Using machine learning for effective marketing

Marketing - especially digital marketing - is one of the most common use cases for machine learning (ML). Sales, and customer support have also seen a significant increase in machine learning solutions as a result of the recent ecommerce boom.

This is understandable as the core functions of machine learning - clustering and predictive models, and optimization algorithms - are well suited to the core tasks of marketing. Finding new customers, recommending new products and services, and controlling media budget efficiency are almost textbook examples of machine learning in action.

Yet the use of machine learning in marketing has not advanced as quickly as one might expect. More than 50% of marketers consider machine learning crucial to their future success in the industry, yet two thirds of marketers say they are only in the research stage for using ML. Only a remarkably low share of 17% of companies report that they have moved from piloting machine learning to scaling up its use.

Why does it take so much time to get moving with machine learning?

The unbearable uncertainty of machine learning (ML)

Surveys on the topic tend to revolve around three factors:

Lack of know-how

In an international survey on the topic, more than 50% of marketing professionals say they themselves are only beginning to understand how to utilize machine learning. A mere 30% of respondents believe they are average in their understanding. It is hardly surprising then, that experimentation remains low, if people don’t feel comfortable with the general principles.

This personal lack of familiarity may also contribute to the difficulty of finding partners with the needed skills - also a recurring theme. It is hard to validate the level of skills of someone claiming to be an expert in something you yourself don’t fully understand. In addition to this, high quality data scientists are quite few compared to the potential amount of work it would take to bring ML to marketing across the board.

While machine learning is no longer as mysterious as it once was, it does still require non-ubiquitous, specialized skills.

The lack of reliable data

The lack of reliable data is a problem somewhat particular to marketing, and this tends to cause some key issues for the use of ML.

Cookie based analytics causes data quality issues, and gaps in the data you have, even when best practices are being followed. But while well known, this is not the only big issue with reliable data in marketing.

The available toolset for digital marketing is overabundant, and combining the use of various tools for a common goal will inevitably lead to some issues with using the resulting data. Additionally the lack of individualizable data from the two largest players, Google and Facebook (Alphabet and Meta) is an issue for marketers who want to tap into all available data to run machine learning with.

Uncertainty of profits

The third major challenge in developing ML solutions for marketing is the uncertainty inherent in calculating returns on the investment. Problems that will benefit from machine learning are almost necessarily complex, and thus difficult to predict in advance. Both results, and the time needed are difficult to predict. This uncertainty is a large roadblock between large scale use of machine learning in marketing.

This issue isn’t, however, entirely unavoidable. Part of the problem of profitability for machine learning projects is the typical approach through large scale projects, and big investments. Starting out with large projects is inherently even more risky, as failure in large projects may be impactful for the company as a whole. This limits the amount of companies that are willing to participate in the implementation of such projects.

How to speed things up?

We Lamiateers believe that machine learning should be systematically demystified. Despite it having some complex mathematics behind it, utilizing machine learning can be done without delving too deep into its weeds and nuances.

The easiest way to bring machine learning within the reach of marketers is to get started through smaller experiments to build up confidence, and familiarity with the use of machine learning. Only then should one advance toward more complex use cases and larger scale implementation.

During the course of 2022 we will publish a series of examples of machine learning solutions through which you can create a bigger impact through marketing. But while you wait for the examples, here are a few things to consider on your quest to utilize ML.

Analytical data structures

Using machine learning in online commerce requires data that can be tied to a specific browser (client) us person (user). Google Analytics 4, Adobe Analytics, as well as several smaller analytics tools have this built in. The most commonly used Universal Analytics from Google will only allow this data to be used via API in the basic setup.

It is possible to work around the limitations, however. You don’t need to start working through API queries if that’s unfamiliar to you. But be sure to implement some solution to get granular data to work with if you want to benefit from ML.

Combining analytics and customer data

Analytics data can’t provide granular data about demographics, or interests. While this data can be provided on an aggregated level, you won’t get access to this information as it relates to individual users.

This problem can, however, also be worked around. Joining analytics data of website use to CRM - or other customer data - can be done as long as you have a key to match the two sources of data. Transaction ID’s can allow you to combine analytics data to your own database of existing customers.

While joining these sets of data are quite insightful in and of itself, for machine learning this is particularly interesting as it will allow you to look for patterns in website use that let you fill in the blanks for website users you haven’t previously identified.

You can test machine learning models like anything else on your site

The special knowledge you need to have in order to use machine learning models isn’t as high as it used to be. The improved speed at which models can be built allows for testing models on a small scale quickly to validate - or discard the approach entirely.

Assuming your data models allow you to build the necessary models, you can attach the predictions and models quite easily to your analytics, and ad platforms, and use the predictions as a type of AB test you would run on your site anyway. Failing fast can be useful in machine learning as well.

Even small scale experiments can move you forward

Even if you don’t succeed at driving mindblowing growth through your first experiments with machine learning, they will likely still be worth it if they help you learn more about the use of ML.

There is no better way of learning than through practical experimentation - trial and error. Failed experiments will bring in valuable learnings about an important topic sure to increase in importance. If the experimentation has not been too expensive, it is likely to repay the investment in the long haul.

Being proactive in learning the approaches that won’t work long term will also allow you to position yourself as an early adopter. All companies need to learn what works and what doesn’t - being proactive shouldn’t be a bad thing.

Just running small experiments isn’t enough

Just because we believe machine learning can be utilized as small experiments, we by no means consider this the ultimate goal. While the experimentation approach is an important tool to get moving on your machine learning journey without taking on huge risks to start, there will be limits on how far it can take you unless you follow up properly.

If we are to use machine learning as experiments similar to AB-testing, the analogy needs to be extended beyond the initial experiment. After running an A/B test on - let’s say - Google Optimize, you would not just let the winning variant run on your site through the test platform. Neither should you try to use experimentational light machine learning solutions in the long run.

You should consider early on how to implement your experimental models into production environments when you find solutions that work. The best part is that the larger implementation can be more straightforward once you have experimental validation to tell you what needs to be done.

Machine learning in practice -series

This blog is only meant to be our opening salvo on the issue of machine learning in marketing. We intend over time to present practicable solutions on how to experiment with machine learning in digital marketing without large up-front investments. We already have some ideas in mind, but we would like to hear what you think. What are some of the problems you would like to try machine learning to solve? Get in touch with me to discuss!