How to boost sales and improve ROI using data and machine learning

Online advertising algorithms thrive on a high number of signals. Industries like e-commerce and mobile gaming provide sufficient data to train these algorithms. Therefore, the approaches that are popularly employed by advertising platforms such as Google and Facebook allow efficient targeting and engagement of desired audiences.

However, in many other industries where the number of signals is lower, the target conversion rate is far below the channels’ benchmarks. If you find yourself in this situation, it is vital to make the switch and leverage your data with the power of machine learning.

We have teamed up with Konstantin Bayandin, the founder of the adTech startup Tomi.ai, to discuss online marketing opportunities that can be unlocked with machine learning.

As CMO and Head of Data Science & Analytics Department of a large marketplace I lead the development of data-driven marketing at the company. In e-commerce there are valuable conversions (purchases) galore, and lots of data to slice and dice. Upon relocating to the US to work for Compass, a high-tech real estate brokerage, I discovered a problem. Having switched to real estate from online advertising and e-commerce, I was shocked by how challenging online marketing was for this industry. The main issue was the tremendously low number of final conversions, since people rarely buy houses or apartments. To put it into perspective, the US has a total of just 5 million real estate transactions per year.

Because of this, conversions are extremely rare. Roughly speaking, you are lucky to even get one transaction per 10,000 website visitors. The rate of conversions into transactions amounts to 0.01% or less. Therefore, real estate marketers are unable to employ smart bidding: the automated bidding strategies offered by Google and Facebook’s advertising platforms. They either have to maximize the number of clicks, or optimize for mid-funnel website events (such as contacts collection). The value of these conversions is not always clear, as people often submit their contact info for reasons unrelated to buying.

So this begs the question, why do some industries have a harder time with online marketing? Why do they have low conversion rates and poor ROI? I believe there are five key reasons for this:

1. A highly specific audience: No matter how much effort a marketer puts in, only 5% of the people targeted will buy or sell real estate in the next 12 months.
2. Having offline components: Such as having to sign a real estate contract on paper and in person.
3. Long sales cycles: According to Zillow, it takes an average of 4.5+ months to buy a house. Meanwhile, in the automotive industry, people typically take weeks or months to choose a car, and an additional couple of weeks to complete the transaction.
4. Large transactions: As you can imagine, all of the above can be quite costly. And, naturally, the higher the amount in question, the longer the decision-making process can take.
5. Regulation: Although this is not necessarily a major consideration in all industries, it certainly is for real estate and finance. These industries are heavily regulated in order to secure equal access to housing and credit. For example, the Fair Housing Protection Act bans the use of most targeting involved in real estate marketing, such as gender, age, race, and geolocation.

If your business has at least three of the five components mentioned above, you probably have a low conversion rate. In this case, you need to be very careful with managing your data. Consider implementing machine learning to your advantage: it can help close the feedback loop for your future transactions, by linking them to the demand that exists at the top of the funnel.

Inspired by this challenge, I created a solution that predicts the probability of future purchases using machine learning and website visitors’ behavioral data. Tomi.ai is a platform helping online marketers today more than double their advertising ROI using machine learning, behavioral 1st party data, and API integrations. It opens possibilities for industries where online marketing is a challenge, such as real estate, banking, insurance, EdTech, and SaaS.

For our customers coming from complex industries, we activate predictive scores into their advertising accounts as custom conversions with an expected value amount. That is, we calculate the anticipated income that could be generated by a website visitor, by looking at their level of website interaction. We treat it as a completed purchase, and then maximize this metric. As a result, our customers are able to show their ads to highly engaged users with a high expected lifetime value.

Machine learning is the engine powering today’s online marketing. In my opinion, marketers need to evolve into analysts or data scientists, because marketing problems frequently have mathematical solutions. Online marketing operations essentially resemble an optimization problem in mathematics, where a certain parameter (e.g. revenue) needs to be increased in a situation of budget, channel and other constraints.

The whole purpose of using machine learning in online marketing is to close the feedback loop from purchases and the top of the funnel events.

It is possible to build predictive models across all funnel stages. We do this by predicting the likelihood of conversion based on users’ early behavior. Then we express it in terms of purchase probability and the predicted value of the purchase.

The challenge presented to martech professionals is to think of a way of closing this feedback loop, and creating value-based look-alike audiences based on the LTV produced by such models. They also need to find an efficient way to inform their ad platform bidding decisions using early signals and micro conversions that happen early in the process. Then blend them with machine learning algorithms to create a uniform metric of predicted conversions and sales, rather than using final conversions (which occur rarely).

If Google and Facebook already have machine learning on their side, why would marketers need to implement machine learning on their own?

1. Machine learning broadens the signal

Typically, a low-conversion website will have many user sessions. For example, an average e-commerce project will have a conversion rate of 2%, meaning that only 2 out of 100 sessions will have a nonzero value. Their signal strength is equal to 100%.

