What is ad revenue?

What drives it

Ad revenue is the product of how much inventory exists, how much of it sells, and what it sells for. On the web those three are tracked as impressions, fill rate and eCPM, and a publisher can move all three.

The first lever is the one that does not transfer to AI surfaces. A web publisher can create more inventory — more pages, more placements per page, refreshed slots. An AI publisher cannot: the inventory is created by users asking questions, and adding more advertising per conversation degrades the product rather than expanding supply.

What a publisher actually controls

• Eligibility. Which conversations may carry advertising — the decision that most affects both revenue and product quality.
• Floors and demand mix. What the inventory is worth and who is bidding on it.
• Format. How natively a placement renders, which drives engagement and therefore price.
• Frequency. Capping protects sessions, and protected sessions produce more monetisable moments — see conversational frequency cap.

How AI publishers measure it

Per-impression metrics describe AI inventory badly, because most conversations carry no advertising at all and impressions are therefore a small and variable share of activity.

The metric that fits is per query: total ad revenue divided by total queries, whether or not they were monetised. It captures both how well the sellable inventory monetises and how much of the total is sellable, which is the number a product decision actually moves. Revenue per query and RPM benchmarks cover where these currently sit.

The restraint argument

On a web page, revenue and user experience trade off gradually — more ads means somewhat more revenue and somewhat worse experience, and publishers tune the balance.

On an AI surface the curve is sharper. Advertising that appears to compromise answers does not reduce satisfaction slightly; it removes the reason the product is used. Publishers running low ad density and strict eligibility consistently report better long-run revenue than those optimising per-conversation yield, because the base keeps growing.