Ad units served inside AI-generated search answers rather than beside a list of blue links, which means the context around your ad is written fresh by a model every time.
Also known as AI search ads, in-answer ads, Highlighted Answers, conversational ad formats
Conversational discovery ads are paid placements that appear inside a generated answer rather than around it. Google moved the format, along with Highlighted Answers, into U.S. mobile and desktop testing in 2026. The user asks something in natural language, the model composes a response, and the ad is threaded into that response at the point where a product or service is relevant.
The mechanical difference from search advertising is that there is no stable page to buy against. A keyword auction assumes a query maps to a results page you can model. Here the query maps to a generated passage whose length, tone and framing change from user to user, so the same ad can land inside a cautious answer, an enthusiastic one, or a comparison that names two competitors first.
Clients are being offered a new inventory type with no case studies, no benchmarks, and no internal expert, and they are asking their agency what to do about it.
You are writing for a slot with unknown neighbours. Copy that reads well in isolation can read badly directly beneath a model-generated caution about your category. The craft question becomes tonal compatibility with text you did not write and cannot see in advance, which is closer to product placement than to search copywriting.
The old measurement frame does not transfer. Impression, position and CTR assume a page layout. In a generated answer, position is a function of how the model chose to structure that particular response. Reporting that treats it like paid search will produce numbers that look fine and mean nothing.
Brand safety moves from placement to generation. Traditional brand safety asks what page the ad sits on. Here it has to ask what the model just said. An ad for a financial product landing inside an answer about debt distress is a different risk profile than the same ad on a finance article, and the existing exclusion lists do not describe it.
A mid-size agency running paid search for a regional insurance brand gets early access to the format and treats it like a standard expansion: same ad copy, same landing pages, same weekly pacing report. Three weeks in, the CTR looks strong and the sales team is unhappy, because a meaningful share of the clicks arrive from answers where the model has just explained why the user might not need coverage at all. The response is not a bid adjustment. The team builds a small library of variant copy keyed to answer intent, splitting reassurance-shaped answers from comparison-shaped ones, and adds a manual weekly review where a copywriter reads twenty live answers and flags contexts to exclude. That review has no line in the retainer, which becomes the actual conversation with the client.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.