AI in Content Marketing: Use Cases, Limitations & More

Before we dive deeper into the impact of AI in content marketing, it’s worth saying a few words about LLMs or, more specifically, Large Language Models in general.

To be precise, we’ll present some basics about ChatGPT, which is undoubtedly the most prominent tool and the basis of most AI generative tools out there.

After all, its growth in demand can be verified by Google Trends too:

Image Source: Google Trends

To take things from the start though, what is ChatGPT?

In a nutshell, ChatGPT is an LLM developed by OpenAI that is designed to respond to natural language inputs in a conversational manner. It is part of a family of language models known as GPT (Generative Pre-trained Transformer) and is built using machine learning algorithms, in order to generate human-like responses to various questions and prompts.

ChatGPT

ChatGPT is a large language model (LLM) developed by OpenAI that is designed to respond to natural language inputs in a conversational manner. It is part of a family of language models known as GPT (Generative Pre-trained Transformer). It is built using state-of-the-art machine learning algorithms and has been trained on massive amounts of text data in order to generate human-like responses to a wide range of questions and prompts.

In order to get a wider overview of how LLMs are being developed, there are 3 main stages:

1. Pre-training: This involves LLMs being pre-trained on large amounts of text data. The model is therefore trained to predict the likelihood of each word in a sentence given the surrounding context.
2. Fine-tuning: Once the LLM has been pre-trained, it can be fine-tuned on a specific task, such as question-answering or language translation.
3. Inference: Once the LLM has been fine-tuned, it can be used to generate responses to natural language prompts. We are currently in the inference stage when using ChatGPT.

So these are the 3 stages that LLMs are being developed on a higher level.

However, if we dig a little deeper, the question of “How do LLMs work?” will arise.
In essence, there are 6 steps for understanding exactly how such models work and operate:

1. Pre-processing: The input text is pre-processed to convert it into a format that can be used by the model.
   For example, the sentence “I need project management software” will be translated by the system into a sequence of numbers so that it understands it better.
2. Encoding: The pre-processed text is then encoded by the model, which involves passing it through a series of layers in a neural network architecture known as the transformer.
   Following the same example as earlier, the system understands that the person inserting that prompt is likely to buy project management software.
3. Prediction: Once the text has been encoded, the model can make predictions about the next word or phrase in the sequence.
   This means that for the sentence “I need project management software for…” the system may predict that the next phrase is “small business”.
4. Sampling: The model then selects the most probable next word or phrase and continues the process of predicting and generating text until the desired response has been generated.
   In the sentence “I need project management software for a small business to…” the system will select the most likely word or phrase to follow this sequence. So we’re talking about a very powerful autocomplete system.
5. Output: The final output is then post-processed to convert it back into human-readable text, applying various normalization and formatting techniques to ensure that the output is coherent and grammatically correct. The output may be the sentence “I need project management software for a small business to help me organize my business.”
6. Refinement: This is the process of refining the final output by giving additional prompts to the system until we reach a satisfactory level. A refined version of the previous output may be “I need project management software for a small business to help me manage client projects and employees.”

It’s worth pointing out that generative AI tools like ChatGPT are only as good as the data they are trained on, which can be considered a major limitation.

To be more specific, when it comes to the kind of training data we typically have the following:

• WebText (dataset that contains billions of words scraped from web pages on a wide range of topics)
• CommonCrawl (dataset that contains petabytes of web crawl data from thousands of websites)
• Wikipedia (dataset that contains a large amount of text from the English Wikipedia)
• BooksCorpus (dataset of over 11,000 books in the public domain that have been filtered)
• English Gigaword (large corpus of English-language news articles that has been used to train language models)
• One Billion Word Benchmark (dataset of over one billion words from news articles, books, and web pages that has been used to train language models)

Of course, there are more kinds of training data, but the ones above are the most prominent ones.

It’s also obvious that many sources, such as Wikipedia, are by no means perfect, which further supports our point that the output will be as good as the input data.

