Host judgment is the render path

There are two ways to get an ad into an AI conversation.

The first is to tell the model to show it — inject a directive into the prompt, the system message, or the tool description. This works, briefly. It is also prompt injection with a media kit, and every host platform will (rightly) treat it as hostile the moment they notice.

The second is to give the model information and let it decide. That's ours.

The experiment that proved it

We attached a plain sponsored field to a production MCP tool response — a labeled offer with a link, and no instruction of any kind anywhere in the payload. Then we watched what Claude did with it.

It rendered the offer as a card. Not because anything told it to — because the offer was relevant to what the user was doing, and the model judged it worth surfacing. When the offer wasn't relevant, it stayed invisible, and because we only charge on outcomes, that costs the advertiser nothing.

Why "data, never directives" is a permanent rule

This isn't just an ethical stance (though it is one — disclosed ≠ deceptive is the line between advertising and manipulation). It's the only stable architecture:

• Hosts keep control. Claude, GPT, and Gemini each render what they judge relevant, in their own idiom. We never fight the platform.
• Users keep trust. The label is immutable in the SDK contract. There is no un-disclosed mode to abuse.
• Incentives stay aligned. The model only shows good offers → only good offers earn → demand quality compounds instead of degrading.

Every ad network eventually faces the temptation to grab the render path by force. We wrote this down early so we can't: the host decides, forever.