For the last two decades, we've accepted an uncomfortable compromise: content that tries to please both humans and search engines ends up underwhelming both. Now there's a third constituency—AI models—and the compromise is untenable.

Marketing’s new audience is AI itself: people now start buying journeys by asking models like ChatGPT, Gemini, and Perplexity, which act as influential intermediaries deciding which brands to recommend. To win those recommendations, brands must treat models as rational, verification-oriented readers—using clear, specific, and consistent claims backed by evidence across sites, docs, and third-party sources. This unlocks a compounding advantage: AI systems can “show their work,” letting marketers diagnose how they’re being evaluated and then systematically adjust content so models—and therefore buyers—see them as the right fit.

Turns out that they can. Anthropic’s 2025 research shows advanced Claude models can sometimes detect and describe artificial “thoughts” injected into their own activations, providing the first causal evidence of genuine introspection rather than post-hoc storytelling—about a 20% success rate with zero false positives. The effect is strongest for abstract concepts and appears to rely on multiple specialized self-monitoring circuits that emerged through alignment training, not just scale. While this doesn’t prove consciousness, it demonstrates that leading models can access and report on parts of their internal state, with significant implications for interpretability, alignment, and how we evaluate future AI systems.