“I know I’m wasting half my marketing budget, I just don’t know which half”. It’s a well-worn marketing phrase and reflects a problem many would like to solve.

The better you understand your customers, the more effective your marketing – so VisualDNA has been looking at how to apply the profiling technology it uses in advertising to CRM, to see if it could predict which ‘unknown’ customers are most likely to buy.

VisualDNA has invested heavily in personality profiling technology. Using patented visual-based personality quizzes, it is able to profile individuals online with great detail and accuracy, and subsequently use behavioural models to extrapolate this detailed knowledge onto hundreds of millions of users. It groups people by various characteristics into ‘segments’ – anonymised groups that can be used by businesses in communications, usually targeted advertising, product recommendation or content personalisation.

There are other applications too, one such area being CRM. Most businesses, whether they are business-to-business, or business-to-consumer, possess a database of customers, or active leads. These customers can be categorised in all sorts of ways, but whichever way you define your customer base, it’s notoriously hard to predict who is most likely to buy.


VisualDNA wanted to test the idea that the inference algorithm it uses to build its high-performing segments could also be used to identify which customers are most likely to buy out of any given data set.

Piano Media is a company that runs a subscription paywall business on behalf of a large number of publishers in multiple countries in Europe; with a massive data set of customers who fit one of three categories:

  1. Unknown / un-registered,
  2. Registered but not subscribed,
  3. Paying subscribers.

Clearly the objective for any business operating a freemium model is, having attracted the customer in the first place, migrating your unknown and registered users into paying customers. The question is, how do you know which of your registered users are most likely to become paying subscribers? At which part of your audience do you direct your marketing?


Working with Piano, VisualDNA drew a random sample of 30,000 registered non-subscribers and compared their online behaviour to that of the entire population of paying subscribers; using what it calls Look-alike modelling.

The audience insights company did this completely anonymously, linking with Piano’s database using randomly generated anonymous identifiers stored in cookies.

From here VisualDNA was able to assign a similarity score to every one of the 30,000 registered users – the score representing how similar users’ behaviour is to the paying subscribers and therefore how likely it thinks this person is to convert to a paying subscriber.

Piano then did three mailshots, under exactly the same conditions, to three separate groups within that population;

  • The top 5,000 (those with highest score)
  • A random sample of 5,000 users
  • The bottom 5,000 (those with lowest score)


The results are hugely encouraging. On the top 5,000 Piano recorded a conversion rate of 1.59%, compared to 0.52% for the random group and 0.22% for the bottom group. In other words, our algorithm correctly predicted those most likely to convert, based on the theory that the online behaviour of a paid-up subscriber was sufficient to build a profile on them and use that as a means to find those with a similar profile. Moreover, it correctly predicted that users with behaviour dissimilar to paying customers are a lot less likely to convert than random users, so that marketers don’t need to spend the money and energy marketing to these.

This result has significant implications. It means VisualDNA can, in theory, take any database, whether online or offline, anonymously profile anyone in that database, and use its inference and Look-alike algorithms to predict behaviour to improve the effectiveness of a company’s marketing – in this case increasing conversion more than three-fold vs a random sample and weeding out those least-likely to buy.