Unsupervised Learning and SEO/SEM

A fair amount of the time when we’re working with AI and machine learning, we’re working with supervised learning. As defined in AI for Marketers, supervised learning is when we’re having our machines look for something we know we want:

• Find a logo in this pile of Instagram photos.
• Find which variable correlates most to an outcome we care about.
• Find the prospects most likely to convert.

However, we don’t talk nearly as much about unsupervised learning, when we don’t know what we’re looking for and we need our machines’ help to find it.

For example, we know from previous studies that Google uses advanced machine learning models of its own, including BERT, to classify text and match it to searcher’s intents. Google’s models look not only at individual keywords, but sentences, paragraphs, and entire documents to judge the relevance of a document in its entirety to what someone searches for.

If we want to do well in SEO, if we want our ads to do well in SEM, we need to understand what it is Google is seeing, and try to extract that information and use it. This is reverse engineering; it’s like trying to take a fully-cooked dish and understand how it was made. Suppose we’re eating a famous chef’s French toast. Can we tell, just by taste alone, what went into it?

To extend the cooking analogy more, there are some things in reverse engineering Google results that are obvious, that you don’t need machine learning for, just as reverse engineering a recipe has obvious points. French toast contains bread; there’s no need to study that extensively when it’s patently obvious there’s a slice of bread on your plate. If we’re marketing a course about data science, data science as one of the core terms is plainly obvious.

Where we need help is in the nuances, the non-obvious stuff. So the question is, how do we tackle reverse-engineering Google’s search results for those nuances? Here’s the good news: Google has technically done the hard work already. When you search for something like, say, data science courses, Google’s language models have already identified the most relevant documents and content. The only thing we need to do is determine what in those documents, from a language perspective, made them relevant.

This is why we call this unsupervised learning. We don’t know what we’re looking for, but our machines will help us make that determination.

Example

Let’s look at an example for data science courses. In the SEO software of your choice, type in your core keyword and see what articles and content show up:

This is a good selection of relevant pages that attract search traffic for our chosen term, data science courses. A quick sniff test suggests these results are relevant for the most part. Our next step is to figure out what all these pages have in common, what’s interesting about them from a topical perspective.

To do this, we vacuum up the page text from each of the pages and use natural language processing to extract out the most relevant terms:

This is where domain expertise matters, where human oversight is important. Some terms are superbly relevant. Some are not. The key is cleaning up our list so that we’re able to generate content from it. Once we’ve done our list cleaning, we’ve got a seed list that we could use for many different kinds of marketing:

• SEO: obvious application. If we’re making content about data science courses, our target pages should check the box on many of the terms that are receiving organic traffic.
• CPC: the same keywords we target for SEO, we should be using in SEM.
• Social media: content we create for social media should echo our keyword list; after all, if the language being used in the most visited content is in our social posts, it reflects our understanding of the discipline and implies our expertise.
• Email marketing: when you optimize for SEO, you are also determining what may work for things like subject lines in email.

The hardest part of this process is doing something with the data you’ve acquired. This level of analysis means nothing if you don’t put it into action.

For an even more detailed analysis, watch this episode of So What?

Can’t see anything? Watch it on YouTube here.

Conclusion

Reverse engineering at least part of the way is within our reach as marketing technologists. We don’t need to have Google’s massive technology infrastructure to discern the content-based features they might be seeing, just as we don’t need access to Gordon Ramsay’s entire kitchen to get a sense of what might be in his chicken marsala. We’ll miss some of the finest details and we won’t see the preparatory steps, but we can still assess the outcome and reverse engineer the most obvious, critical parts.

You might also enjoy:

• You Ask, I Answer: AI Works And Copyright?
• You Ask, I Answer: AI Music Collaborations and Copyright?
• Fireside Chat: Geraldine Deruiter on Food, Feminism, and Fury
• You Ask, I Answer: Retrieval Augmented Generation for Tax Law?
• Mind Readings: Hacking Social Media Algorithms

Take my Generative AI for Marketers course!

Join my Analytics for Marketers Slack Group!