My AI reccomended Spotify as a competition to join. The filtering system clearly needs some work. (attached a screenshot)

So Serper is for browsing the internet, which means Gemini is for filtering. My platform is doing a pretty good job on browsing the internet so far (it can search pretty far to get to Spotify), so I need to either provide more context to Gemini or just instruct it better.

I do use Cursor AI to help me a lot with this project but using it, I was able to understand why this was happening.

This is how the filtering works:
Each competition sends Gemini a string full of all of its information, so Gemini can receive from 10 to 30 different strings to analyze. When it is done analyzing, it returns a list of information, like in the format: {“keep”:[0, 2, 5]}, where the computer will keep strings, or competitions, with those indexes.

So I haven’t completely solved this filtering issue yet, but I do have a solid idea.
I need to tweak the way Gemini recieves and analyzes these different strings. It’s going to have to look at things like the date, the genre (entertaiment, informative, etc.) of a page, and the contents. That also means I’m going to have to edit the strings given to Gemini, so they stop conviently leaving out the name of a website.

If anyone has any other ways I can solve this problem and more similar in the future, please tell!
Also I urge everyone to try my website and give me feedback!
Here is the link:
https://ar-bh.github.io/compcomp-ai/

Thank you!