There’s a special type of satisfaction we get when things fall into place. It gives us a sense of calm, of things well done, or maybe even restores hope in humanity if the stakes are high enough.
… finding what you need on the first Google search.
… spotting the jacket you’ve been eyeing on sale (and they have your size!).
… getting a delivery earlier than you expected.
As customers, these experiences are the standard. This is how we judge and evaluate the brands and companies we buy from. When we get to work and switch roles, we might get caught up in our burning priorities and unwittingly forget that. However, our customers never do and KPIs clearly indicate it.
This is why understanding your customers’ context can become an ace up your sleeve to meet their expectations (and your managers’).
To make things even more exciting, you get to work with technology only the biggest companies in the world could afford until recently. You can now leverage machine learning and AI-powered optimizations to take your marketing performance to new heights.
Make subtle improvements with a big impact
Predictions around the next phase of marketing technology converge around a common idea: shopping “will be more digital, but will feel more human” because “technology will be less visible – but far more empowering”.
Using AI doesn’t make a difference for your customers unless their experience is better, faster, more relevant, and more gratifying. This is precisely what we work to help you accomplish. Being able to predict the next shopping stage helps you make it easier for them to convert.
You can achieve this by giving your users:
- customized offers
- personalized discounts
- more relevant recommendations for matching products
- targeted notifications.
What’s more, you can get a better ROI for your marketing budget by targeting users who are more likely to purchase through remarketing campaigns. Here’s how AI is making this entire process easier, faster, and more effective.
Engage users you never even knew wanted to buy
The idea of building customized experiences for online shoppers isn’t new.
However, more often than not, this type of personalization translates into segmentation. This involved identifying common characteristics for groups of customers and targeting them based on those features.
Traditionally, marketing specialists rely on creating a fictional customer profile to define an ideal user who is most likely to convert. They base their assumptions on criteria drawn from research and analytics such as:
- Device used, etc.
Given this approach has become a standard, everyone tends to use similar tactics. The toolbox includes:
- online ads to engage these users
- email notifications to prompt them to finalize their purchase for items added in the cart or marked as favorites
- general discount rates
- rule-based calls to action.
Besides the high competition for customers’ attention, this approach has other, less obvious disadvantages.
It completely misses outliers. These users are equally committed but don’t check the same boxes. These ghost users can become a source of growth for your business if you can find a way to reach them effectively.
What’s more, segmentation has other downsides. This may have happened to you too: customers who fit the same profile will receive the same deal or call-to-action irrespective of their journey through the conversion funnel or value of the future purchase.
Shortcomings like these lead to lost opportunities, wasted budgets, and sometimes even to frustrated users who lost their loyalty or any change of becoming a loyal client. Thankfully, we know AI can fix that. That’s why our approach to segmentation is more nuanced.
Instead of making assumptions about the customers’ personal characteristics, we focus on analyzing their activity and history. We still compare users to identify similar patterns, but we calculate their probability of converting individually. This results in a new criterion for targeting users: converters vs. non-converters.
Here’s how this works.
Spot users who are most likely to convert
If you’ve been working in ecommerce for a while, you’re most likely familiar with what Google Analytics offers. You can generate activity reports or segment users to better understand who is using your website and why.
In addition, in the Enhanced Ecommerce module you can find an overview of your sales funnel.
This funnel – as presented in the Conversions > Ecommerce > Shopping Behaviour section from your GA dashboard – gives you the high-level view. But it doesn’t provide any qualitative information about why some users convert and some don’t.
You can certainly dig deeper into each report. For example, the Audience > User Explorer report is particularly useful to analyze the behavior of individual users. GA identifies them through their browser cookie or user ID, which is available if your e-commerce application has user accounts. But no matter how close you examine your GA data, the challenge remains unsolved:
How can marketers identify those users who are close to converting?
Use the customers’ context to see what they’re going to do next
Our assumption was that the users’ activity and history are the most relevant indicators for their conversion probability.
The learning algorithm behind MorphL predicts the outcome of a user’s browsing session based on information from previous sessions. We extract the data from Google Analytics Reporting API v4. MorphL plugs directly into the data source and the same model can be applied to Google Analytics 360 or BigQuery, depending on your setup.
Going beyond splitting users into converters vs. non-converters, we help you achieve much more specific targeting. This gives you the possibility to engage users depending on the most likely shopping stage they’re about to transition to.
MorphL automatically calculates the probability for each of the following scenarios:
- continue browsing through products
- add a product to the cart
- finalize transaction.
