Attribution was built on a simple idea: a person sees an ad, clicks, browses and buys, and each step leaves a trace. Agentic commerce changes who takes those steps. An AI agent reads product pages and feeds, compares prices and policies, and increasingly places the order on someone's behalf. The person may never see your ad or your site.
What breaks
- Last-click and multi-touch models assume a human path. Agent sessions look like bots, arrive with no referrer, or are filtered out entirely.
- Platform-reported conversions over-credit what they can see and miss what happened inside the assistant.
- Identity gets murky: who delegated the task, with what instructions, and was a person involved at the end?
Measuring agent-mediated demand
Treat agents as a channel of their own. Separate known AI agents and assistant referrals from generic bots in your analytics and server logs, and tag the conversions they touch. Make product data, pricing and policies machine-readable so agents can act on them, and log those reads. Where the path cannot be observed, use experiments: holdouts and staged rollouts show what changed when you improved agent readiness, without needing to see every step.
Label every number
The most useful habit is labeling each figure by how it was produced: observed, sampled, experimental or modeled. Leaders can then weigh a server-side fact differently from a model estimate, instead of averaging them into a single, misleading total.