You can already use proprietary cloud-based LLMs like OpenAI's ChatGPT or xAI's Grok. If you explain to them in the prompt what "niche, enthusiast, passionate websites" are and how to find them, they can definitely help you and give you much better results than Google even in their current state. "Hallucinations" are a complete non-issue. If the LLM gives you two non-relevant links out of ten, with the rest being correct, that is still better than Google, where you might only get one relevant link out of fifty.

Now, thankfully, you do not have to rely on the cloud. If you have some DIY skills and a fair amount of computing power at home, you can run a setup locally that rivals cloud-based LLM searches in performance.

Unfortunately, it is somewhat of an arms race as you said. Advertisers and marketers aim to target people who stick to defaults: the ones who search for "top 5 password managers" on Google and click the first result. That is their audience. LLMs are not a complete solution. There are clever ways to use them with well-crafted prompts, and there are simpler, less effective approaches. Those who remain with default behaviors will be absorbed by the system; those who make the effort to resist stand a better chance of avoiding marketing influence.

As an example, some people began adding "site:reddit.com" to their searches in an attempt to get real opinions from real users. I can assure you, marketing firms have caught up with this tactic. Due to widespread astroturfing, I no longer consider Reddit a reliable source.