Same Ads, Different Room

Thu, 15 Oct 2026 08:30:00 +0200

TL;DR: Ads in AI answers work much like search ads in their publicly documented buying mechanics, with three changes. The system reads the whole conversation to pick an ad, where search starts from a short query, and in the format I model, the ad sits next to one synthesized answer instead of ten links. No publisher stands between the platform and the user. The buying side is familiar: auctions, cost-per-click and outcome bidding, pixels, attribution partners. I built a small simulation to see whether this creates new ad money or only moves it. Publishers lose in every scenario I tried, because the answer sends fewer people to their pages. In my model, total ad spend only grows if the answer engine converts about 1.45 to 1.67 times better than search. Up to that point, the better conversion only makes up for showing one ad where search showed several. The numbers are my assumptions, and the code is on GitHub so you can change them.

The earlier posts in this series, written over the past two months, were about ads the model reads: a publisher serving sponsored text to crawlers [9], an experiment showing that text can tilt an agent’s answer [10], the ad-free tier that follows from that [11], and the question of who pays for the agentic web [12]. This post is about ads people see in AI products, which just became a big business.

schristoph.online

Same Ads, Different Room