Five pins before an agentic measurement tool writes

Measurement is getting agents. Agents are also showing up in the traffic being measured. Both happened in the same two weeks.

What shipped (only what the sources say):

• Lifesight (blog). An agentic unified marketing measurement platform: a causal engine plus agents across MMM, geo-incrementality and incrementality-adjusted attribution. A human checks the evidence and decides.
• Snowflake AI Gateway for Advertising (preview). Cortex Agents connect to Meta and TikTok via MCP and join campaign data with inventory and transactions. They recommend budget actions, campaign managers approve, and there’s a shared audit record. Their example: ROAS fell while CTR looked fine, and the agent traced it to checkout payload, catalogue and stock problems.
• Contentsquare Experience Platform (beta, CX World). It separates LLM and AI-agent traffic from human shoppers, attributes AI-referred journeys, and tracks WebMCP.
• ChatGPT Ads. Visual formats in test, and measurement partners including Hightouch, LiveRamp and Tealium. WeightWatchers reported a 15.3% lower attributed CPA than its paid-search blend. That number is partner-attributed, not independent. Agencies still cite attribution gaps as a cap on spend.

The direction is right. The approval gates are in the launch copy. What I care about is what has to be pinned before any of it gets write access to budget.

Five pins:

1. Split human vs agent sessions in the spine. Every session row gets an actor: human, agent, or agent-for-human. Not a filter on a dashboard. A column in the record. If it’s missing, every conversion rate downstream is a blend of two populations.
2. Own the events. Spend, sessions, orders, returns and stock land in a store you control, with your definitions. Network-reported numbers come in as inputs to reconcile, not as truth. The Snowflake diagnosis only worked because the ad data sat next to inventory and transactions.
3. Causal check, not only MTA. Multi-touch attribution splits credit. It doesn’t prove lift. Any recommendation over a set threshold needs an incrementality read behind it (geo test, holdout, MMM calibrated to one) before it ships.
4. Human approves, audit row written. The agent proposes and a named person approves. The row records what changed, the evidence cited, who approved, when, and the expected effect. It’s append-only and the agent can’t edit it.
5. Don’t let the optimizer grade itself. The agent that recommends the move doesn’t get to score whether it worked. A separate check, ideally on a held-out metric the optimizer never sees, reads the result.

Tells it’s not ready:

• Agent traffic shows up as a channel row instead of an actor field.
• The only conversion source is the network’s pixel.
• “Lift” means MTA credit moved.
• Approval is a toggle with no name attached.
• The weekly report on the agent’s results is written by the agent.

How it maps to what we run: Atrilyx’s AI Agent at Ai Media Group runs on reconciled attribution, not network-reported numbers. It watches pacing and conversion anomalies and recommends, the team reviews, and it makes no autonomous budget changes. That’s pins 2 and 4 by design. Pin 1 is the one I’d audit first on any stack as agent traffic grows.

Let the agent read everything. Pin what it can write.