Texterro
**Background:** Our client operates a vast marketplace that features millions of product pages across various categories. Previously, the SEO department and copywriters handled the creation of SEO content for these pages. However, this manual process was costly and time-intensive. Furthermore, the department faced restrictions in terms of the number of unique articles that could be produced, prompting the client to seek optimization solutions.
**Solution:** We initiated a research and development (R&D) process, experimenting with systems such as ERNIE, BERT, RoBERTa, and various models of GPT-2 (including 124m, 345m, 774m, and 1558m). Through a series of experiments and fine-tuning, we discovered a method for generating lengthy texts (up to 3,000 characters or the length of an A4 page) while maintaining contextual integrity. We employed a Generative Adversarial Network (GAN) approach to ensure the quality of the generated content and filter out undesirable material. Additionally, we created an interface that allowed users to choose a theme and input specific keywords.
**Results:** After seven months of R&D and three months dedicated to developing a minimum viable product (MVP), the client received a system capable of generating thousands of unique SEO texts for their marketplace in a very short timeframe. Unlike spinning technologies like Article Forge, our texts were original. The SEO department benefited from having a unique, fully customizable tool to meet their requirements. One of the models was even made available as a public demo for anyone to try out.