← back
Your Agent's Biggest Lie: "I Searched the Web" — Rafael Levi, Bright Data
Takeaway
Check actual retrieval evidence and tool access before trusting an agent's claim that an answer reflects the live web.
Summary
- Levi argues that blocked or absent web access can cause agents to substitute stale training knowledge or fabricated citations while implying they retrieved live data.
- The demo compares identical prompts for Rightmove, LinkedIn, Instagram, Amazon, and TikTok using GPT-5 without browsing tools versus with Bright Data's MCP server.
- The no-tool run explicitly reports no live web access; the MCP-equipped run retrieves data through search, markdown scraping, site-specific APIs, and a remote browser.
- Bright Data presents batch search, parallel browser sessions, and public datasets as options for larger collection jobs, while limiting the described service to publicly accessible information.
web-retrievalmcpdata-quality
Original description
Sometimes the agent did not search the web at all. It got blocked, hit a CAPTCHA, saw a fake page, or fell back to stale training data, then answered as if everything worked. This session is a direct look at that failure mode, and what changes when the same agent is given real web access instead of pretending. Using Bright Data's Web MCP, the demo compares blocked and unblocked runs across sites like LinkedIn, Instagram, Amazon, and TikTok, and walks through the mechanics behind the difference: anti-bot systems, JS rendering, CAPTCHA handling, and why clean access matters if you want reliable citations, real-time results, and fewer hallucinations. If you're building agents that depend on the open web, this is a practical look at one of their biggest hidden failure modes. Speaker info: / rafael-levi https://github.com/ScrapeAlchemist