Balázs Turán's two-week tracking experiment reveals a finding that reshapes how you should think about ChatGPT optimization: switching from fast to reasoning mode changes which sources ChatGPT cites more than it changes what ChatGPT says. In fast mode, ChatGPT averages 2.6 sources per answer with a 50% citation rate. In reasoning mode: 4.5 sources per answer, 68% citation rate — a 73% increase in source count. Only 25.6% of cited domains appear in both modes.

The source composition also shifts dramatically in reasoning mode: government and academic sources jump from 1.9% to 8.8% of citations, while Reddit's share falls from 15% to 7%. This means optimizing for ChatGPT fast-mode users (who tend to get Reddit and news sites) and reasoning-mode users (who get academic and government sources) requires materially different citation infrastructure. The report also replicates the 97% llms.txt finding: in an Ahrefs study of 137,000 sites, 97% of llms.txt files received zero crawler requests from AI bots.

Esteve Castells (LLM Pulse) lays out the measurement framework that makes Perplexity the most actionable AI surface to monitor: mention rate, citation rate, citation order, share of voice, sentiment, and competitive overlap. Perplexity places numbered source links directly in answers — making it uniquely auditable compared to ChatGPT or Gemini, which surface mentions without reliable URL attribution. Practical threshold: 50–200 prompts per tracking cycle for statistically meaningful data. Key freshness finding: Perplexity consistently favors content updated within 30–90 days for time-sensitive queries, making it the most recency-sensitive of the major AI engines.

"The reasoning mode finding from Creative Data Engineers is the thing I keep coming back to. Only 25.6% domain overlap between ChatGPT fast and reasoning mode means you can be optimized for one and invisible in the other — on the same platform, on the same day. That's not a gap you close with a single content strategy. The Perplexity data compounds this: 57.8% citation rate vs GPT-4o's 15.6% on identical content. We are past the point where 'optimize for AI search' is a coherent instruction. You need a position on each engine's retrieval architecture, not one GEO strategy applied uniformly across all of them."