Table of contents
Machine learning in digital marketing: Examples of use cases
Vlada Malysheva, Creative Writer at OWOX BI
Vadim Onyshchenko, Digital Analyst at OWOX BI
A couple of decades ago, the first thing that came to mind when you heard the words “artificial intelligence” was likely the rise of the machines and the Terminator with a sawed-off shotgun. Today, this term has rather positive associations. Almost everyone encounters machine learning in ordinary life. For example, you might communicate with a chatbot on a website, be shown promotional offers that correspond to your hobbies, or set up a spam filter in your email service.
For marketers, machine learning is an opportunity to quickly make crucial decisions based on big data. In this article, we’ll talk about what decisions you can make based on big data.
What is machine learning?
Let’s start with a little terminology. According to Wikipedia, machine learning (ML) is a class of artificial intelligence methods characterized by their not providing direct solutions to problems but rather training systems to apply solutions.
There are many methods of machine learning, but they can roughly be divided into two groups: learning with a teacher and learning without a teacher.
In the case of learning with a teacher, a person supplies the machine with initial data in the form of situation–solution pairs. The machine learning system then analyzes these pairs and learns to classify situations based on known solutions. For example, a system can learn when to mark incoming messages as spam.
In the case of learning without a teacher, the machine receives unsorted information — situations — without solutions and learns to classify those situations based on similar or different signs without human guidance.
Machine learning in online marketing
Marketers use machine learning to find patterns in user activities on a website. This helps them predict the further behavior of users and quickly optimize advertising offers.
What is the potential of behavioral data?
In psychology, a pattern is a particular set of behavioral reactions or a common sequence of actions. Therefore, we can talk about patterns with regard to any area where people use templates (which is most areas of life).
Consider the example of a pattern used on websites. If the user isn’t interested in the offer in the pop-up window shown below, they can close this window by:
- clicking on the X sign
- clicking No thanks
- clicking anywhere on the site that’s outside the pop-up window.
In addition to these three actions the user can take, the pop-up window will close on its own after a certain period of time.
So we get four possible user actions:
- Click X — Can be true/false
- Click No thanks — Can be true/false
- Click past the pop-up — Can be true/false
- Pop-up viewing time is 5 seconds
When hundreds of such parameters are collected, the collected data gains value because it contains patterns of behavior and dependencies. It hides the enormous potential of behavioral data, allowing us to supplement user data with the missing parameters based on the data we already have for other users.
For example, the simplest way to define a target audience is by gender and age. But what if users fill out this data only in 10% of cases? How can you understand how many of your website users fall into your target audience? Patterns of behavior can help.
You can use gender and age data from 10% of users to determine patterns specific to a particular gender and age. Then you can use these patterns to predict the gender and age of the remaining 90% of users.
Having complete data about gender and age, you can now make personalized offers to all website visitors.
The effectiveness of marketing depends on the quality of the data on which the models are trained. Attribution model based on only company's data loses to models built using market data. OWOX BI uses data from tens of thousands of projects to make the machine learning model take this knowledge into account in predictive functions. Book a demo to learn more about how OWOX BI and ML can boost your company's profits.
Why machine learning is effective in marketing
The role of machine learning in marketing is to allow you to quickly make decisions based on big data.
The algorithm for the work of marketers is as follows: Marketers create hypotheses, test them, evaluate them, and analyze them. This work is long and labor-intensive, and sometimes the results are incorrect because information changes every second.
For example, to evaluate 20 advertising campaigns considering 10 behavioral parameters for five different segments, a marketer will need about four hours. If such an analysis is carried out every day, then the marketer will spend precisely half their time assessing the quality of campaigns. When machine learning is used, evaluation takes minutes, and the number of segments and behavior parameters is unlimited.
With machine learning, you can respond faster to changes in the quality of traffic brought by advertising campaigns. As a result, you can devote more time to creating hypotheses rather than to carrying out routine actions.
The value of your results depends on the relevance of the data on which the analysis was conducted. As data becomes obsolete, its value decreases. A person simply can’t process the volumes of information that are collected every minute by analytical systems. Machine learning systems can process hundreds of requests, organize them, and provide results in the form of a ready answer to a question.
