Anyone that has spoken to me over the last nine months will know I have been deep down a rabbit hole of AI. I have completely engrossed myself in how AI works, and at Ellipsis we’ve rapidly adopted it into our processes.
The biggest beneficiary has been our SEO Content work, which makes up most of our day-to-day work. We now have a set of internal tools which are optimised specifically for creating content for WordPress businesses. This set of tools is significantly better than anything available commercially, and helps us ensure our content gets rankings and conversions.
We call the system FALCON, and with it we have improved SEO Content results by about 30%. This is how it all works.
This post is interesting as a snapshot in time! But it is out of date now. FALCON AI is now massively more powerful, gets even better results than what we’ve trailed here, and up-to-date information can be found on this dedicated page: https://getellipsis.com/falcon-ai/
Overview: what can AI do?
We’re interested in four types of AI here:
- text classification (what category is a text in?)
- text generation (create a mostly-original response)
- prediction (what will happen next, based on past data?)
- natural language understanding (what does text mean?)
AI in 2021 is well suited to these use cases. Each of these types requires a “model” that’s used to generate the output. There are two types of these: self-trained and pre-trained.
Self-trained are custom models built on your own data for your use case. This is powered by Machine Learning (ML), so you often see ML used interchangeably here. Pre-trained models are trained with somebody else’s data. These tend to be much more general, but a good general model is very powerful.
If you want to know whether a new contact form enquiry is a sales email or customer support request, you’re probably fine using a pre-trained model. If you want to know which team a specific support message needs to be routed to, you’ll probably need a self-trained custom model.
The big breakthrough in the last year or so has been pre-trained models getting significantly better. The likes of Open AI’s GPT-3, which does text generation extremely well, have opened up many more possibilities. This has been combined with training custom models becoming much more accessible (shout out to Ellipsis client Akkio): previously you needed a Data Scientist to do this.
This is the context on what AI can do, and how we can use it. Let’s now look at the specific use cases Ellipsis has for each of these in our SEO Content process.
Machine learning to increase the success rate of keyword and title combinations
One of the first stages in the SEO Content production process is keyword selection. We’ll identify which topic cluster we want the client’s content to fit in, and then look for a good target keyword for the post. If you pick the wrong keyword, nothing you can do with content creation will make the post work.
With FALCON, we’re using a custom Machine Learning model to predict how well a keyword and title combination will perform, based on our internal data on previous post success. This custom model alone is single-handedly responsible for the lion’s share of the 30% boost to our SEO Content Performance.
We have four versions of this, to check:
- Existing keywords
- Related keywords
- Long tail keywords
- Keywords on content that is already live
The first one does a simple check of the keyword you’ve given it, and it can check up to 100 keywords at a time. The second looks up related keywords and runs them through the prediction at scale (up to 100 at a time), showing you related keywords that would work. This in turn can handle 100 base keywords at a time, so we’re checking 1000 keywords in minutes. The third is similar but goes deep into long tail keywords. The final version checks to see if a different target keyword on an existing post would perform better.
Through our use of FALCON, we’re seeing both increased performance of the content we produce — as we have high confidence the post is going to rank — and improvements in the keywords we can find, as the AI makes it easier to surface keywords.
Classifying search intent using BERT
Google is interested in providing the best possible user experience to the searcher. This means providing a complete answer to whatever they’ve looked for, quickly. In order to do this, Google has to understand what the searcher is looking for, and if the results it’s showing provide the answer to this. We call this search intent.
We understand what the searcher is looking for by looking at the results Google shows.
BERT is a Natural Language Understanding (NLU) methodology open sourced by Google in 2018. BERT is used by Google to understand search intent in context: if you search for “Queen albums” it’ll understand that Queen in this context means the band, not Queen the person.
Google is using the methodology to understand what the searcher is looking for. We’re using it to do the same with FALCON, by looking at the results and using them to classify the search intent. We’re using a pre-trained BERT model to do this.
This comes into our keyword research process. You can do this manually, but manually it tends to be done on the search query and not the search results: it’s fairly intuitive “buy WordPress themes” is a purchase-intent keyword. Google increasingly throws unexpected results, though, as it’s responding to real-world users, so using BERT lets us look at the top 10 results and make a classification based on all of them – and it does it in about half a second.
Natural Language Processing for identifying topics
Google wants to understand what the searcher is looking for. In order to determine where to rank content, it needs to understand what’s on a page so that it can understand to what extent the result matches the searcher’s need.
