Joining the Dots: The Big Data Problem

Joining the Dots: The Big Data Problem

November 2019 | Marketing Operations

Bad data quality is a constant challenge for all marketers. Alas, there is no easy fix for bad data but AI does help.

Data is at the heart of all digital marketing activity. In an era of data-driven decision making, the information used to make investment and operational decisions must be as accurate and as comprehensive as possible. Yet, far too many B2B marketers pay little to no attention to the quality and completeness of the data they have available. More often than not, bad data inevitably leads to bad decisions.

Ultimately, few marketers truly understand what good data looks like and the difference it can make. It is often only talked about in terms of segmentation and targeting possibilities, yet a comprehensive marketing database also acts as the foundation for detailed reporting. That's because there is more to a contact profile than an email address and job title. Activity data and Predictive scores are all part of that profile too.

Knowing what a contact is interested in, the kinds of messages that resonate and the channels they're engaged with is vital in an omni-channel world. Marketers spend a lot of time trying to guess all these things based on gut instinct when a comprehensive data analysis can give you a scientific answer to any questions you may have on these topics. The trouble becomes actually making sense of all that data available. The typical enterprise actually knows a lot about their customers, despite the fragmentation and inconsistencies that exist within the business. For one thing, much of it will be held outside the marketing department in ERP systems or customer services. There are limitations surrounding how that data can be used for marketing in a GDPR world, but it can definitely be utilised in some capacity.

Categorising Complexity

Even if purchase history and support patterns can't be analysed at the contact or account level, they should definitely be aggregated. A lot can be learned by examining purchase history over the span of multiple years to discover buying patterns by industries or buying groups. This is an area where leveraging AI models and machine learning to tease out the trends and then map them to similar accounts. This is basic pattern matching at enterprise scale, an area where AI analysis has long been superior to human interpretation. A machine can analyse data to a much more granular level than a human and will be quicker to spot clusters of purchasing patterns in specific industries, company sizes and geographies. Predictive lead scoring has been doing it for the past decade.

The complex part of this kind of analysis is aggregating the clusters of potential customers into a set of account types or buying unit categories which make sense to business users. That's an area where AI is less useful. Lead and account scoring is often seen as the best way of doing this because it best fits the way that predictive scoring works. However, this limits the usefulness of the output to marketers, who can struggle to interpret a score and what the score means in practice. That is because scores are now being used to measure more than just the difference between a hot lead vs a cold lead. They are being used to measure the level of potential fit for specific personas or product categories.

Filling the Gaps

In a marketing context, predictive data analysis is often only associated with scoring, but it doesn't have to be. CRMT's Normalator data normalisation service uses predictive AI to convert job titles into normalised job level and job function fields. This kind of data normalisation is a rule based process. A data analyst identifies specific keywords or phrases in job titles that can then be mapped to levels and functions. This is fine for the most common job titles in the most widely spoken languages, but can't account for the long tail of one-off job titles in niche industries that typically aren't mapped to a Job Role or Job Function, even when they should be. This is an example of the 80/20 rule in action.

The typical data normalisation process will match 80% of your database, leaving the other 20% as blanks which can't be used for campaign segmentation or data analysis. Normalator uses AI based pattern matching to fill in these blanks, inventing the missing normalisation rules using similar rules for related words or phrases. This does leave the potential for error, so a confidence score is assigned to each match that measures how accurate the assigned value is likely to be. We typically recommend reviewing matches with low confidence scores and manually updating the normalisation rules to fix any incorrect matches. This teaches the normalisation algorithm the correct normalisation for a particular phrase, improving match rates and data accuracy over time.

I've seen similar techniques used for de-duplication and other data cleansing tasks, taking days or hours off what are otherwise very laborious and error-prone manual maintenance processes. The sheer amount of data being stored in the marketing database means that cleansing and categorisation is essential, and such tasks can not be carried out manually. AI is an immense help in performing these tasks, but it cannot be used to interpret the output within a business context. That does not make AI totally useless in contextualising some kinds of data.

The Power of AI

Technology firms are now adding AI based recommendation engines to their products. Pardot can now tell you the best emails and subject lines to send to a particular audience. Eloqua and Marketo will be getting similar features soon. It sounds revolutionary, but in fact such recommendation engines aren't that new. Pardot's campaign recommendations are simply a more advanced form of pattern matching driven by advances in text and image recognition. Salesforce's Einstein AI is simply reviewing campaign results and identifying the best performing emails for a particular list. Emails that have performed well with similar lists are then recommended to marketers as a potential future campaign for that audience.

Such recommendations seem cool but are actually of limited utility to a B2B audience. Marketers have spent the last decade trying to move away from random acts of email marketing to an orchestrated campaign strategy. Sending emails to one audience just because they worked for a different audience goes completely against this. The message has to fit in the overall communication strategy and the marketing calendar because the buying cycle takes longer than a single campaign. Opens and clicks are not the correct metrics to identify a successful B2B email blast. Much more useful would be identifying the specific messaging and campaign features that have worked to a specific audience over multiple campaigns. However, that is a harder task that AI isn't yet capable of assisting. It's only a matter of time.

Written by
Marketing Operations Consultant and Solutions Architect at CRMT Digital specialising in marketing technology architecture. Advisor on marketing effectiveness and martech optimisation.