Your Kubernetes network policy example is spot on. I've seen that exact failure with AWS Lambda quotas - ask about the per-function concurrency reservation and it'll cite a generic EC2 instance page because both contain the word "limit".
The verification steps are basically what I do, but I'd swap the order. I always open the link first, especially for cloud services. The domain tells me almost everything I need to know. If it's docs.aws.amazon.com, there's at least a chance. If it's a "cloud数字化转型" blog post, I close the tab before the snippet even loads.
That last point about the answer standing on its own merit is critical for cost questions. If I ask about Savings Plans vs Reserved Instances and the explanation can't articulate the payment flexibility difference without citations, it's a useless answer.
Exactly, that Kubernetes network policy example is the perfect illustration of the problem. It's not just a broken citation, it's a conceptual mismatch that reveals how the retrieval works - it's matching keywords, not intent.
I do the same verification dance, but my first check is always the domain. I can't count how many times I've asked a nuanced question about a CRM's API rate limits and gotten a citation to a three-year-old company blog post about "the power of customer data." The link works, but it's just fluff.
Your final line hits the nail on the head. I've started treating the entire answer as a draft hypothesis. If it can't stand up without the citations, it's usually wrong. The citations are, at best, a starting point for my own search. For anything in marketing automation, the real answer is always in the platform's own changelog or developer docs, never in a top-of-funnel article Perplexity loves to grab.
Happy testing!