Google uses Natural Language Processing (NLP) for this. NLP lets Google take a URL and understand the entities and topics contained within the page. Once it knows what’s on the page, it can understand where to rank it.
NLP has long been a cutting-edge area of SEO, and it’s been a core part of our content process for the last three years. Commercial tools like Clearscope, MarketMuse, and Frase have popularised the methodology: they’ll look at the first 10-30 results for a target keyword, and then aggregate the topics within those results. That will give you a list of 30-50 topics for you to cover in the post, and covering them lets you ensure you’ve done a good job of answering all the questions a searcher will have.
This is a tried-and-tested process, and we’ve had good results from this for a number of years.
Commercial tools have some limitations, though. The primary one is the AI: NLP is good, but it has limitations. The tools are reliant on good results out of the NLP they use. Google, for example, is only able to identify about 18% of the topics on a page, and it often misidentifies them.
We see the same with commercial tools: it’s typical for them to use one NLP provider, so they’ll miss out on topics and entities posts need to cover. If you’re basing your content off an incomplete list of topics, you’re missing out.
As you can see in this example, I’ve run Google’s NLP demo on the top ranking post for “best WordPress hosting”:
The classification has mostly worked, but Google thinks that Hostinger is a person and not a company.
This has led to us to develop our own FALCON NLP solution, to better get WordPress-specific topics.
With our in-house NLP solution, we can get better topic identification. This lets us produce more complete content even than competing content produced with commercial tools.
If you can get better identification of topics to cover and make sure your content lets Google identify your topics, you’ve got a competitive advantage. This is what our content has.
We’re also using NLP in our keyword research stage to automatically group keywords into topic clusters. Previously this was a slow manual process, so the automation is extremely useful.
Custom GPT-3 models for title generation
I highlighted GPT-3 above. GPT-3 is remarkable technology made by Open AI. It does text generation, and it does a good job of it. GPT-3 is trained on 175 billion parameters and can write short text that is indistinguishable from what a human can do. It’s pretty incredible.
We’re not using GPT-3 for any long-form content generation, as I’ll get to later. What we are doing, though, is using custom models for specific parts of the content process. This is pretty powerful stuff.
A lot of the hype for GPT-3 comes from what you can do with it. There’s a long list of apps built on it. “AI copywriting” is one of the most obvious areas for GPT-3, and there’s currently an arms race to build out tools doing these. All of these tools are built on the GPT-3 API from Open AI.
The limitation with commercial tools is that you’re reliant on generic prompts. The “blog post title generator” you’ll get from an “AI copywriter” SaaS needs to work with all types of titles. We need something specifically optimised for WordPress content.
We’re using AI to generate extremely well SEO-optimised titles. With FALCON we’re able to take a target keyword, look up the top results, and then use GPT-3 to generate titles similar to the top ranking content results (we filter out non-blog post results). The output is a title perfectly optimised for Google, as it’s based on what’s already ranking.
This is where AI is much better than humans: the AI can look at the nuances of the results and generate multiple versions on the fly. We’re thus generating multiple title options and running them through the machine learning step described above: FALCON will then output the winning title.
We’re not delivering any time or cost savings here (if anything it’s more time-consuming and expensive), but we can now evaluate many more options at scale and give our clients the best ones. My thanks to Dr Oliver Crook at Oxford University for support on this.
An obvious question is: what about AI generated content? I’m pretty bullish on this being a terrible idea, as I’ll get to next.
What about AI generated content?
We are NOT using GPT-3 for long-form content generation. GPT-3 is terrible for long form content as it doesn’t know what’s it’s talking about. It has no notion of what the truth is: it can connect words together, but it has no idea what they mean.
When Google values subject matter authority more and more, using an AI that has no idea what it’s talking about is a recipe for disaster.
We’re therefore using text generation at the margins of our content process to make improvements and find efficiencies, but we’re not using GPT-3 for long-form content generation.
If anything, the rise of this makes us happier to double down on working with subject matter experts. Those experts are expensive, but if others start trying to auto-generate content that could be nonsense, we and our clients will have even more of a competitive advantage. Bring it on!
The FALCON system and how you can get it
Our results so far have shown a 30% improvement in SEO Content results since we started seriously implementing AI into our content process. Since the start of this year, we’ve started realising those results for clients.
The system as a whole is labelled the FALCON system. It’s our collection of in-house tools designed to increase the success of the SEO Content work we do for clients.
FALCON is now included in all of our Content Growth packages for SEO Content. You can see detail here, or get in touch to see how we can help you.