Just came to a similar conclusion after testing different ways to get AIs to use my llm.txt content.
Far from exhaustive, but these are my early observations:

1. Simply adding an example.com/llm.txt file did not change how ChatGPT talked about my brand in my tests. Clearly not yet treated like a standard such as robots.txt.

2. Linking to the file from the footer also had very limited visible impact. From what I can see, the model seems to grab specific chunks of a page and does not reliably follow internal links the way a classic web crawler might.

3. Trying to “force” it with inline instructions also failed. I tested messages in the source and even tiny low contrast text telling the model to read the file. None of that made ChatGPT read my text and shape the answer based on it.

4. The only approach that consistently nudged the output was putting my llm.txt style markdown directly on the homepage inside an accordion. Only then ChatGPT started using the positioning I wrote when asked about the brand, although it still cherry picks. For example, I listed a long list of publications that featured our studies and it only wrote about it when asked very directly.

All of this is based on very limited testing and one site, so treat it as anecdotal.