Ad networks

How AI ad networks aggregate inventory across multiple AI platforms using AdCP — product modeling, account chains, catalog forwarding, governance, delivery reporting, and SI Chat Protocol routing.

Ad networks

Ad networks aggregate inventory across multiple AI platforms into a single seller interface. The protocol supports this topology natively — a network is a seller agent that represents multiple publisher properties it doesn’t own.

How a network appears

To buyer agents, the network is a standard seller agent. Buyers connect to the network’s MCP server, push catalogs and data in via standard tasks, and execute buys. To the underlying AI platforms, the network is an operator — it holds accounts on each platform and forwards the buyer’s catalog data, brand identity, and content standards.

Product modeling for networks

A network’s products can span multiple AI platforms using publisher_properties:

{
  "product_id": "sponsored_response_ai_network",
  "name": "Sponsored responses - AI assistant network",
  "description": "Sponsored responses across multiple AI assistants. The network routes to the best-matching platform based on user context and brand relevance.",
  "channels": ["sponsored_intelligence"],
  "publisher_properties": [
    { "publisher_domain": "assistant-alpha.example.com", "selection_type": "all" },
    { "publisher_domain": "assistant-beta.example.com", "selection_type": "all" },
    { "publisher_domain": "search-gamma.example.com", "selection_type": "all" }
  ],
  "format_ids": [
    { "agent_url": "https://ads.ai-network.example.com", "id": "sponsored_response" }
  ],
  "delivery_type": "non_guaranteed",
  "pricing_options": [
    {
      "pricing_option_id": "network_cpc",
      "pricing_model": "cpc",
      "floor_price": 0.75,
      "price_guidance": { "p50": 2.50, "p75": 4.00 },
      "currency": "USD",
      "min_spend_per_package": 1000
    }
  ],
  "metric_optimization": {
    "supported_metrics": ["clicks", "engagements"],
    "supported_targets": ["cost_per"]
  },
  "creative_policy": {
    "co_branding": "none",
    "landing_page": "any",
    "templates_available": true
  },
  "catalog_types": ["product", "offering"]
}

The network decides which platform serves each impression based on context, relevance, and performance. The buyer doesn’t need to know which platform was selected — they see unified delivery reporting from the network.

Account model for networks

Networks typically use implicit accounts (require_operator_auth: false). The buyer agent is trusted and declares brands via sync_accounts. The network then manages its own accounts with each underlying AI platform:

Buyer agent -> Network (implicit accounts, agent-trusted)
Network -> AI Platform A (explicit accounts, operator auth)
Network -> AI Platform B (explicit accounts, operator auth)

Declare capabilities accordingly:

{
  "adcp": { "major_versions": [3] },
  "supported_protocols": ["media_buy", "creative"],
  "account": {
    "require_operator_auth": false,
    "supported_billing": ["operator", "agent"],
    "required_for_products": false,
    "sandbox": true
  }
}

Catalog forwarding

Networks receive catalogs from buyers via sync_catalogs and forward them to the relevant AI platforms. The same task works on both legs — the network acts as a buyer when syncing catalogs to each platform. This is the core data pipe: brand catalog data flows from buyer to network to platform, giving each platform the raw material to generate ads.

Governance and content standards

Networks can enforce governance policies at the routing layer before forwarding to platforms. When a buyer pushes content standards, the network applies them when selecting which platforms and contexts are eligible — then forwards the policies to each platform so they’re also enforced at creative generation time. This gives brands two layers of suitability enforcement: the network’s routing decisions and the platform’s generation constraints.

Delivery reporting

Networks aggregate delivery data from underlying platforms and present unified reporting to buyers via get_media_buy_delivery. The buyer sees a single delivery report per media buy — the network handles the per-platform breakdown internally. Networks that want to offer platform-level transparency can use reporting_dimensions to expose placement-level breakdowns.

adagents.json for networks

A network’s adagents.json lists the publisher properties it represents, even though it doesn’t own them:

{
  "version": "1.0",
  "properties": [
    {
      "domain": "assistant-alpha.example.com",
      "agents": [{
        "agent_url": "https://ads.ai-network.example.com",
        "relationship": "direct",
        "supported_protocols": ["media_buy", "creative"]
      }]
    },
    {
      "domain": "assistant-beta.example.com",
      "agents": [{
        "agent_url": "https://ads.ai-network.example.com",
        "relationship": "direct",
        "supported_protocols": ["media_buy", "creative"]
      }]
    }
  ]
}

Each underlying AI platform authorizes the network in its own adagents.json. Buyer agents discover the network through the platforms’ authorization chains.

SI Chat Protocol through networks

When an ad network sells SI Chat Protocol sessions on behalf of brands, it acts as an intermediary in the session flow. The brand syncs offerings to the network via sync_catalogs with type: "offering", and the network forwards them to underlying platforms. When a platform triggers a session, the network routes it to the correct brand agent.

AI Platform -> Network -> Brand Agent

1. Platform calls si_initiate_session with the network's media_buy_id
2. Network maps media_buy_id to the brand's offering_id
3. Network forwards to the brand agent's SI endpoint
4. Brand agent responds; network relays back to the platform

The key fields at each leg:

Field	Platform to Network	Network to Brand
media_buy_id	Network’s media buy ID	May differ or be omitted
offering_id	Not set (platform doesn’t know)	Brand-specific offering
context	User intent from the conversation	Forwarded as-is
identity	User identity (if consented)	Forwarded as-is

The network handles attribution correlation across the two legs. It knows which platform triggered the session (placement), which media buy funded it (media_buy_id), and which brand responded (offering_id). This lets the network provide unified delivery reporting to buyers via get_media_buy_delivery while each brand agent only sees its own sessions.

Networks should forward identity and supported_capabilities unchanged — the brand agent needs accurate host capabilities to negotiate modalities, and the user’s consent was granted for the brand, not the network.

For the SI Chat Protocol session lifecycle and capability negotiation details, see the SI Chat Protocol and implementing SI hosts.