Now that we have seen some of the most basic things about Large Language Models and how they work, the question is: Why is all this important for content marketing?

The way we see things, besides specific use cases for generative AI, content marketing is heavily based on text-based content.

Since one of ChatGPT’s main use cases is text generation, it’s inevitable that generative AI and, more specifically, generative content are becoming important for content marketing.

Of course there are many content types like videos and images, but for the most part we’re focusing on written content, such as blog posts, for which generative AI tools can contribute, considering they’re text-based.

When it comes to the adoption stage, by taking the diffusion of innovation theory to make things clearer, we would say that we’re somewhere between the innovators and early adopters.

Whether this will reach mass adoption is hard to tell. We’ll have to wait and see and monitor all changes that happen, especially in the content marketing industry.

It’s no secret that there are many use cases that ChatGPT can be used for.

After all, we’re certain you see them almost every day on LinkedIn, YouTube, and other channels.

However, below we’ve listed some of the most prominent ones that are used in the content marketing industry, including their limitations.

Use Case #1: Blog topic ideas

The first use case that generative AI can contribute to is the generation of blog topic ideas.

As you can see below, the prompt we used is: Give me a few contrarian blog post titles for the topic “ecommerce marketing”.

From the 8 results, we’d say that only idea #6 is worthy; the rest are rather unhelpful and even “dangerous” for an eCommerce business to write about, one might say.

What we can do however is ask ChatGPT to refine its output by prompting it to focus on eCommerce marketing and truisms, and present an antithesis to these truisms.

As we can see, the results are much better. Besides ideas #4 and #8, the rest are rather helpful and could easily be used by a content marketing team of an eCommerce business, to write some blog posts.

Use Case #2: HTML table

Our second use case involved generating an HTML table.

Our prompt to ChatGPT was: Can you create an HTML table with as many as 10 comparison points on the topic of “mountain bike vs road bike”?

The HTML code looked good, so what we did next was to validate the output to make sure there were no errors.

Note: The tool we used to validate the code was Code Beautify.

Indeed, the code had no errors so ChatGPT’s output was great.

Use Case #3: Schema Markup

A third use case such a tool can prove to be useful for is the generation of Schema Markup.

What we basically did was ask ChatGPT to create an article schema markup for our Clearscope guide.

At a quick glance we notice that the tool’s output looks fine.

Let’s validate it nevertheless with Schema Markup Validator.

At first sight, we notice that the output can be used as it is.

However, after taking a better look at the output, we notice that the date published and the date modified are wrong.

We asked the tool the reason for this and whether it scrapes the content to provide the schema, and the answer was essentially no.

This is based on the data the tool has been trained on and it won’t crawl the content to give us a schema markup that’s tailored to a specific piece of content, which is of course a limitation, since we can’t use it for something dynamic.

Use Case #4: Keyword research

As you already know, keyword research is an important part of a content marketing process, so it only makes sense it’s one of our use cases.

After conducting customer research for one of our clients, we asked ChatGPT to give us a few keywords we should target to reach our audience, which includes:

• HR specialists
• HR managers
• Online training providers

And more.

As you can see below, the output wasn’t very satisfactory, considering the keywords were too generic and couldn’t help us with the task at hand.

So what we had to do was dive deeper and refine the results, to optimize the tool’s output.

That’s why we took one of the terms it provided us with (employee development) and asked it to dive deeper into that topic, thus presenting us with various semantically relevant keywords.

What we did then was take those terms and insert them into Ahrefs to validate them.

It turned out that the terms existed on Ahrefs and had a decent search volume for the most part.

Although this is a rather simplified way of doing things, it’s a good starting point.

Let’s move on to the next use case.

Use Case #5: Audience research

Audience research is an integral part of our process here at Minuttia.

After conducting the same customer research as earlier, we asked the tool to give us websites these people tend to visit, what social media accounts they follow, media publications they consume content from, and basically anything else to help us understand where that audience spends its time online.

