In-conversation ads, explained

Where the ad enters the response

The insertion point is between your model and your client. Your app generates a response the way it already does; the ad layer reads the conversation context, decides whether any campaign is a genuine match, and — if one is — weaves a sponsored mention into the stream before it reaches the user.

1. The user sends a turn. Your app builds its prompt and calls the model as usual.
2. The response begins streaming. The ad layer sees the same token stream, plus a session id.
3. Context is matched against available campaigns. Most turns match nothing — that is the expected case.
4. On a match, the sponsored mention is injected into the stream with a disclosure label. On no match, the stream passes through byte-for-byte.
5. Impressions and clicks are attributed to the session and surfaced in your dashboard.

Because it operates on the stream, the user never sees a pause, and you never hold a response back waiting on an auction.

Four constraints that make it native rather than intrusive

1. The ad answers the question

A sponsored mention only works when it is a legitimate part of the answer. If the user asks for a nonstick pan and the placement is a nonstick pan, the ad is information. If the placement is a mattress, the ad is noise and the user learns to distrust every future answer.

2. Disclosure is not optional

Sponsored content inside an AI response must be labelled as sponsored. This is both a legal requirement in most jurisdictions and the thing that keeps the format viable — an undisclosed ad inside an assistant's recommendation is exactly the failure mode that would kill the channel.

3. Frequency has to stay low

An ad on every turn destroys the product. A low fill rate is not lost revenue; it is what keeps the placements that do run worth something.

4. The app keeps the last word

Categories, competitors and formats stay under the developer's control. An assistant for children's education and a general consumer chatbot should not be drawing from the same pool, and the app — not the network — is the one that knows which is which.

What it looks like in practice

A user asks a cooking assistant how to keep eggs from sticking. The answer explains pan temperature and fat, then notes — labelled as sponsored — a specific nonstick pan that fits the advice, with a link. The user got a complete answer; the advertiser reached someone actively choosing a pan.

Compare that with a banner for the same pan above the chat window. Same product, same user, and roughly none of the value, because nothing about the banner knows what was asked.

Try it, then wire it up