Machine learning helps us to understand the likelihood of conversion from early events in each session, and assign a nonzero value to it. We can identify 25% of website visitors that have a signal strength varying from 2% to 10%. This approach gives us a wider audience with a weaker signal that can be used for segments creation.

2. Machine learning is able to calculate the expected incremental LTV

This helps you understand how well the campaign is performing, and whether it will reach its target value in the future – even before the actual conversions take place.

3. dLTV may be used for ad bid optimization

You can optimize your bidding using predictive attribution powered by machine learning to understand your campaign’s expected performance.

Let’s assume you are launching an ad on Google with lots of keywords. You have spent a certain amount on each keyword, achieved a certain sales figure, and divided your sales by the ROI. You have sorted these keywords by arranging them in descending order by their ROI, then plotted 100% of the costs (cumulative) on the X-axis, and plotted all business outcomes in the form of sales or profit on the Y-axis. Having done so, this will give you your ‘performance curve’:

The purpose of using marketing technology is to remove unnecessary controls. In this case, we need to remove poor keywords that are negatively impacting your ROI. You cannot afford to invest in these keywords, therefore their respective budgets can be cut.

By employing new and more efficient methods, you can change the performance curve itself. You can achieve higher performance with the same budget and keywords. In 2016, we used machine learning-powered bid modifiers for search ads and thus helped Google attract higher quality repeat audiences.

As a result, your revenue grows, you earn a higher marginal ROI than your target figure, and it helps you unlock new opportunities. For instance, by investing in new keywords and increasing spending, you can achieve higher performance on your target ROI figure.

Perhaps we are the pioneers of machine learning for predictive marketing in industries such as real estate, banking and SaaS. The very idea of scoring users and using this score for optimization purposes comes from the gaming industry. Mobile games are often free, but come with options to buy in-game. “Whales,” as they are referred to in the gaming industry, are a group of players who comprise only 1% of the player population, yet are capable of paying the cost of all players through their in-game purchases.

All games of this sort are designed to grow its share of high spenders, and most importantly, identify prospective high spenders. For a long time in the industry, product managers have been known to use data to predict LTV in order to inform their decisions.

Mobile gaming apps have already had integrations with Google, and later with Facebook, to transfer online event data. To assess the likelihood of a player becoming a “whale,” marketers use data from a player’s first few days of gameplay. They then produce a “quality install” synthetic conversion, submit this event to Google UAC, and then optimize for quality install, rather than regular install.

We implement the following standard integrations for our customers:

1. Tomi.ai Pixel: You need to install a Tomi.ai pixel on your website to collect behavioral data: such as pages visited by users, length of stay on certain pages, and clicks. The machine learning algorithm will sort through all of these signals to find the ones that provide the most accurate predictions of future conversions. For example, people who really intend to buy a house will use a mortgage calculator and look up schools nearby. These factors will yield the variables that will be used by machine learning to predict the likelihood of conversion into purchase.
2. CRM Integration: We ask customers to provide their completed transaction feeds (historical data). Using hashed emails and phone numbers submitted on the website as anonymous identifiers, we are able to link offline transactions to people’s online activity. We use a one-month (or longer) historical data sample to train our models on people’s behaviors and the likelihood of subsequent purchases.
3. API Integration: Finally, we feed this data back into the advertising accounts. We request access to marketing APIs for us to connect and upload synthetic conversions and predictive audience data to Google and Facebook.

The most important factor for machine learning is the number of positive outcomes that took place – how many conversions we have been able to capture. Theoretically, 20 purchases are enough, but practically, we attempt to pinpoint at least 100 purchases that we can link to online behavior. Generally, we aim to train the model using several hundred or thousand purchases. It is important that we achieve at least the first 100 positive outcomes during our data collection period, which typically occurs in the first month of work (or longer if conversions are scarce).

From our experience, the main criteria for data quality is the share of offline orders that we can link to customers’ online activity. The main idea is not only to integrate the website and the CRM system but to also collect anonymized cookie identifiers. Google Analytics ID is usually used for this. You need to collect users’ Google Analytics IDs and then plug this identifier into the CRM at all subsequent stages: lead qualification, sales support, contract execution, etc.

The value of machine learning does not lie in the data contained in a single table. It lies in the quality of the link between multiple tables.

The value of data is not in the data, but between the data.

This means that the value is not in simply having the tables, but connecting several tables together in order to unlock extra value. Being able to link these online behavior tables with order tables is extremely important for marketers.

If advertising campaigns never progress beyond the learning phase (and this can be the case with extremely rare conversions), you can do something about it with the power of machine learning.

If you are still maximizing the number of clicks on the website, you’re already behind the times. If you are optimizing your ads for contact collection, and these contacts are poor quality (only 5-10% of contacts converted into a sale), machine learning is highly likely to improve that.

I want to emphasize that you don’t have to be a data scientist to get started. Get your feet wet and see what machine learning can do for your marketing campaigns. As industries continue to become more interconnected, now is the time to bring new tech innovations into your toolkit.