Although some of the suggestions seem promising, we still have to validate them for relevancy and other factors (such as if they’re still active) to include them in our audience research.

With our second prompt we try to get podcasts as one of the elements we want to include in our audience research for our client.

After some back and forth, we would say that the results look quite promising:

Obviously they need further validation such as to confirm they’re still active, but it’s clear that audience research is a good use case for ChatGPT to assist us with.

Use Case #6: Content optimization

While content optimization involves various aspects, in this case we’ll focus on the aspect of inserting semantically related keywords and entities into a piece of content, to ensure it’s relevant for search engines.

Assuming we’re writing a guide on how to create an online course, we asked ChatGPT to give us as many entities and semantically relevant terms as possible, to include in the piece.

Some of the keywords seem promising, but we asked ChatGPT for more keywords and entities as we wanted to dive deeper into the topic of creating an online course.

We were provided with various other terms, so what we did was to generate a report with Clearscope to identify common terms between the two tools.

After two prompts, ChatGPT generated 120 terms while Clearscope generated 65 terms (custom terms excluded). Based on our calculations, there were only 8 common terms (exact match) between the two lists and the total overlap was 12.71%.

To be fair, the overlap percentage is relatively low. Since Clearscope is a state-of-the-art tool and many people trust it to get the best entities, then this is an indication that ChatGPT’s recommendations aren’t as good.

We actually took things a step further and asked our friends at LearnWorlds to determine the relevance of ChatGPT’s output.

The (strict) relevance distribution among the terms is as follows:

1. Not Relevant: 21 or 17.50% of all terms
2. Somehow Relevant: 47 or 39.17% of all terms
3. Relevant: 52 or 43.33% of all terms

Note: You can find the Google Sheet we used to make the calculations here.

This shows that in general ChatGPT’s terms are in the right direction, but it’s important to keep in mind that, according to the people at LearnWorlds, these 120 terms could be covered in over 2 or 3 blog posts.

Use Case #7: Formulas & Python Regex

This use case is quite different from the rest and involves generating formulas and Python regular expressions.

We actually asked ChatGPT to create a regex that extracts the name “George” from a piece of text, in order to set up a Zap.

We had several back-and-forth discussions with the tool in order to refine the result since the first output wasn’t right.

However, after a few refining prompts, we managed to get the result we needed from Zapier Formatter!

For someone not very familiar with Python formulas and such, ChatGPT can prove to be quite useful.

Use Case #8: Customer research

Our next use case involves customer research.

Here, we assumed that one of our competitors is project management software Asana, so we asked ChatGPT to give us a few weak points of Asana regarding its pricing, product, customer support, and more.

All in order to use those points in a comparison page.

The tool gave us 5 different points that we could potentially use.

However, after we asked the tool where those points come from and if they can be trusted, the response showed that we can’t blindly trust such systems when doing customer research.

At the end of the day, the ultimate source of truth for this use case is the customers themselves.

Use Case #9: Blog post outline

Creating blog post outlines is another use case worth looking into.
What we did is ask ChatGPT to create an outline for the keyword “what is field service management”.

Although it looks fine at a first glance, the issue is that if many people use the same prompt, they’ll get more or less the same output, thus resulting in a large amount of content being more or less the same.

This essentially means that the value of content will periodically decrease, which is actually one of the predictions when it comes to the impact of AI content.

After a bit of digging, we notice that many of ChatGPT’s suggestions are based on existing knowledge and lack product knowledge and information gain, perspective, experience and other elements which are essential for differentiating ourselves from what’s already out there.

Note: You can find the brief ChatGPT created here.

Overall, we’d say that when it comes to blog post outlines from generative AI tools, this wouldn’t be something helpful to stand out from the crowd.

And if you decided to use ChatGPT for this, then significant input from a human would be required to make it reach a satisfactory level.

Use Case #10: Content repurposing

When it comes to a use case like content repurposing, things are more simple.

