I would separate training permission from search eligibility in the conclusions. GPTBot and OAI-SearchBot have different jobs: blocking GPTBot is not a guarantee of zero ChatGPT search traffic. A browser-UA fetch of robots.txt also cannot establish what the CDN serves to the actual crawler.
For a follow-up, test representative cited URLs rather than only homepages, include OAI-SearchBot, and accept sitemap indexes as well as urlsets. Then measure citations on a fixed prompt set separately from those access checks. That would make the comparison much stronger without treating llms.txt as a ranking requirement.
I work on QuanticData, where we have an AI crawler checker. Keeping a rules check separate from evidence of a real bot visit or citation is an important distinction for this kind of audit.