Unlike standard patent search, “trend analysis” is difficult to benchmark. Give the same dataset to three different experts, and you’ll likely get three different perspectives based on how they categorize technologies. There is rarely a single standard answer.
So, how do we test an AI? I fed the exact same trend analysis prompt into Perplexity Patents and ran it 3 times in parallel. The result? 3 completely different answers.
To dig deeper, I changed a topic and asked: “Please provide the most important patents published by Apple Inc. in 2025.” I ran this twice.
Run 1: It cited 9 patents. Run 2: It cited 17 patents. The Overlap: Zero. None. The identified technology trends were also completely distinct.
Perplexity Patents isn’t analyzing; it feels more like “spray and pray.” It grabs whatever data points it hits first, without a consistent underlying logic.
Does this mean it’s useless? Not necessarily. Since patent analysis is subjective, it can serve as a random idea generator. But relying on it for accurate landscape analysis is risky when we can’t see the logic (or if there isn’t any).
A final thought: If AI eventually evolves to fully automate patent analysis, skipping the research and jumping straight to the conclusion, is that actually a good thing for us?