A Practical Guide to Artificial Intelligence for Publishers
Raab Associates recently took an informal census of vendors using artificial intelligence for marketing or customer management. The list, which is nowhere near complete, included more than 120 vendors offering 58 distinct capabilities within nineteen categories. Exactly one of the nineteen categories (biometrics) had no obvious application to publishing (and even biometrics has some non-obvious applications).
One lesson from this exercise is that “artificial intelligence” is too vast to treat as a single topic. Rather, publishers need to look at how individual applications can support specific business needs. These applications may draw on some shared resources, such as consolidated customer data, content libraries, and publishing platforms. Their success also depends on organizational capabilities including change management, performance measurement, and data-driven decision-making. But it’s still important to understand how these applications differ in order to make wise choices about what to do and how to do it. Following is a quick guide to artificial intelligence and how publishers can apply it to their businesses.
What Is Artificial Intelligence?
Let’s start with a bit of background. Artificial intelligence has many definitions but the one I prefer is “machines that learn over time,” with the understanding the “learn” really means “learn to make decisions that achieve specified goals.” It’s the ability to learn that sets AI apart from conventional computers, which follow fixed rules provided by human programmers. Details vary, but most AI systems today are “trained” by giving them examples of inputs and outputs (i.e., correct decisions), and letting the system generate its own internal rules to predict the output from the input.
After initial training, the system is fine-tuned by feeding it new inputs and letting it generate its own outputs. The user then tells the system whether its answer was correct and the system adjusts its rules based on the new information. This process continues until the user decides the results are acceptably accurate or that the system won’t work for their particular purpose. The system’s internal rules are often incomprehensible to humans, so checking output is the only way to ensure it is functioning as desired. Because the rules will change over time in response to new information, it’s essential for users to keep a close eye on outputs after deployment.
Many AI categories apply to the business side of publishing, such as ad sales, circulation, and customer support. For this article, though, let’s look at AI applications related to the reader experience.
Publishers can use AI to understand topics that readers are consuming on their own sites and elsewhere. It involves selecting the content to analyze, scanning it, extracting key features, and then analyzing the results. Typical content analysis questions include what’s most commonly presented, what’s most often read, what’s new, and what’s changing.
The simplest form of extraction is to scan for specific words or phrases. Such key word analysis doesn’t require much in the way of artificial intelligence, but it’s also very limited. One problem is that it only captures terms that users know about in advance, so it can easily miss new topics or trends entirely. It also can’t automatically associate similar terms, find relationships between terms, understand context, or measure sentiment.
The form of artificial intelligence known as natural language processing does all those things, providing much richer data to analyze. From an editor’s perspective, content analysis provides guidance about what to create next. (Outside of editorial, content analysis keeps an eye on competitor’s ad campaigns and marketing materials, providing useful market intelligence.)
While most content analysis is still based on text, AI systems are becoming better at interpreting images, video, audio, and other formats. Like text analysis, these help to understand what readers are being offered and responding to. Also like text analysis, they are far from perfect. Sentiment and emotion analysis in particular are still very limited; any system that offers them should be evaluated closely before you trust their results.
Vendors in this space include Sysomos and LiftMetrix for social media analysis, Amobee and Stackla for web content consumption, and Captora, Unmetric and Crayon for competitive marketing information. Image and video processing vendors include Alphonso, LinkInfluence, Clarifai, Ditto Labs, and Cluep.
We’ve all heard about AI systems generating financial and sports stories that most people can’t distinguish from the output of professional journalists. But using AI for content creation also extends to email subject lines, social media posts, web pages, and videos. These are mostly used to create advertisements, but they can also help publishers boost consumption of editorial products such as email newsletters.
Content creation systems and content analysis systems can be thought of as mirror images. Both rely on large databases of words (or sounds or images) that have already been classified, plus semantic engines that understand the rules of a given language. But while content analysis systems use this information to pull apart existing sentences, content creation systems use similar information to assemble new ones. In both cases, the quality of the databases and subtlety of the semantic engines are critical to success. Users must often create new vocabularies and semantic rules that are tailored to their particular application.
