> ## Documentation Index
> Fetch the complete documentation index at: https://docs.adgentek.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# How Adgentek works

> The request lifecycle from conversation context to measured outcome.

The Agentic Ad Server was built from scratch for AI surfaces. Legacy ad servers were designed for web pages and video players, where ads occupy fixed rectangles or pre-roll slots. On an AI surface, the "page" is a live interaction, and intent is expressed in natural language. That requires a different architecture.

## The request lifecycle

<Steps>
  <Step title="Context in">
    Your integration passes the conversational context for the current turn: the user's message and your AI's response.
  </Step>

  <Step title="Signals derived">
    The Ad Server classifies the intent into one of nine buckets (see [Intent classification](/publishers/concepts/intent-classification)) and extracts structured entities: product categories, price ranges, comparison contexts.
  </Step>

  <Step title="Demand evaluated">
    Derived signals, never raw conversation data, are evaluated against the four-tier demand waterfall: Direct, Programmatic, CPC, and API/CPA. Which tiers are eligible depends on your integration path. See [Choosing your integration path](/publishers/choosing-your-integration-path).
  </Step>

  <Step title="Format selected">
    The Ad Server picks the format that fits the surface and the conversation state: a Spark interactive unit for a consideration moment, an action card for a transactional one.
  </Step>

  <Step title="Rendered and measured">
    The ad renders natively in your surface. Impressions, engagements, clicks, and conversions are measured per the rules in [Tracking and measurement](/publishers/concepts/tracking-and-measurement).
  </Step>
</Steps>

## What never leaves your surface

The Ad Server transmits only derived signals to demand partners: IAB content categories, extracted keywords, the intent bucket, and entity types. It never transmits raw user prompts, conversation transcripts, PII, or session-level behavioral data.

This is structural, not a policy toggle. Sensitive data is never collected into the bid path, so it cannot leak downstream. Details in [Privacy architecture](/publishers/concepts/privacy-architecture).