AI assistants are becoming the front door to the internet. People ask ChatGPT, Gemini or Perplexity for recommendations and act on the answer without ever clicking a result. For anyone building a product or a brand, this raises an uncomfortable question: what do these engines say about you, and can you even know?

Reach is a Lisbon-built platform that answers this at scale. This talk is a technical walkthrough of the system behind it: a pipeline that queries the major AI answer engines thousands of times a day and turns messy, non-deterministic output into structured brand analytics.

We will cover how the pipeline works end to end, why extracting brand mentions from AI answers is harder than it sounds (models invent name variants, citations lie), how sampling deals with getting a different answer every time you ask, and the cost surprises of running an LLM-heavy pipeline in production, including thinking tokens quietly becoming the biggest line item.

The session is aimed at engineers, founders and product builders working with LLMs in production, and at marketers who want to understand how AI visibility can actually be measured. No hype, no pitch: real architecture, real failure modes, and lessons you can apply to any system that extracts structured data from LLM outputs.