The retrieval pipeline matters far more than whether an LLM can technically parse an accordion.
I tested this myself. One model couldn't fetch the page at all. Another surfaced the content through a completely different route. Both are frontier models, but the tools underneath behave totally differently. So how the content gets accessed makes all the difference.
The retrieval could be a live HTML fetch with no JavaScript, a partial render, a cached snapshot, or a search pipeline that returns only relevant snippets instead of the full document. Depending on which path gets used, the model might see everything or nothing at all.
That's why "optimise for LLMs" feels premature. It's not one crawler - it's a fragmented set of systems with no shared standard.
On the accordion question, you don't necessarily need to flatten everything. Google spent years getting good at rendering messy frontends, and these pipelines will probably catch up. But right now, anything that depends on JavaScript-triggered state changes, hidden DOM nodes, or user interaction introduces risk depending on the pipeline.
There's also a versioning issue people aren't talking about much. Some systems hit an older cached version. Some hit a partially indexed snapshot. Some do a live scrape of whatever's immediately available. Looks like the same page, but they produce completely different outputs depending on how it's accessed.
It's messy because it's just a pile of loosely coordinated hacks that people call a strategy.