Vendors in the content creation space include Narrative Science, Automated Insights, and Arria NLG for article generation, Persado and Phrasee for ad copy, Adext for web pages, and Magisto and Wibbitz for video.
This is probably the best known AI application for publishers. Reader-related applications include recommendations for content to read, watch, or listen to; products to buy; and actions to take in a game or mobile app. Business-related applications include recommended marketing, sales, and service messages.
This category overlaps with predictive models, which often provide inputs to a content recommendation and themselves may be built by AI systems. But content recommendations go beyond predicted response rates to factor in the long-term impact of a recommendation on lifetime value, retention, service requests, and other future behaviors. Predicting and optimizing for these items is where the artificial intelligence comes in.
AI is also often required to identify the variables affecting likely response, such as relevant product features (topic, brand, style, quality, function, etc.), context (location, time of day, day of week, etc.), customer relationship (new, satisfied, recent problem, responsive, etc.), and customer attributes (demographics, lifestyle, persona, etc.).
There are also many different types of recommendations and ways to make them, including recommendations for similar products, complementary products, most popular products, or best values; recommendations based on the individual, segments, or the entire customer base; recommendations choosing from a few options or a huge catalog; and recommendations in response to a search request. Most publishers will be more interested in recommending content than products, but many will need both.
Some recommendation systems can handle multiple types of recommendations. But most publishers will use different systems for different purposes. This can deliver better results but adds complexity and, often, a need to coordinate outputs from different systems. Because recommendations are so varied, it’s especially important to check how well any system matches your needs before making an investment.
Web Conversion Optimization
This goes beyond content recommendations to optimize every aspect of the website, including page layout, copy, images, and offers, with the goal of funneling more users toward a subscription or purchase. AI-based optimization sets up tests of these elements, monitors results, and deploys the winner. In some systems, users define the versions to be tested and the system simply executes the tests. Other systems create the test versions themselves by combining prebuilt components.
AI systems can manage many more tests than humans but must learn to create combinations that make sense in subtle ways and to build on previous results. Some AI systems continuously create new versions to test against the previous winners. This lets them improve performance and adjust automatically as new components become available and as customer behaviors change. In some cases, the AI recommends new components for users to create, such as additional products, topics, or offer types.
All optimization programs need a goal to measure against. Many systems target a specific behavior such as user registration or placing an order. Others target longer term outcomes such as attracting high value customers or increasing lifetime value. Since the long term outcomes cannot be measured immediately, the AI may create a predictive model that estimates them.
Traditional web optimization systems find the single best configuration to serve all customers. This distinguishes them from personalization tools, which create different versions for different segments or individuals. AI-based systems often combine these capabilities, simultaneously finding customer segments and creating new site versions tailored to these segments. Site optimization is also often combined with content recommendations: for example, the optimal site design for current customers may include a recommendation box that is populated with the best content for each individual.
We’ve just scratched the surface of what’s possible with today’s artificial intelligence systems and there’s no doubt that tomorrow’s systems will do even more. But the complexity of the topic is no reason to avoid exploring AI. In fact, the large number of applications means publishers can start with whichever they find most appealing. Nor should they be deterred by the technology itself: vendors have hidden nearly all the details from users, who only need to provide training data and review the results. In many ways, it’s best to think of AI as a smart new employee who will do great work once you teach her your business and adjust your processes to take advantage of her skills. In short, AI is a management challenge more than a technology, and publishers who learn to manage it successfully will reap great rewards.
David Raab is a consultant specializing in marketing technology and analytics. His clients include major brands in publishing, retail, financial services, telecommunications, technology, and other industries. His early career was spent in magazine circulation and direct mail continuity marketing. He is founder and CEO of the Customer Data Platform Institute.