Why AI personalization needs to open the black box to increase adoption
An interview with Jean-Noël Rivasseau, CTO, Kameleoon
AI – the key to scalable, real-time personalization?
98% of website visitors leave without having either bought anything or moved forward in their customer journey. Contrast this with the experience in physical stores, where visitors are much more likely to engage and make a purchase, thanks to trained, skilled staff who understand their behavior, personalize their approach and deliver what they are looking for.
That’s why brands are continually looking to extend this same level of personalization to the online world. However, applying this approach digitally, at scale, in real-time, has previously been difficult, if not impossible. To achieve this brands have to be able to personalize for huge numbers of visitors at the same time, often with little information about their wants and needs.
AI offers a solution to this problem, matching intelligence with real-time speed in order to deliver personalized experiences that will increase engagement and sales. However, the black box approach of most AI personalization solution vendors means that brands have no idea what is happening and why particular decisions are being made. They therefore worry about a lack of control and transparency. The big issue is that marketers are unable to understand exactly what has led to results so cannot analyze if their overall strategy is being effective – they are therefore wary of letting AI define their targets and tactics unless they know exactly how it works.
How can brands overcome this challenge? We talked to our CTO Jean-Noël Rivasseau about how you can use AI personalization to drive results while ensuring control and transparency.
What are the challenges facing marketers embracing AI today?
Brands understand that AI has the potential to deliver major benefits, particularly around personalization, but do have concerns in three areas:
- Choice. There are already a huge number of marketing technology platforms available, and the number is growing all the time. Picking the right one for your specific needs is difficult, particularly as from the outside many claim to do a similar job.
- Hype. AI has become a real buzzword. It can feel that every vendor is now offering an AI-powered solution, and that they promise to deliver amazing solutions to every problem. This makes it hard for brands to identify the best use cases where AI will make a difference and drive results for their individual needs.
- Results and Ethics. Brands need to know that the AI they choose will deliver results – and deliver them ethically. Will it achieve ROI and can you understand the approach and models used? Can you therefore be confident that it will meet your needs? Brands obviously want to achieve their marketing objectives, but not at the price of damaging their reputation through unethical or discriminatory algorithms. They want to retain control of their strategy and tactics.
How does AI personalization work in practice?
AI personalization relies on Machine Learning artificial intelligence. This essentially works by analyzing data and learning rules based on the patterns it finds in this data. It then applies these rules as algorithms on new data, improving its predictions as it goes. The key thing is that it learns itself, which makes it extremely powerful – you feed it data in order for the system to train itself.
You can see this machine learning in action when it comes to technology such as image and speech recognition for example. I’ve been working in Machine Learning since 2002 and have been involved as the technology has grown and moved beyond image and speech recognition into a huge number of applications, particularly in the past few years.
With AI personalization, systems essentially analyze data around customer behavior to predict what consumers will want. This can be “hot” data that is being collected about what they are currently doing on the website, or existing “cold” data which covers what you know about them. This could be preferences and previous orders from systems such as CRM, if the visitor is known to the brand and is logged into the site. The aim is clearly to increase conversions by offering them the information and experience that will best meet their immediate needs.
Kameleoon uses neural networks to analyze both types of data, in real-time, automatically finding correlations between visitors in order to define a target group of people with often very complex characteristics. Showing the power of AI, these characteristics are often too complex for a human mind to easily grasp! So, for example, it could be used to analyze which visitors require an incentive to convert, and which will convert anyway without receiving a special offer. We will explain how we use information, and how it sits with data privacy, in more detail in a future blog.
What makes Kameleoon’s approach different?
I’ve always been interested in Machine Learning and started to develop Kameleoon’s AI personalization platform with the team in 2012. AI is not easy, and this gives us an enormous depth of experience in the field and helps us to both understand current needs and predict future trends.
Essentially, there are three important ways that we differ from other vendors:
- Our platform is usage centric – it was designed to answer any business need, focusing on onsite web personalization in real-time. I truly believe that the value of any technology only resides in the use we can make of it, which is why we designed our platform to adapt to ANY use case. We’ve developed this further with our 450+ clients, and typical use cases within this include:
- Personalizing brand universe and content depending on visitor’s sensibilities and needs at given time. So this could include creating an individualized homepage, with header, images and content personalized for every user, dependent on what appeals to them.
- Triggering marketing and commercial actions dependent on purchase intention. For example, offering a discount to sway undecided purchasers.
- Generating more qualified leads by identifying hot prospects within a targeted segment and providing them with a tailored experience that pushes them to engage and move forward on the customer journey.
- Activating data by engaging with cold prospects that have been defined via a Data Management Platform (DMP) or a Customer Data Platform (CDP).
- It focuses on being customer-centric. Traditional retail product recommendation platforms work through a graph method. A visitor views products A, B and then C, and the platform aims to predict which product D to then show to have the greatest chance of generating an upsell. By contrast the input data we use is almost all customer-centric, based on what they have done, or what you know about them.
- Our AI brings real-time into onsite personalization strategies. This means it is very specific to the actual visit to a website and the real-time actions that can be triggered when they are browsing on that site.
The Kameleoon platform is ideally suited to when you have a specific goal or offer, but are unsure who the target audience should be. It uses Machine Learning to define the characteristics of this target audience, and then finds them from amongst your visitors, in real-time, in order to provide the personalized experience that is likely to lead to conversion.
Opening the black box – Kameleoon’s algorithms
One of the main concerns from marketers is that AI is not transparent, and they cannot understand how algorithms arrive at particular results.
Like most AI vendors, Kameleoon uses proprietary algorithms that are created by learning from the raw data. However, the key value we provide is in the architecture around them. These algorithms are supervised – the intelligence comes from the training. How does this work?
1. Learning phase
When settings are changed, a new supervised learning phase is triggered. The data model is trained using the data of all visitors to the site. This trains the neural network by inputting visitor session data, correlating this with goal conversion and thus identifying signals that positively or negatively influence conversion. When its predictive performance is considered satisfactory (normally after 2,500 conversions/around 100,000 visits), the data model can be deployed. However, the learning phase itself never ends as all new visitor data keeps teaching the algorithm, further improving its prediction capabilities and adapting to changes in context.Given that the percentage of visitors that convert on any website is very small (around 2%), the natural tendency of algorithms is to predict a no-conversion. Kameleoon’s heuristics therefore compensate for this – we encourage the algorithms to take risks by magnifying positive signals, based on experience and previous use cases.
2. Deployment Phase
When deployed the data model monitors the visitor’s behavior in real-time and then calculates their current level of intent to convert – the raw score. This internal metric is then normalized into the Kameleoon Conversion Score (KCS), which is an external metric that can be easily understood by marketers.This is a major differentiating point for Kameleoon. We don’t have a black box approach – our algorithms provide a score that makes it straightforward for marketers to see why particular actions have been taken for particular visitor segments, increasing confidence in the system through greater transparency.To find out more about Kameleoon’s technology and how it delivers transparency for digital marketers download our technical whitepaper: