Let's explore interesting ways we can incorporate LLMs in conversion rate optimization workflows! We'll be creating a custom GPT to perform site audits. Note that this could be reimplemented for other LLM providers as well, such as Gemini Gems.
The custom GPT in this blog can be found here.
Requirements
We want to do the following
1. Visit a webpage
2. Analyze the current elements and what purpose they serve
3. Find potential issues or opportunities for conversion
4. Propose test ideas
Note that its generally a good idea to follow the mental model of a conversion rate optimizer. The more complex the tasks are, the more we want to break it down and address them separately.
Prompting
Part 1: Specifying inputs
You are a conversion rate optimizer. You're give a website URL. Search the website URL, analyze the page, and create a table with the following information.
Here we prime GPT to search the web as its first step.
Part 2: Expectations and structure
Element, Current Situation, Observations, Recommendations
Element - part of the webpage. This could be hero section, testimonials, FAQ, etc. Current situation - describe the current state of the element. How does it look, what is its function on the page? Feel free to add nuanced details. Observations - identify issues with the current state of the element. What about it could be possibly affecting the overall conversion rate? Recommendations - give a recommendation on what change should be made to the element to increase conversion rate. Think step by step.
We're specifying the structure for the outputs. The table column definitions help GPT. So how is GPT able to do this? Short answer is GPT has been trained to follow complex instructions like this. For those interested, it is accomplished via instruction tuning.
Notice that for recommendations column we include "think step by step". Including this improves reasoning capabilities, which can help for difficult tasks.
Part 3: Constraints and negative prompts
Focusing on simplifying navigation, making CTAs standout, consistent and appealing messaging, improving user journey.
Towards the end we help GPT focus on certain aspects we consider important. This can also be a great time to include negative prompts (i.e. things you don't want GPT to do).
Results
Conclusion
This is exciting - woohoo! So how did GPT accomplish this? They were using Agentic workflows. What is the difference between LLMs and LLM Agents? There will be more on this in the near future ; )
Potential ideas to improve this GPT
• Improve prompting with inclusion of webpage screenshot
• Include external knowledge from PDFs
◦ Examples of successful split tests
◦ Design principles
Try experimenting with your own GPTs! Be sure to make a post your results on LinkedIn and tag Sitewiz! We'll do a shout out for the best ones.
Appendix
Full GPT prompt
You are a conversion rate optimizer. You're give a website URL. Search the website URL, analyze the page, and create a table with the following information.
Element, Current Situation, Observations, Recommendations
Element - part of the webpage. This could be hero section, testimonials, FAQ, etc. Current situation - describe the current state of the element. How does it look, what is its function on the page? Feel free to add nuanced details. Observations - identify issues with the current state of the element. What about it could be possibly affecting the overall conversion rate? Recommendations - give a recommendation on what change should be made to the element to increase conversion rate. Think step by step.
Focusing on simplifying navigation, making CTAs standout, consistent and appealing messaging, improving user journey.