This article originally appeared in Experfy on July 17, 2020.
In the race to become data driven, it’s clear that some industries have a tougher time than others. Namely, industries where human intuition is still of prime importance such as journalism or PR (public relations) are particularly difficult to quantify. To be certain, PR has metrics. However, it’s not quite clear what measures like outlet circulation/UMV or social shares for company content actually equate to in terms of business value. More importantly, they offer little in terms of predictive value. What quantifiable elements in a blog post, for instance, are predictive of how many times it’s shared on Facebook? Answers to such questions remain elusive.
Furthermore, many of the metrics in PR are difficult to apply to standard attribution models. For example, while it’s been shown that a positive article about a company or brand can boost favorability, it’s difficult to tell for sure if a well-placed story on your brand actually leads to a certain action being taken. There’s no direct trail between the two, as there often is in paid marketing tactics such as online advertising.
We also have a lot to learn about what shapes media trends. Why, for instance, does one story get covered by every major outlet in the world while another story–which may appear to the objective eye to be just as compelling–go totally ignored.
The current state of analytics in PR is almost entirely in the “descriptive” phase. In other words, it tells you what happened, but offers little insight into why it happened, or how likely it is to happen again.
While the “art” of PR–things like a well-crafted blog or a compelling pitch to a reporter–may not be made obsolete in the foreseeable future, there are a lot of untapped areas where machine learning can add value. While it won’t replace intuition, it can enhance and support it in ways that ensure creative efforts are directed toward the right target.
Predicting the success of content with Natural Language Processing (NLP)
In all content marketing efforts–whether it be a blog post, article or video–the dream is for it to ‘go viral’. Unfortunately, no one really has any idea why one piece of content goes viral while another falls flat. However, predicting virality in content isn’t totally out of the question. If a Natural Language Processing (NLP) model, such as Naive Bayes and Random Forests, can be trained on a sufficient volume of successful content, it may be possible to predict the spread of new content as it is generated.
We’re seeing NLP applied to PR data right now with platforms like Meltwater and Cision, which analyze the sentiment of articles toward a brand. It’s my observation that, at the moment, these predictions aren’t trustworthy. It may be, however, that they lack predictive power because they’re being trained on aggregate data based on of many industries. As the language used across industries varies tremendously, it’s unlikely that such models would discern the nuance required to predict sentiment on a brand-by-brand basis. If these models were trained on news articles from specific industries, I suspect they would more accurately predict sentiment. From there, it’s only a small leap toward developing NLP models that predict more complex things, such as how many times a blog will be shared on social media, based on analysis of its text.
Correlations between standard metrics
Machine learning techniques can be used to discover correlations between standard metrics. In assessing the effectiveness of a campaign, a PR team might look at metrics like the unique monthly visitors (UMV) of a publication where a brand was featured, or the number of social shares of the article. But is there a relationship between the two? Does an article published in an outlet with a higher circulation produce more social shares? Regression techniques as well as Neural Networks can be used to discover such relationships and, possibly even more importantly, determine if tactics can be implemented that boost one or the other. For instance, my team analyzed about 400 articles and discovered that–counterintuitively–the size of a publication has only a very slight impact on the number of social shares that an article receives. On the other hand, in the same study we found that articles with higher numbers of social shares tend to produce higher MOZ scores for brands, indicating that investing resources in sharing articles through social channels may be an effective tactic for companies (who may see a 3 to 7 point boost in the article’s MOZ score for 100 shares).
Finding the right reporters
The ‘art’ of PR involves convincing journalists that a particular story angle is newsworthy (typically something that favorably leverages your brand). This often requires manually sifting through scores of journalists’ news coverage to determine that ‘perfect fit’. Machine Learning can speed up this process. For instance, Market Basket Analysis, which is used by e-commerce recommendation engines to predict purchase preferences based on previous purchase patterns, can be applied to reporter coverage preferences. For example, my firm ran a study that looked at coverage patterns from thousands of reporters covering infant health issues, and found a reporter who had covered antibiotics and constipation was 3.7 times as likely to cover probiotics. If you were interested in landing a story about your probiotics brand, this process might help you more quickly identify the journalists most likely to bite on your news angle.
Which outlets are most important to your brand? Unsupervised Learning can Help
Getting your clients good press coverage means identifying the outlets that are most relevant to their brand. It’s tricky, however, because there are a number of metrics that can be used to determine relevance, including Unique Monthly Visitors (UMV), backlinks, authority score and relevant search terms for the site. It’s easy enough to compare outlets one metric at a time, but in order to get a full picture of value, you have to look at these metrics as a whole.
Unsupervised learning techniques such as K-means or Hierarchical Clustering, can group outlets by multiple attributes, providing an aggregated picture of the kind of value that a publication might bring for your brand, and allowing you to focus your attention on the most important outlets. For example, we used Hierarchical Clustering to examine 15 different tech outlets for one client, and determined, based on the factors mentioned above, the first, second and third tiered outlets in terms of value, enabling us to prioritize our efforts.
Reading “between the lines” with Association Rules
Many have wondered whether hidden patterns exist in news coverage. Machine learning reveals that they not only exist, but also offer predictive insights. My firm, for instance, analyzed 6000 election-related articles using a model based on Association Rules, and uncovered some interesting patterns. For example, an article with the terms “Mike Pence” and “Kamala Harris” was 7 times more likely to contain the term ‘racist’, and an article containing “Biden”, “Castro” and “Klobuchar” was 6.8 times more likely to contain the term “socialist”.
In my opinion, this is only scratching the surface of the potential for Machine Learning in PR. In an industry that has its finger on the pulse of the news cycle and social media, the potential for mining data to uncover relationships and patterns is massive, and the future is exciting.