Marketers see great potential value in using artificial intelligence (AI) to support the use case of recommending highly targeted content to users in real time. That use case scored the highest among 49 use cases presented to marketers in the 2021 State of Marketing AI report by Drift and the Marketing Artificial Intelligence Institute.
That use case scored a 3.96, putting it on the cusp of “high value” (4.0), with 5.0 being “transformative.” The AI marketing use cases that trailed in the top five include:
- Adapt audience targeting based on behavior and lookalike analysis (3.92)
- Measure return on investment by channel, campaign and overall (3.91)
- Discover insights into top-performing content and campaigns (3.86)
- Create data-driven content (3.82)
“Most websites you go to today for businesses, a human is writing the rules to say which content to recommend,” Paul Roetzer, CEO and founder of the Marketing Artificial Intelligence Institute, told CMSWire in a CX Decoded Podcast. “What are the related articles? There is some basic tagging system for if they read this, then read that. Most of them are human-powered. They don’t have a Netflix or a Spotify type algorithm that’s actually learning preferences, knows the last 15 articles someone read, and how far along he got into them. It’s not pulling any other kind of behavioral intent data. Most aren’t doing that.”
The Data Conundrum
Therein lies potential, however it’s something marketers and customer experience professionals remain hopeful about: 54% of them told CMSWire researchers in the State of Digital Customer Experience 2021 report they see AI having significant impacts on digital customer experience over the next two to five years. And most of them see “gaining actionable customer insights” (27%) as the area where they see the most potential.
Roetzer said it is hard to find really good solutions to do this out-of-the-box. Noz Urbina of Urbina Consulting agreed, calling the technology nascent.
The bigger question for marketers beyond what kind of tools are out there is do we have the data to support the use case, according to Roetzer. And do we have a strong foundation of metadata, content tagging and content taxonomies, according to Urbina.
“You need enough data, for one,” Roetzer said. “Sometimes the problem is smaller data, not necessarily the cost. It’s do you have enough data to make it worthwhile to try and use a machine learning algorithm to do this better than a human would? Do you have enough traffic coming to your site to justify it?”
Related Article: CX Decoded Podcast: Practical Use Cases of AI in Marketing
Build or Buy?
Does it make more sense to custom-build a solution on AWS or Google, or is there an out-of-the-box solution to go plug in for a couple grand a month that will learn our users and start making recommendations? These are some questions marketers should be asking when considering using AI for targeted content recommendations, according to Roetzer.
“A lot of people are actually building on GPT-3, a technology that came out of OpenAI, which was sort of a lab that was developed to rapidly advance AI technology and then share it with the world thus Open AI in the name,” Roetzer said.
According to OpenAI, nine months since the launch of the first commercial product, the OpenAI API features more than 300 applications. Those 300 or so companies are building language generation abilities on the backbone of GPT-3, according to Roetzer. He cited conversion.ai and copy.ai, the latter which secured $2.9 million in funding in March. “What they (copy.ai) do is they have a bunch of pre-trained models so you just get a subscription, and you can actually go in, feed it some inputs … and it’ll actually write ad copy for you, email copy. Very interesting.”
OpenAI officials cited the example of Algolia, which partnered with OpenAI to integrate GPT-3 with its advanced search technology in order to create their new “Answers” product that better understands customers’ questions and connects them to the specific part of the content that answers their questions, according to OpenAI officials.
“Algolia Answers helps publishers and customer support help desks query in natural language and surface nontrivial answers,” they wrote. “After running tests of GPT-3 on 2.1 million news articles, Algolia saw 91% precision or better and Algolia was able to accurately answer complex natural language questions.”
Related Article: 8 Considerations When Selecting an AI Marketing Vendor
Responding to Behaviors
Urbina said the most popular method for generating targeted content in real-time through AI is through recommendation engines. According to Google developers, content-based filtering uses item features to recommend other items similar to what the user likes, based on their previous actions or explicit feedback.
“Rather than only responding to a query, which is a definition of a search engine, recommendations engines are responding to behaviors,” Urbina said. “Your location, your activities, your previous search behaviors, all of these things are ambient data that search engine technology can be used to then be turned into a recommendation engine. And artificial intelligence and machine learning, of course, are foundational for that. They will find patterns in the data and then they will make appropriate recommendation.”
The most common phrase after recommendation engine that marketers need to know about is “next-best action,” according to Urbina. Marketers want to usher people along the journey, and machine learning helps to determine various next actions such as sending an email, an SMS or offering recommended content that pops up.
“And it is optimizing this at scale beyond human capacity,” Urbina said. “So the AI has to observe the user activity, and, based on the data of that user correlated against what all the other users say, determine the next best thing that I could suggest to move them along the journey. So that’s basically the main area that we need to focus on for content recommendations: finding out how we can correlate the user’s behaviors, against trends to establish the next-best action, which can be recommended content.”
AI Can’t Do It Alone
What marketers often struggle with is leaving the AI to do all the heavy lifting, according to Urbina. There exists the need to have a solid structured content plan in place: labeling and applying metadata on existing content and then building content taxonomies.
“One of the most successful things we’re doing now is building out the taxonomy so that the recommendations engine has something to work with,” Urbina said. “A taxonomy of personas. A taxonomy of business scenarios. A taxonomy of challenges. A taxonomy of benefits. A taxonomy of functions. A taxonomy of content types. A taxonomy of channels. If you haven’t actually got this taxonomy in place that defines these buckets, what can the AI work with?”
With a foundational content structuring program, marketers can define what makes a white paper, what makes a case study, what makes a brochure, what makes a product overview and so on, according to Urbina.
“Then you unleash the AI over all of the content that exists … and it reads it, uses natural language processing to see what the subjects are and the common words are in your content and then you can organize into your actual taxonomy. So before you’ve got any of this going, you could actually use AI to figure out what your taxonomy is and what your categorizations should be.”
Why a ‘Brute Force’ Approach to AI Won’t Work
Most marketers just want a data scientist to do all of this, but that’s not always possible. Further, they often don’t actually realize that if they participated in the organizing of the content around taxonomies and metadata and tagging, the whole operation will be much more effective.
“And that’s where I see this technology is absolutely nascent. And its effectiveness through brute force approaches is the thing that’s mostly slowing it down,” Urbina said.
Can this AI tech supporting content-recommendation AI engines pick up brand tone? Roetzer said it’s getting there.
“It’s made massive leaps forward in the last three years,” Roetzer said. “2013 was sort of this tipping point where the AI caught up with the promise of what it could eventually do, and language is at the core of that. It’s why voice assistants have actually gotten good. It’s why some conversational agents have gotten good and why language understanding and generation have gone to entirely new levels in the last few years. And so the ability to understand and replicate tone, if it’s not there, it’s coming. And there are a lot of people putting a lot of money behind that sort of thing.”