One of our clients generates leads via their website which are then nurtured offline until sale. There are several types of leads generated including brochure requests and showroom bookings, and the final sale value of the lead understandably varies depending on the type of request, the customer demographic, and the online behavior leading up to the lead.

The lead time between a visitor submitting an inquiry and making a purchase was several weeks. Not only does this make using actual sales to target auto-bidding less effective, it meant we were unable to respond quickly to campaigns that were driving low quality leads.

We needed to find a way of predicting the value of a lead in real-time so that we could use this information within auto-bidding technologies. To do this, we trained a machine learning model to value form leads based on the eventual sale value from historical form submissions:

We analyzed historical sales events and the form submissions that preceded them using a BigQuery machine regression model. This model took into account both form features (such as store location, contact type, etc.) and the user’s onsite behavior prior to the form fill

We then used the trained model to estimate the value of new forms as they are submitted by taking the associated data and passing it through the trained model. This returns a predicted eventual sale value for the form fill

The value is sent back into Google Analytics using BigQuery’s measurement protocol where it can be used to feed into the client’s auto-bidding platform and create audiences to value post-inquiry, but pre-sale, users

The model is retrained daily using the newly available data.

Using machine learning to value website users

The same client also required us to be able to better predict the ‘quality’ of users (in terms of how likely they were to eventually convert). A typical user would visit the website several times before submitting an inquiry, and not all visitors would lead to leads (or sales). It was important that we were able to prioritize the remarketing budget towards the most valuable visitors with the highest chance of converting.

We were able to use ML to identify users that are most interested in the product so we can place them in suitable audiences for retargeting.

We exported Google Analytics 360 (GA360) data into BigQuery and then designed and built a custom recurrent neural network and trained it using this dataset. This trained a model to ‘score’ all sessions based on all behavior events and pageviews for each user in the current and previous sessions, and the meta data for those events and pages. All sessions were scored even if they did not lead to an inquiry.

After a session is ‘scored’, we pushed the event back into Google Analytics along with custom dimensions and metrics describing the probability to submit a form in the future and the likely value if so. This is pushed back into GA within four hours of the session, meaning we hit timestamp to associate the event as part of the original session and can get a session scoped dimension against that session, as well as the place metrics in the session ‘score’.

We are now able to create audiences for the different tiers of website users based on how likely they are to convert and users with a higher chance of buying can be targeted more aggressively. This is a more efficient and effective way of using the client’s remarketing budget and means we can deliver a higher ROI.

Getting started with machine learning

Building a ML model requires both data scientists who are skilled at finding insights in data, and developers or engineers who are skilled in writing algorithms and engineering a solution that enables an automatic use for the output from a ML model. If you want to invest in machine learning, ensure you have team members who have these skills, as well as the ability to work together to create a machine learning tool, is crucial. There are many tools, such as TensorFlow, available for developers to practice creating machine learning models – take advantage of these! Play around with creating different models and designing solutions for different problems.

Machine learning can be a great tool to improve your digital marketing efforts, however a ML model is only as good as the data it’s trained on so the first step to using machine learning in your digital marketing strategy is ensuring that you have access to high-quality data, and lots of it. This means that there are certain steps to take before creating a machine learning model, such as:

Setting up tags to correctly track and capture onsite user behavior

Joining up online and offline behavior by linking up analytics and CRM data

De-duplicating data across multiple sources

Housing all data from different sources in one central place such as BigQuery.

Once your data is in order, you are in a good place to start building machine learning solutions to any big data problems you have. If you are interested in using machine learning to improve your digital performance and would like help with data capture, analysis or developing a machine learning model, get in touch.