Key benefits of machine learning in marketing:
- Improves the quality of data analysis
- Enables you to analyze more data in less time
- Adapts to changes and new data
- Allows you to automate marketing processes and avoid routine work
- Does all of the above quickly
Examples of machine learning in marketing
1. Recommendation systems
The essence of a recommendation system is to offer customers products they’re interested in at the moment.
What a recommendation system predicts: Goods that a customer is likely to buy.
How this data is used: To generate email and push notifications as well as “Recommended products” and “Similar products” blocks on a website.
Result: Users see personalized offers, increasing the likelihood of their making a purchase.
Common algorithms for this purpose: K-means clustering.
2. Forecast targeting
In general, the essence of all types of targeting is to spend the advertising budget only on target users.
Most used types of targeting:
- Segment targeting — Show ads to groups of users with the same set of attributes
- Trigger targeting — Show ads to users after they take a certain action (for example, viewing a product or adding an item to the shopping cart)
There’s also predictive targeting, in which you show ads to users based on the likelihood of their making a purchase.
The main difference between these types of targeting is that predictive targeting uses all possible combinations of tens or hundreds of user parameters with all possible values. All other types of targeting rely on a limited number of parameters with certain ranges of values.
What forecast targeting predicts: The probability that a user will make a purchase in n days.
How this data is used:
Example 1: To launch advertising campaigns. For this purpose, create segments based on the probability of a purchase and upload those segments to Google Ads, Facebook Ads, and other advertising systems.
OWOX BI automatically imports audiences from Google BigQuery into advertising services. This allows you to automatically create, update, and upload audiences to ad services. Manage data-based bids, boost ROI and conversions, and save your advertising budget.
Example 2: To analyze the effectiveness of advertising campaigns. For this purpose, create segments based on the probability of a purchase and upload those segments to Google Analytics and use them to analyze the effectiveness of advertising campaigns (which campaign leads to the most conversions).
Result: Advertising is shown to a more targeted audience, increasing the effectiveness of campaigns.
3. LTV forecasting
The best-known methods of calculating lifetime value, or LTV, are based on knowledge of the total profit from a customer and the time for which the customer has been interacting with the business. However, many modern business tasks require you to calculate LTV even before a customer leaves. In this case, the only solution is to predict LTV based on available data.
What LTV forecasting predicts: The LTV of each user by segment.
How this data is used:
- Segments are loaded into push notification or email services and used for mailing to reduce customer outflows (the churn rate).
- Segments are uploaded to Google Analytics and used to analyze the effectiveness of advertising campaigns based on predicted LTV.
Result: The advertising budget per user is determined based on LTV, which improves the effectiveness of campaigns.
4. Churn rate forecasting
In marketing, the concept of churn or outflow refers to customers who have left the company and the associated loss of revenue and is usually expressed in percentage or monetary terms.
Churn rate forecasting allows you to respond to a customer’s intention to abandon your product or service before they actually do.
What churn rate forecasting predicts: The probability of users leaving by user segment
How this data is used: Segments can be uploaded to email or push notification services as well as to Google Ads, Facebook Ads, and other advertising systems. You can also pass this information to the retention department so they can personally reach out to customers with a high probability of leaving.
Result: Retain customers.
How OWOX BI uses machine learning
OWOX BI has developed our ML-based solution, which calculates the probability of purchasing considering the purchased orders for each website user. Based on this calculation, you can create audiences, use them to target advertising campaigns, and increase ROI by two times, as one of our customers did.
Features of ML segmentation from OWOX BI
The OWOX BI model can be trained on data from different sources: CRM, websites, and mobile applications. Our solution allows selecting any targeted action: transactions, purchased goods, phone calls, adding goods to the cart, etc.
You can also set any period relevant to your business as a conversion window, depending on the timing of the purchase decision.
You can use the calculation results in different advertising services: Google Ads, Facebook, Instagram, etc.
You can use conversion prediction results to:
- Bid adjustments in contextual search ads.
- Targeting in media campaigns.
- Exceptions in media campaigns.
- Increase incoming new traffic, buy more widely and cheaply, and adjust after user assessment.
Book a demo to learn more about ML segmentation opportunities for your business.
OWOX Predicted Conversion Lifts attribution
In OWOX BI, you can connect any standard attribution model to your reporting. Also, our analysts can configure a data-driven model based on conversion forecasting from OWOX, or a custom model for your rules and your sales funnel.
You can look at campaigns from different angles, compare the results of calculations on several attribution models and choose one that meets your goals.
