AI Content Generation: Our Research
We faced a challenge – automating the creation of high-quality articles that are well-optimized for SEO, have low AI-detector scores, accompanied by appealing title images, written in a natural, human-like style, cost-effective, and well-structured.
Finding the right solution required significant research and testing. Today, we’re ready to share our findings and offer effective strategies for content automation!
Choosing a Content Generation Approach
- Using a Donor Article: This method involves rewriting an existing article to fit your niche. It’s the simplest approach but requires access to relevant content sources.
- Generating Content Based on Keywords: In this method, the AI model analyzes provided keywords and creates unique content. The challenge here is ensuring variation in word choices and avoiding repetitive structures.
- Creating Articles Based on Company Activities: This is the most complex approach, as it requires an in-depth analysis of company data to generate highly relevant and personalized content.
Mathematical
General-Purpose
Text
Selecting the Right AI Model
The quality and cost of content generation largely depend on the AI model used. OpenAI currently offers three main types of models.
Often, the best results come from a combination of several models. For example, using a mathematical model to analyze the keywords or data of your business and determine the structure of the article, and a text model for generating content. This approach allows you to create meaningful,
well-structured and optimized articles that not only attract an audience but also improve the visibility of your site in search engines because tasks are distributed between models that work well in a certain type of tasks.
Crafting Effective Prompts
Prompt engineering is a crucial step to ensure AI delivers the expected results. Different models interpret prompts in unique ways, so there is no universal format.
When defining your requirements, consider specifying:
– Article length
– Content structure
– Writing style
– Number of links
– Use of examples
– Target audience
– Language preferences
– SEO requirements
. . . and other
We will optimize these specifications into a prompt that AI understands to ensure high-quality, tailored content.
Using Batch API: Pros & Cons
Batch is a way of processing data when queries are performed in groups, not in real time. This reduces the cost of requests to API, but at the same time increases the processing time. In the context of AI content, Batch API is used for mass generation or rewriting of texts, which makes it effective for processing large amounts of data, but less convenient for creating unique content due to lack of memory of previous requests.
The Batch API can reduce costs by up to 50%, but it has significant drawbacks:
– Longer processing time
Article generation can take up to 24 hours, making it harder to track progress.
– No memory retention
The AI does not remember previously generated articles, which can result in repetitive content if articles are based solely on keywords.
The Batch API is best suited for rewriting donor articles. For unique content creation, we generally do not recommend using it.
Automated Image Selection
To enhance articles with relevant visuals, we explored three approaches:
1. AI-Generated Images
While unique, AI-generated visuals may contain artifacts (as unreadable text, graphics or irregular shapes of figures) and require manual review, which can impact SEO.
2. Image Search via Google Search API
In this method, the AI generates a search query based on the article topic, retrieving relevant images. This approach is more SEO-friendly.
3. Template-Based Image Editor
This system assembles a final image from multiple elements (background, logo, text). While more stable, it works best for specific niches.
For most cases, we recommend the second approach—searching for images via the Google Search API.
Reducing AI Detector Scores
To ensure AI-generated content appears natural and undetectable, we employ two key strategies:
1. Highly Detailed Prompts: We provide extensive examples, rules, and stylistic requirements.
2. Third-Party Rewriting Tools: These tools refine AI-generated text to achieve a more human-like style.
While the second option is more expensive, it significantly improves SEO performance and enhances content authenticity.
Enhancing the System
Automation Platforms
- AWS: Ideal for managing complex tasks such as parsing donor articles or processing additional data.
- Make.com: A simpler solution focused on workflow automation and AI optimization. While efficient, it may not support highly complex systems.
Conclusion & Recommendations
For News Platforms:
– Utilize general or text models based on your budget.
– Retrieve images using the Google Search API or make use of existing media.
– Build your system on Make.com for automation and AWS Lambda for parsing.
For Keyword-Based Content Generation:
– Employ a mathematical model for keyword analysis and structuring.
– Use a text model for content creation.
– Access images via the Google Search API.
– Base your system on Make.com.
For Company-Specific Content Generation:
– Use a fine-tuned text model trained on your company’s data, along with an additional model for improved analysis.
– Retrieve images through the Google Search API or create them using an editor.
– Construct your system on AWS for regular data updates, formatting, and AI requests.
Other Cases