Our prompt is based on the intro of a blog post we published a few weeks ago on our blog.

So we asked ChatGPT to create a LinkedIn post based on that intro.

In general, we would say it looks OK.

Although some refinement would be great, the post is good enough to consider content repurposing a use case of generative AI.

Use Case #11: Email outreach

When it comes to email outreach, our prompt included the URL of a case study we made and want to promote, as well as a set of instructions for crafting the email.

The instructions also include the prompt of including a subtle joke on SEO.

After a lot of back and forth and several refinements, there was a satisfactory result.

However, we would say that email outreach as a use case of generative AI is quite questionable, although some elements can be used.

Use Case #12: Title tags & Meta descriptions

Title tags and meta descriptions are definitely a great use case.

After all, it’s something that can easily replace the manual work of a human.

In this case, we asked for a title tag and a meta description for the keyword: online collaboration tools.

We would say that the output seems great.

It’s inclusive, to the point and could easily be used for a piece of content!

Use Case #13: Content creation

Last but not least, content creation is yet another use case, although complicated.

Someone could say:

Although the statement is understandable, we have 3 main arguments:

1. The data these systems are trained on make the probability of getting cookie-cutter content that’s not original really high.
2. The output has to be refined (edited, fact-checked) which significantly increases the time-to-final-draft and increases the final cost per content piece produced.
3. There’s a high chance that content that’s created with the sole purpose of getting organic visibility and clicks will be depreciated by Google moving forward.

Since Google is the most prominent search engine, you should take into account the probability of creating something that could put your website at risk.

Plus, another question that arises is who’s going to read AI content and why?

Image Source: Twitter

So far, it’s clear that AI content has both pros and cons.

Although it can be helpful in some use cases, in others it shows significant limitations.

In our opinion, there are 4 major limitations to take into consideration:

1. Micro-tasks: Generative AI is good for a variety of micro-tasks. However, for many of these tasks there’s usually a better replacement (such as Schema Markup Plugins).
2. Mistakes: In many cases it makes mistakes, has serious limitations (such as knowledge cutoff date) and needs refinement.
3. Originality: It doesn’t have original thoughts, opinions or ideas. It’s based on existing data and knowledge.
4. A tool that improves: It’s a tool; not the be-all and end-all to all our problems and challenges. The more we interact with it, the better it gets.

As we mentioned, that was our own opinion.

When it comes to the limitations of ChatGPT, what better way to find out than to ask the tool about them?

According to ChatGPT, its limitations include:

• Lack of real-world knowledge
• Biases in the training data
• Inability to understand the context
• Limited ability to handle complex reasoning
• Lack of emotional intelligence

Despite the limitations we just mentioned, it’s only natural for people working in the content marketing industry to be skeptical about their careers.

Should content marketers be afraid?

The way we see things, they should only be afraid if…

• …they’re topic-agnostic writers not specializing in a specific topic
• …their content creation process doesn’t take into account the product/service they’re writing about
• …their content creation process doesn’t involve interviewing customers, product owners, and other stakeholders
• …they’re a content writer who only works in specific content formats (like SEO content)
• …their content creation process starts and ends with words, not expanding into other areas like graphic design
• …they’re reluctant to monitor the changes in an industry and interact with these systems

Meaning that for content writers to keep their seats at the table they need to constantly stay up-to-date with what’s happening, work on other content formats, talk to their customers, and last but not least reskill and upskill.

Let’s wrap things up with some final words.

Whatever your stance on AI in content marketing may be, one thing’s for sure: You can’t ignore it.

After all, one way or the other, it will continue to impact the world of content marketing.

What you can do is see this as an opportunity to improve your skills and excel at your content creation processes, in order to be one step ahead.

That’s what we also do here at Minuttia; we’re staying up to date with everything that’s happening in AI content, but at the same time we won’t integrate AI content into our service line either.

We want our human content writers to keep their seats at the table, while embracing the changes happening around us!