To make predictions, you need either some ancient and divine wisdom or well-prepared data – and, at least, data we have in spades. Then, you need an event to predict. However, this needs to be related to the web analytics data you have, such as a conversion – not the next World Cup winner. Other useful predictions in web analytics are the return and churn probability of users: Will they be back or are they gone for good? Furthermore, the next basket value: If a user makes a purchase, what will be their basket value? Finally, the prediction of the customer life time value. This is the value of all conversions of a customer within the next 12 months - past purchases are not included in this prediction and it is rather hard to do.
Now we know what to predict, we need the data. However, handle it with care – if you torture it long enough, it will tell you anything! To make it tell the truth, we divide it into training and evaluation sets. The training set is used to create a model that makes the prediction, while the evaluation set is used to test the result of the model versus the truth. Before I dive deeper, a short disclaimer: This is not a data science course but an explanation to give you a conceptual idea of what is happening behind the scenes when predictions are made.
You can imagine a model in data science as a formula, like 2x+3y/4z = Prediction. The creation of the model is the hard part: You try to come up with a formula that produces most often the correct result when you replace the variables with your data. In web analytics, the variables can be something like the number of page impressions, session duration, referrer, browser, product views, past purchases, and so on. These are also called features of the model.
Whenever you have developed a model based on the training data that seems promising, you try it out on the evaluation data. You put the variables for each user into the evaluation set and compare the result with the truth you already know, that is e.g. whether the user converted or not. A model, however, is not really a formula like shown above, it is more complex, and the features are not just added or subtracted. But the important part is that you can enter data points into a model and get a result. Once the evaluation phase shows good results, you got a model that can be applied to unknown users.
Predictions – or how to figure out valuable users
The first level in web analytics is to figure out what happened on your page. For the second level, you figure out why things are happening and that’s where most companies are. The next level are predictions - figuring out what will happen and that’s what we are currently working on. We have some exciting news to share - We now provide predictions on:
•Return and churn probability (user will or will not come back within 90 days)
•Next basket value (if a user converts, how high will be his basket value, regardless of conversion probability)
•Customer Lifetime Value
These probabilities are based upon analytics data and can be used to create segments for analysis and for exporting segments.
In addition, Webtrekk also provides real-time predictions which are based upon in-session data and calculated while the user is active on the website. The real-time predictions are available in Webtrekk’s Marketing Automation tool to trigger actions, e.g. adding users to remarketing lists, or showing them layers to offer goodies, ask for contact details, etc. Real-time predictions are currently in private beta mode and we are going to test them with a small group of customers. If you are interested, get in touch with your consultant.
After completing the first round of case studies with our customers, we are excited to present what you can do with these predictions.
Use case: Mailing coupons
A travel vendor wants to remind his users of their services by sending them gift coupon of 10€. But is the coupon really necessary, if the user is likely to convert anyway? To find out if this is true, the following experiment took place: A group of users with a high likelihood to convert was divided into two groups. One group got a reminder email with the gift coupon of 10€. The other group got just a reminder email. Which group has the higher conversion rate? Please place your bets now.
It turned out that they are virtually the same! Once users are likely to convert, the additional incentive did not contribute to increasing the conversion rate in this group. It is not necessary to send out expensive gift coupons to all customers, just to the small segment where it actually makes a difference. With Webtrekk predictions, you can save money and make better, data-driven decisions.
Use Case: Efficient (Re)Marketing
Remarketing can be expensive, but predictions help you to make the most out of your remarketing budget. Instead of hitting every visitor with your Google AdWords ads, create smart lists that distinguish users by the metrics that matter most – conversion probabilities and return probabilities. This will save you a lot of time: You don’t need to create proxy segments for which you have calculated (or, let’s be honest, intuited) success rates, so you can concentrate on optimizing the bidding for the predictive segments.
1. For loyal customers, who have a very high return probability, you can stop bidding on your brand. They will click your first organically ranked page anyway.
2. Users with very low conversion probability can either be excluded or put on a list with small bids. No need to spend money on them.
3. Users with high conversion probability get on the list with a high bid.
4. Users who are about to convert get on the list with highest bid.
With predictions you can create lookalike audiences provided by Facebook and AdWords, for example, based on your list of users with high conversion probability. You are probably already working with lookalikes for buyers (if not, then do it now!), however, this is usually a small list – at least compared to the high conversion probability list. A bigger list means that more candidates will be matched, granting you a bigger reach to attract new customers.
Adding users to lists on AdWords or other services can be done with Webtrekk Audience Stream. If you haven’t used it yet - give it a try, it is part of Webtrekk’s Customer Analytics Platform, alongside other marketing tools.
Use-Case: Triggering On-Site actions
Webtrekk’s in-session conversion prediction can also act as a trigger for onsite actions, e.g. triggering an incentive, a sales-agent interaction or special offers. Imagine a user with an average conversion probability and exit-intent behavior. You can try to induce a conversion with a voucher – but save money on the users who are likely to convert anyway!
Use-Case: Exit-intent replacement on mobile
On mobile, exit intent doesn’t work in the same way as on desktop devices since there is no mouse cursor that would be moved to close the page. Instead, you can look at the drop in conversion probability. The longer the session, the unlikelier the conversion without additional prodding. You can start to interact with the user either by prompting for the email address, a voucher, or creating a new account.
There are various use cases for predictions, so you will find further updates on the Webtrekk blog. If you are interested, contact firstname.lastname@example.org and ask for early access. We are always looking for data-driven customers who like to dive deeper into their data and predict the future with us.