Valuable marketing insights are hard to find in the mountains of data many marketers are now faced with – resurfacing them requires more sophistication than manual methods can deliver.
Machine learning and AI has come to the fore in this regard. So much so that 92% of companies have increased their investment in AI and big data this year. But things are rarely simple when it comes to data analytics. Even early adopters admit that they are yet to become fully ‘data-driven’:
- 53% of respondents state that they are not treating their data as a business asset.
- 52% cannot compete on data and analytics.
What’s the issue for such discrepancy between investment and tangible results?
In most cases, the lack of a proper AI adoption strategy is to blame.
Embracing AI is a bottom-up process that requires building a solid base first. That is determining which data sources you should choose for analysis and mapping a supporting infrastructure around those. So let’s take a closer look at the building blocks you should use.
1. Your end-goals will define the data you need
Machine learning and AI now has a variety of use cases in marketing. The CMO Survey 2019 lists the following as top uses of AI in marketing:
- Content personalization
- Predictive analytics for customer insights
- Targeting decisions
- Customer segmentation
- Programmatic advertising and media buying
- Improving marketing ROI by optimizing marketing content and timing
- Conversational AI for customer service
- Next best offer
- Augmented and virtual reality
- Autonomous objects / systems
- Facial recognition and visual search
Knowing what outcomes you want to gain will be key in deciding which data may be more useful to collect and operationalise.
To better articulate the use case, you should do two things:
- Audit the data you have and assess how easy it would be to consolidate it.
- Rank it against the common types of the machine learning algorithm, supporting the common use cases.
Broadly speaking, ML algorithms can now tackle the following tasks:
- Classification – answering yes / no questions or providing multi-class classifications e.g. segmenting different groups of keywords to pursue. In this case, you’ll need to provide the algorithm with a labelled dataset highlighting the right answers.
- Clustering – the algorithm devises the rules of classification and an appropriate number of classes itself. In this case, you do not provide the system with explicit rules for the division. Clustering is a solid method for segmenting your customers/audiences on a more subtle level to provide personalised offers.
- Regression – prompts the algorithm to come up with a certain numeric value. For instance, pricing decisions depend on many factors (internal/external) and regression algorithms can estimate the optimal value for different types of customers or enable dynamic pricing.
- Ranking – an algorithm ranks some objects by several features. Such algorithms are used by Netflix and Spotify to suggest similar content, as well as by Amazon to pitch up-sells / cross-sells.
Most likely your end-goal can be fulfilled by either of these approaches to data analysis. So your next step is to start gathering a relevant dataset.
2. Conduct an audit of internal sources first
Most companies are already sitting on a goldmine of valuable big data sources such as:
- transactional data/purchase histories (CRM & ERP systems)
- data about service/product use (CRM, customer support system, internal product analytics)
- online behaviour data (Google Analytics, Google Webmaster Tools + additional martech platforms).
Clearly, you have a lot of data stashed inside, but you do not need every bit of it (at least yet). When getting started with analytics, it’s best to avoid over-complex problems and pursue one manageable use case at a time.
For example, as an ecommerce company, you may be interested in modelling your customers’ lifetime value to anticipate their future behaviours and devise better marketing offerings.
The basic data prerequisites for predicting this parameter are purchase count, purchase value and purchase frequency rate.
This data is relatively easy to get from any modern CRM system and can be used for creating a basic predictive Pareto / NBD model for estimation (there’s a chart illustrating this model on this article fromOracle Data Science).
Once you have a solid base (a 3-parameter CLV model), you can add on additional attributes for analysis to increase the quality and accuracy of your predictions.
For instance, your next goal is to single out the high spenders in your online store. Age, location, loyalty status can be good indicators for that but collecting other values such as bounce rate can increase accuracy in predicting conversions for this group. To get access to this data, you will have to check if your CRM customer records are associated with web analytics and converge them into one data lake.
If you want to analyse more than one marketing parameter, you can then build a custom data-driven attribution model that will leverage:
- Google Analytics
- Google Tag Manager
- ecommerce tracking/goals data
Such a model will improve your visibility into conversions for different engagement, acquisition, and retention channels. Afterwards, you can enhance this model with CRM data to expand the analysis scope over the entire customer life cycle.
3. Consider external data sources too
Internal data sources can help you learn who your best customers are and how they interact with your brand.
But every company ultimately wants to find more ‘lookalike’ prospects in other markets and crack the code for attracting new business on command. That’s when you may want to add some external data sources to your analytics mix.
Let’s say you want to launch a ‘luxury product line’ and promote it to a new audience segment. How do you learn who’s most likely to buy from you?
- Leverage public data sources. By analysing Google Trends data you can gauge the demand/seasonality for certain goods among different demographics. This way, for instance, you can learn that an ‘organic beard oil for men’ has a good chance of flying off the shelves right now. As well, you can tap into UK Data Service datasets to better understand the needs of a certain demographic. Other good sources to obtain general market/consumer insights include Data.gov.uk and European Union Open Data Portal. GitHub also hosts a big list of cross-industry public datasets.
- Social data mining. Cluster analysis can help you determine the common points of discussion on social media in your niche, and match the popularity of some topics to a specific platform (Instagram, Facebook, Twitter etc). This way you can learn which content to serve and where. You can also use sentiment analysis to understand exactly what people like/dislike about your company or your competition. Every major social media network now has open APIs to use as an analytics source.
Moving forward with advanced analytics
The key to finding the most suitable data sources for marketing is learning to ask the right questions. By clearly articulating the problem you want to tackle, you will be able to determine what type of data you need and explore where it is stored. However, you will likely need to hire data scientists to help you automate the data collection process, set up appropriate infrastructure and assist with scaling. This way you can move away from data chaos and pave your way towards becoming a data-driven organisation, one model at a time.