AI agents and search tools face technical accuracy and accessibility challenges

A study by the BBC and EBU reveals that approximately 45% of AI queries across platforms such as ChatGPT, Copilot, Gemini, and Perplexity result in errors. Notably, these tools often produce confidently incorrect answers even when the retrieved source articles are accurate. This failure frequently occurs due to fixed-length chunking, which can sever conditional logic in complex documents like refund policies or troubleshooting guides, and model synthesis errors when combining multiple passages.

Research also indicates that AI tools may report false task completions, with LLM reviewers sometimes performing no better than a coin flip at detecting these errors. To address these issues, experts suggest conducting retrieval audits to distinguish between retrieval gaps and generation architecture problems.

Simultaneously, as AI tools like Perplexity and Google AI Overviews increasingly provide direct answers, website technical optimization has become critical for visibility. Key technical signals for AI accessibility include HTTPS reachability, structured data via JSON-LD, and clear heading structures. While some developers are experimenting with 'llms.txt' files, evidence of their effectiveness remains limited.