Evaluate channel contributions to ongoing and future conversions with OWOX Predicted Conversion Lifts
The Predicted Conversion Lifts attribution model is based on ML and shows the incremental contribution of each channel and campaign to the sale. Thanks to this, you can immediately analyze the launched campaigns and not wait a month or two to draw a full conclusion, even if the sale takes place only in the future.
For example, you have launched campaigns that work on the top or middle of the funnel. And you need quick feedback on the impact of these campaigns on sales over the past week. This assessment can be carried out with Conversion Lifts as early as the first week after launch. You purchased traffic today. It has yet to bring in sales but already has a predictive estimate of the contribution to future sales.
If you want to see how OWOX BI attribution works, sign up for a demo. Our colleagues will show you real examples of how to apply attribution and demonstrate how it can be useful for your business.
Machine learning in attribution
Why is machine learning needed and how does it help you solve the attribution problem? This is a topic for a separate article (which we’re already preparing).
In this article, let’s figure out at what level decisions are made using attribution. We’ll compare these levels based on several criteria:
- The level itself
- Key decision-makers
- Types of decisions made
- Tools used
- Attribution models most often employed
Levels at which attribution-based decisions are made:
1. Vision. A company’s vision is formed by its board of directors, CEO, and marketing director. Vision-level decisions are associated with investments in the brand and balancing budgets between online and offline. The tools used to make such decisions are market research and consultants. There’s rarely a place for classical data-based attribution models, since the data on which decisions are based is not digitized enough.
2. Strategic. Strategic decisions are made monthly, as a rule, by marketing and ecommerce directors. These decisions are devoted to the allocation of the budget between channels and the definition of the top-level KPI. The tools that help in making strategic decisions are Visual IQ, OWOX BI, or casual models. Here, the business uses data-driven attribution, variations on the theme of the Shapley value and Markov chains, or funnel based attribution. At this level, it’s important to understand the mutual influence of channels and make strategic decisions on their development.
Compare the pros and cons of the best-known attribution models, from standard models to Google Analytics, Markov chains, and the Shapley value.
3. Tactical. Usually, tactical decisions are made weekly or even more frequently by the manager of paid traffic acquisition. Budget allocation occurs between campaigns and ad sets, and decisions are aimed at clarifying KPI and campaign goals. For making tactical decisions, you can use Google Sheets or OWOX BI. Often at this level, specialists work with Google Analytics. For assessing the impact of media advertising, most attribution models use associated conversions, time decay, and post-view.
The peculiarity of this stage is that the budget for a channel has already been allocated. So at this point, it’s important to understand what campaigns to spend it on, control the results, and quickly turn off inefficient campaigns.
4. Execution. This is when the decision to evaluate the contribution of a particular announcement or keyword occurs in close to real time. Such decisions are typically made inside advertising services (Google Ads, Facebook Ads). In fact, the customer doesn’t care what optimization mechanisms are used here, as they look at the results of each service separately.
As you can see, machine learning is most useful for strategic and tactical tasks. Sometimes it’s also applied at the execution level, but the general trend is that advertising systems develop fast and have a lot of data. The internal algorithms used in these systems to manage advertising campaigns produce better results than an external model based on machine learning.
The reason is that in order to apply machine learning, it’s necessary to export large amounts of data from the advertising service quickly and then quickly import results back. Technically, this is a difficult task to solve on an industrial scale. Therefore, at the execution level, marketers tend to rely on internal algorithms for optimizing advertising services.
To use machine learning to solve tactical and strategic problems, you need to ensure the completeness of your data. You can do this with OWOX BI. OWOX combines your data from your website, advertising services, and CRM so you can create a funnel that takes into account the peculiarities and efforts of your business and is aimed at attracting customers and growing sales.
What is machine learning?According to Wikipedia, machine learning (ML) is a class of artificial intelligence methods characterized by their not providing direct solutions to problems but rather training systems to apply solutions.
Why machine learning is effective in marketing?The role of machine learning in marketing is to allow you to quickly make decisions based on big data. With machine learning, you can respond faster to changes in the quality of traffic brought by advertising campaigns. As a result, you can devote more time to creating hypotheses rather than to carrying out routine actions.
Examples of machine learning in marketing1. Recommendation systems
2. Forecast targeting
3. LTV forecasting
4. Churn rate forecasting