By comparing these probabilities between them, the AI model helps you identify which shopping stage is more likely to happen.
As a result, you can correctly place the user in the corresponding segment. For example, a segment can be “users likely to add a product to the shopping cart”.
What happens behind the scenes
In the background, we divide the data extracted from the Google Analytics API into 3 categories:
- User / Browser– Information about the user, identified in this case by cookie ID. It includes details such as browser, mobile device, etc.
- Sessions – Incorporates information such as shopping stage, session duration, number of transactions and revenue, days between sessions, etc.
- Hits – These are the details about browsed products, products added to the cart, checked out or purchased (product price, name, category, etc.) products and other events from Google Analytics.
Users have a variable number of sessions and sessions have a variable number of hits.
Our model has a recurrent architecture. This means it uses the data in a temporal sequence. The sessions and hits data is organized from oldest to newest and fed into the AI model one at a time.
After processing all the sessions for a particular user, the model calculates the probability for each shopping stage to occur in a future session. So far, we have conducted lengthy experiments with two attribution models:
- Linear attribution modeling where a transaction is equally important for all sessions, regardless of the time passed between the first and last sessions.
- Time decaying attribution modeling which places more value on the latest sessions by applying a half-life weight based on how old the session is.
We described this process extensively in a research paper you can download from our website.
Our real-world AI experiment to predict the next shopping stage
To show you how this practical AI application works in real life, we created an experiment based on data from an ecommerce furniture website. Our experiment data consisted of 24,188 users. Out of these, only 98 users performed a transaction in the last session, with a total revenue of $151,674. This brings us an average transaction value of $1,547.
In real world situations, each targeted user has a customer acquisition cost. For this experiment we assumed it to be $50. Please note we didn’t take into account the gross margin for the calculations below.
The point of the experiment was to evaluate the efficiency of different targeting methods:
- 2 of them based on the machine learning model (with both attribution models) described in the above section
- 1 random
- and 1 using a statistical method.
For each scenario, we calculated the percentage of true positives (correctly targeted users) and false positives (incorrectly targeted users).
The experiment can be described as follows: given each user and their history, we calculated a transaction probability between 0 and 1. If the user was targeted and they made a purchase, then we have a true positive. However, if the user was targeted but didn’t finalize the transaction, we have a false positive.
Here is what the experiment’s results look like:
|Method||Targeted users who made a purchase (TP)||Targeted users who didn’t make a purchase (FP)||Total projected revenue||Targeting costs at $50 / user||Revenue – costs|
|ML, time decay attribution||20.02||135.72||$21,842.74||$7,787||$14,055.74|
|ML, linear attribution||34.04||98.93||$39,524.48||$6,648.5||$32,875.98|
TP = True Positives, targeted users who have made a purchase
FP = False Positives, targeted users who didn’t make a purchase
We can observe that using a random scheme, we get the most true positives (correctly targeted users), but the large cost means that profit is not in the books. On the other hand, the linear model, while targeting a smaller group of users, maintains a low percentage of false positives (0.41%). From these projections, you can clearly see this is the most profitable targeting method.
Random targeting – larger volume, most correctly targeted users, big cost, NO profit compared to spending.
ML, linear attribution – smaller volume, low percentage of users who didn’t make a purchase, small targeting cost, BIG revenue compared to spending.
Using MorphL shopping stage predictions to drive ecommerce performance
Once you have your next shopping stage predictions, there’s a lot you can do to optimize your ecommerce website and marketing flows.
You can become a more relevant and helpful shopping destination for your customers by:
- Displaying pop-ups with personalized calls to actions
- Sending personalized email alerts
- Remarketing the specific users who are more likely to convert and a lot more.
Using an AI-as-a-service platform like MorphL means you can save a lot of time and effort spent pouring over GA reports. Rather than be stuck in reporting meetings, your team can get aligned faster and spend their time putting these insights to good use.
What’s more, MorphL predictions are exposed via an API you can easily integrate into your reporting dashboard. Plus, we fully automate the model training and predictions in our infrastructure.
Here’s how the MorphL integration works in an ecommerce scenario:
- User goes to your ecommerce website
- The website identifies the user by their browser cookie and calls the MorphL API
- MorphL returns the probability of completing a transaction for that particular user
- The website displays a popup with a personalized message, depending on the transaction probability. The popup can include a call to action message, a discount code or any creative trigger you can think of.
There is no doubt that technology is a powerful drive for better customer experiences, no matter which side of the fence we’re on.
What it takes to get there is an open mind and a willingness to experiment with new ways of looking at the data you already have.