Sponsored Turns: Ad Units Built for Assistant Surfaces

A conversational unit has one hard constraint and it is not creative. It has to resolve inside the token stream. We design to a 400 millisecond decision-and-render budget, because beyond roughly that the assistant has finished its sentence and the unit lands as an interruption rather than as part of the answer. Format, disclosure wording and the handoff all follow from that budget, and any of them can be vetoed by it. The number we design against: Click-through on a unit inside a conversation tells you very little. A user who clicks has left the assistant, which is the one thing the publisher cannot afford at scale. We hold every unit to turn continuation: does the user take another turn after seeing it. A unit at 4 percent click that ends the session has cost the publisher more than it earned, and we will say that in the readout even when the click number is the one the advertiser wanted to hear. Formats we build: An inline mention with a source chip, where the commercial element is a named entity inside the model's own sentence and the disclosure sits on the chip. An appended card after the answer completes, which is the safest and the one most publishers should start with. A comparison block the user can query further without leaving the turn. And a handoff, where the assistant offers to continue the task inside the advertiser's own tool — the highest-intent format and the one that fails hardest when the tool behind it is slow or demands a login. Each format ships as a specification rather than a picture: states for loading, timeout, no-fill and refusal; behaviour at three viewport widths and in a voice-only surface where there is no card at all; the payload fields the decisioning layer must return at request time; and a QA script your engineers run against a staging build. Disclosure: The label goes before the commercial content, in the same type size as the body, and it names the relationship rather than gesturing at it. We write the wording against the FTC's guidance on clear and conspicuous disclosure and the identification requirements in Article 26 of the Digital Services Act where the surface reaches the EU, then we test it: 150 to 250 users, unaided, asked what they just saw. Anything under 80 percent correct identification is rewritten and retested. That test is the part clients try to cut and the part we will not run the project without. What we refuse: We do not write units in the assistant's voice. If the model appears to recommend the advertiser in its own register, the unit is deceptive whatever the label says, and it is the fastest way to lose the surface entirely once the platform's policy team catches up. We do not design units that trigger on symptom or dosage questions in health surfaces. And we do not assemble health claims dynamically: copy reaching a wellness audience is fixed, reviewed before render, and versioned. Not included: We do not build the ad server, the bidder or the decisioning logic. We specify what they must return and inside what budget. We do not run media under this scope. We do not design for assistants whose audience is meaningfully under thirteen. Who this is not for: Advertisers who want the unit to be indistinguishable from the model's own answer. Publishers who will not expose a fallback slot, because a unit with no defined no-fill state degrades into a blank gap mid-sentence. And anyone who needs this inside a fortnight: the comprehension testing alone takes ten working days, and skipping it removes the reason to hire us.