For the larger part of my SEO career, I was leading a team of a dozen marketing specialists working on multiple SaaS or affiliate projects. At one point, I asked myself whether I could utilize my data science expertise to get better marketing results.
Obviously, the answer was ‘yes’, but what surprised me was the fact that I got some outcomes far better than what I had expected.
While I’m sure many SEO specialists or even advanced tools are doing what I did — in one way or another — I also feel the techniques I’m about to describe aren’t as popular as they could be. Here is how I used machine learning to effortlessly drive organic traffic to my client’s websites.
Keyword Research & Multi-armed Bandits
Keyword research is a quintessential part of almost any SEO task. Whether you’re planning new content, optimizing old posts, or trying to outsmart your competitors, you have to know your keywords.
Usually, this boils down to comparing what works long-term, what works short-term and why do keywords even stop being relevant. When you have been doing keyword research like this for years, you start to feel a bit like a robot. So, you wonder ‘could a robot be doing this instead of me?’ It could, but not just any robot.
The first multi-armed bandit I built for this purpose was quite simple. It compared the search engine rankings of the five most recent posts of five websites; one that we were optimizing and four that were our top competitors.
The bandit simply tried to predict what type of keywords are going to rank best when first published on one of the sites. It helped me discern that we should shift our focus to covering how-to articles in the niche, even the lowest-volume ones.
Two months into employing this keyword research approach and our sitewide average search engine position compared to the four competitors was #2.
Before the bandit implementation, the site was averaging #4.
Imbalanced Data is Your Greatest Ally
It’s important to understand that you don’t have to build a classifier to notice imbalanced data. If you see a single page on your website driving 95% of your website traffic, you have an imbalanced dataset on your hands.
I did build a few classifiers, however, and they did extraordinary work. At first, we used some simple ones that labeled pages based on their real traffic relative to their keyword volume and expected traffic. If I had insight into competitor traffic, we could make some real marketing magic happen with classifiers but, alas, that’s not how SEO works.
Afterwards, more complicated classification problems were created with the help of some colleagues using varied datasets (not all of them strictly related to SEO) and we managed to make some changes that had an incredible impact on site performance and ROI.
Our readers responded infinitely better to pop-up offers when visiting monetization pages compared to informative ones — 85% CTR vs. 10% CTR— so we removed the pop-ups from all guides and how-to articles.
Some of the high-volume pages were ranking quite well, but had 700% less links coming their way on a monthly basis than similar content on our sites. This helped me discern that those particular pages lacked content that people would ever re-share. We updated them with original images and relevant stats which in turn granted us a steady influx of natural links improving website performance all-around.
The most important classifier I built for this purpose was analyzing search engine rankings of our tutorials. It helped me realize that five of our ~100 pages were severely underperforming. I looked more closely into what the classifier had learned and noticed that the common denominator for the low-performance pages were random strings of numbers in place of meta titles and descriptions. I had no idea who caused this particular issue, but upon fixing it the page rankings shot upwards and our traffic had increased.
AI Generated Content — Do’s & Don’ts
Now this is a hot topic today, but that wasn’t the case back in 2019. Even so, many SEO experts I knew advocated for the use of AI-written content to quickly test keywords and determine whether these had any potential for ranking and bringing in more traffic.
I was and still am a strong opponent to this line of thinking.
About 50% of any AI-generated article on a simple topic, such as how to comb your hair, is complete gibberish. This number goes up to 90% once you change the topic to any basic-knowledge data, crypto, or finance article.
I simply don’t believe Google’s algorithm is so simple to be taken by these cheap tricks. If anything, you might be causing harm to your website by opting to avoid keywords that would have ranked properly if written by an expert.
On the other hand, AIs have proven quite useful when I utilized them to generate writer briefs and outlines for our content writers. A few clicks instead of doing 15–45 minutes of research adds up to a lot of time saved if you publish 100+ articles a month.
I believe that SEO and machine learning are going to get a lot more intertwined in the future and the process has already begun.
New SEO tools are being developed at breathtaking speed, all of them utilizing above principles (and others) to save valuable time for their customers. Similarly, SEO specialist and getting craftier and using AI to improve their marketing performance and gain an edge over their competitors that are still sleeping on machine learning.
After all, the all-powerful entity that rules over the entire theory of SEO is in fact an AI.