There's more to say: How AI content generation using GPT-3 can enrich existing content and boost SEO

What is this about?

Content is the essential aspects of Search Engine Optimization (SEO). It is the content on your website that helps search engines understand what your website is about and how it can help users. Creating high-quality, helpful, and relevant content is the key to a successful SEO strategy. This means making content that is not only keyword-rich but also informative and engaging.

And what’s the problem?

Content is a challenge in SEO because it is challenging to produce quality content that is keyword rich and informative. In addition, it isn't easy to get other websites to link to your content if it is not of high quality. Furthermore, content production is expensive for a variety of reasons. First, it requires significant time and effort to research, write, and edit quality content. Second, it often requires specialized knowledge or skills, driving up costs. All these factors can make content production a significant expense for businesses. However, there are a few ways to offset these costs. First, consider producing content in-house. This can save on costs associated with hiring outside help. Second, take advantage of free or low-cost resources, such as open-source software or free images. Finally, be efficient in your content production process to minimize waste and maximize results.

Here, efficiency can also mean completely rethinking the content creation process.

AI-powered content enrichment

In a rapidly developing world, it's crucial to provide accurate and up-to-date information. Unfortunately, this isn't always the case. Many corporate websites are dead empty, and the product details page consists of just a single table of product properties. In the business world, AI has the potential to transform how companies create content. Here in three ways AI can change the content game:

Smarter content creation

AI can help businesses create more targeted and relevant content for their audience. By analyzing data points such as customer behavior and demographic information, AI can help marketers understand what type of content is most likely to resonate with their target audience. This ultimately leads to more effective content marketing campaigns and a higher ROI.

Increased efficiency

AI can also help businesses save time and resources for content creation. For example, AI-powered tools can be used to automate the research and writing process. This can free up valuable time for marketers and allow them to focus on other aspects of their job. Additionally, AI can be used to help identify and correct errors in the content before it’s published, which can save businesses from embarrassing and costly mistakes.

Improved personalization

With AI, businesses can create more personalized content experiences for their customers. By leveraging data points such as location, past interactions, and browsing behavior, AI can help businesses deliver content that is more relevant and tailored to the individual. This type of personalization can lead to increased engagement and conversions.

AI is quickly changing the landscape of content marketing. By understanding how AI can be used to create more thoughtful, more efficient, and more personalized content, businesses can stay ahead of the curve and position themselves for success.

Text generation with GPT-3

OpenAI GPT-3 can be a game-changer in the world of artificial intelligence. It’s a deep learning neural network model with over 175 billion machine learning parameters and the third generation of OpenAI’s Generative Pre-trained Transformer (GPT) model. GPT-3 can be used for a variety of tasks, including natural language processing, question answering, and machine translation, e.g.:

What sets GPT-3 apart from other AI models is its ability to generate human-like text. This is because GPT-3 has been trained on a large dataset of natural language text. While it is still in its early stages, it has the potential to revolutionize the field of artificial intelligence.

The interaction with GPT-3 is done via an API. The model itself cannot be downloaded. The interaction with the model is done via natural language. It can be divided into three types of requests (“learning within the context”), whereby it always depends on the task, the data, the target, and the model which request leads to the best result. Typically, the following settings are used for performing a task with the language model:

Zero-shot (Source)

Zero-shot is a quick to implement and pleasant way of performing a task with GPT-3. Only a natural language instruction describing the task is given, and the model tries its best to disambiguate and complete the task. In domains with technical language or other specifics, more context should also help the model answer if specific output formats or restrictions are to be observed.

One-shot (Source)

One-shot means that that only one example and a natural language description of the task as input is allowed.

Few-shot (Source)

In Few-shot settings, the model is given a few demonstrations of the task at inference time as conditioning. The resulting context makes it easier for the model to resolve ambiguities and better understand the intent behind the request. Due to the model architecture, the context size is limited in terms of the maximum input size.

Another way is to fine-tune the models, which I will present in a separate article on this blog.

aiSEO: Content generation strategies for GPT-3

The presented ways of interacting with the language model allow for different strategies on how to adapt, create or extend content for SEO. In this introductory article I want to focus on Zero-shot, but there is much more to explore. In particular, the latest Davinci models, which also support the “Insert” and “Edit” modes in addition to the familiar “Completion”, open new variations.

Software solutions that help supplement or transform existing content are often requested. I will look at the cases of “Headline generation”, “Abstract generation” and “Article generation based on keywords” in the following. Furthermore, let’s see how we can transform tables into text.

In the following we see in each case an input window in which I formulate my prompt. The green text components have been generated by the text model. Each time the first generated variant was taken over without changes. GPT-3 makes it possible to repeat the generation as often as desired and finally to take the best result. The individual results can be completely different.

Headline generation 

Abstract generation

Article generation based on keywords

Turn tables into content

That GPT-3 is powerful is also shown by the fact that not only pure text inputs can be interpreted. Tabular data can also be included in the query and combined with a description of the task to form a prompt that GPT-3 can understand. This is valuable for generating small texts from tabular data such as product information, etc. The following example shows the task of generating a text from a table of city names, population sizes, and sentiment values. Again: The green part in the image was generated by GPT-3.

As you can imagine, these are just elementary examples that are possible with this text model. However, they should suffice to present the use case SEO and AI and to go into the implementation in the following.

Implementation

The integration of GPT-3 into your software projects is straightforward. A registration with OpenAI is required. Various SDKs are available, including Python and node.js. However, requests can also be made simply via cURL since it is ultimately an HTTP-based API.

In addition to the examples (depending on the type of query) and the natural language description of the task as a whole package in the form of a prompt, parameters (source) must also be defined that influence the behavior of the model and which, apart from certain guideline values, can best be defined by trial and error.

If this request is transmitted in the form shown, GPT-3 generates a completion of the input. The result looks as follows:

“Dr. Frederik Bäumer is a Senior Manager at Arvato Systems. He has a Ph.D. in Information Systems from the University of Münster and has been with Arvato Systems since 2006. In his current role, he is responsible for the development and operation of Arvato Systems' SAP-based solutions. He has also worked on several international projects, including the development of a global SAP template for a major automotive manufacturer. Dr. Bäumer is a strong advocate for the use of SAP software in the automotive industry. He has presented at numerous conferences on the topic and has published several papers on the subject. He is also a member of the German SAP User Group (DSAG) and the German Association for Information Systems (GI). In his free time, Dr. Bäumer enjoys spending time with his family, traveling, and playing tennis.”

What a beautiful text. Except for the fact that the non-bold text parts in represent correct language but incorrect information. The quality of the generated texts strongly depends on the question asked and the examples provided to the AI. Therefore, some practice is required to be able to generate good texts. I also chose this example to show that when you ask a question, you already must consider whether the model can know the answer at all. Especially the latest facts or world events cannot be known to the model.

Sum up

As we have seen, AI is not the magic bullet in content creation. Nevertheless: It speeds up the creative process immensely and leads to more content while reducing costs. However, an editor is required to review and approve the content - on the one hand, for self-protection in case of incorrect information, and on the other hand, because OpenAI prescribes this for the productive operation of GPT-3. AI-generated content must not be published without a review process. In the AI world, we call this “human in the loop”.

That’s it for now.

If you liked this article, you might be interested to know that AI wrote 46% of it. Fascinating, isn't it?