What is an AI DSP?

What an AI DSP does differently

A conventional DSP answers one question very fast: given this impression and what I know about this user, what should I bid? The inputs are an identifier, a page context and an audience segment.

An AI DSP answers a different question. There is no durable identifier, and the page context is a live conversation. What it has instead is the user's stated intent, in their own words, at the moment they stated it. Bidding is therefore against the strength and specificity of that intent rather than against the value of a known individual.

In practice that inverts the optimisation problem. A conventional DSP finds the right person and accepts whatever context they are in. An AI DSP finds the right moment and accepts whoever is in it.

What it optimises on

• Intent strength. How close the conversation is to a decision — see intent density.
• Prompt relevance. How well the advertiser's offer answers what was actually asked, not what category it belongs to.
• Conversational position. A recommendation early in exploration and one at the point of choosing behave very differently.
• Eligibility and safety. Whether the conversation is one an advertiser should appear in at all, resolved before bidding rather than after.
• Frequency across the conversation, not across pageviews — see conversational frequency cap.

The latency constraint

A display auction runs while the page is loading and has roughly 100ms. An AI DSP has to return a decision inside the window in which the assistant is generating its response — the user is already watching text appear, and a late answer is simply not rendered.

That is a harder engineering constraint than display, and it shapes the architecture. Evaluation has to be cheap enough to run inline, which favours pre-computed eligibility and compact context representations over heavyweight per-request modelling. Our ad infrastructure page covers how the decision path is arranged.

Do you need an AI DSP?

For most advertisers today, no — an AI ad network or a managed buy reaches the same inventory with far less operating overhead. A DSP earns its keep when you are running enough volume that bid-level control changes the outcome, and when you have the team to use it. Our self-serve and managed service options set out where the line usually falls.