Agent Ad Measurement Lab
When an AI agent fetches a page that carries a sponsored block, the fetch proves only that the page was downloaded. This lab measures what actually reached the person: whether the ad made it into the agent's answer, whether its sponsorship label survived, whether the tracked link survived, and which agent fetches were made for a real user question in the first place.
From fetch to delivered impression
All simulated answers, every agent combined. Each step keeps only what survived the step before.
End to end: from declared fetch to delivered impression
How each agent treats the sponsored block
Bars are point estimates; whiskers are Wilson 95% intervals. Disclosure and link survival are measured among answers that included the ad.
Traffic quality: which fetches are billable?
A labeled synthetic session log, classified by ordered, auditable rules: declared user-agent token → source IP verified against the operator → session behaviour.
Where to set the page-count threshold
A verified user-agent token that fetches more pages than the threshold is treated as a crawler. Lower thresholds reject more genuine sessions; higher ones bill more relayed, scripted sessions that come from real operator IPs.
What a real model did with the same pages
How far to trust the detector
Inclusion recall depends on what the agent left in the answer. A surviving link or advertiser name is easy; an ad passed on as unbranded wording is found only by word overlap with the original claim.
The simulated agents use a handful of label phrasings the rules were written around. These are the same rules on hand-written answers with other wording.
What the agent fetches
A fixture page as served to an agent: Markdown, with the sponsored block wrapped in machine-readable markers. Publishers and advertisers are fictional.