I've built an MCP for LM Studio for apps like Claude Desktop (or your IDE) to offload development/coding / creation tasks to LM Studio using the LM Studio SDK. You can also send a custom_prompt from Claude, so you can have AI optimising its own prompts to get outcomes from teh local LLM.
This is a fascinating presentation from Cindy Krum, who uncovers how Chrome (installed locally) appears to be used by Google to render a webpage (with JS etc) locally to save their resources and track user data to improve ads and their own ML ranking algorithm.
Hm. "Found" does not mean "OK". If this were a serious, site wide proposition in a competitive vertical on a high ranking domain, you'd be risking everything going down this route.
A much nicer way to solve the problem I think these tests are trying to address is to implement pre-rendering of your ajax pages. Take a look at something like prerender.io.
It's awesome is what it is. I wonder if you can pull the 1,2 and 3 word terms via the Google Analytics API? I'd betempted to use that data to generate popular search pages / navigation and internal link changes.