Improving SEO with Generative AI

Challenge

Improving SEO score and website traffic of more than 300.000 businesses using GenAI

In today’s digital marketplace, it is crucial for businesses to optimize their online presence, not only to attract new customers but also to enhance online visibility of their existing customers. Our client, who is a leading digital marketing solutions provider, empowering businesses to enhance their online presence and connecting with customers effectively, is facing these challenges on a daily basis. The client asked us to come up with a way to boost the online presence of their customers, by making use of the power of genAI and LLM’s. The provided solution was to automatically improve the online business descriptions of their 300.000 customers, to increase SEO (Search Engine Optimization) scores and website traffic. We had to act within a short timeframe and limited budget, and like many genAI use cases, we had to find the right balance between cost and quality, given the large volume of information.

Approach

Unleashing the power of LLMs, while reducing hallucinations to the bare minimum

We started our challenge with a POC on a smaller dataset, making use of the good old local laptop with jupyter notebook. We connected with different LLM’s using API’s, such as OpenAI, to finetune the solution. We selected the best LLM for our case, taking into account cost, quality and the current technical environment of the customer (mostly Azure based). While a local setup allows you to test a number of things very fast, it’s not suitable for a production setup for different reasons (performance, availability, processing power, etc.). We thus designed and implemented a target architecture, consisting of PySpark, Azure Synapse and Azure OpenAI that was suitable for processing high volumes of data. Also other options such as Azure Functions were considered.

Since we were dealing with genAI, we also faced challenges such as hallucinations: the LLM made up information about certain businesses that was not in the input data. We applied a number of techniques to reduce hallucinations to a minimum, such as advanced prompt engineering, limiting the allowed length of the business descriptions, and parameterization of the input data, to make it more understandable for the model.

Impact

A step forward in improved online presence

The first results were very positive, and while we are still monitoring the impact on SEO and website traffic on the long term, our solution has already made a significant contribution to the quality of the content of the clients website.

This challenge demonstrated how data-driven strategies and the use of AI models can also be applied within a short time frame and limited budget. For our client this was the first succesful AI delivery in a production setting and this project has laid the foundations for other GenAI projects at scale with our client.