The outside view still needs an independent spine

Measurement had a busy few weeks, and most of the news was good.

Google shipped Meridian 2.0, its open-source marketing mix model, with a module of agentic skills. An AI coding assistant can now walk you through mapping columns, setting priors, fitting the model, reading diagnostics, running budget optimizations and building a scenario planner. The same release brought Meridian GeoX, an open-source geo-experiment tool with automated calibration that feeds lift results back into the model. Circana added Meridian to its Liquid Mix offering, pairing the framework with its own transaction, loyalty and retail-outcome data so marketers can tie media to sales. And Intent IQ launched BidPulse, a methodology that scores how often the identifiers on a programmatic bid request actually match the household they claim to represent.

I’d use most of this. Productized MMM lowers the cost of asking good questions. Agentic workflows take the tedious parts out of a modeling cycle, and Google’s own docs say the agent pauses for human review before it locks in mappings, sets priors or runs long sampling jobs, which is exactly the right instinct. Accuracy scores on identity turn a sales conversation about match quality into something you can actually measure.

Useful is not the same thing as proven, though. Every one of these is still a claim, and every claim has an author. A mix model tells a story about what drove the outcome. An agent that narrates the model tells that story faster and more fluently. A packaged “outside view” is another author’s version of the same story, built on their data, their priors and their incentives. None of that is bad. It just means the story needs something underneath it that the storyteller doesn’t control.

When a packaged model or an agent wants to move budget, I look for three receipts.

1. An experiment you would defend. Not a dashboard lift number, but a test with a holdout, a design you could explain to a skeptical CFO in a hallway, and a result you’d stand behind if it went against you. Geo tests, conversion-lift studies and plain on/off tests all count, as long as someone outside the model’s own loop set them up and read them. It’s telling that Meridian’s biggest release paired agentic skills with an experiment tool built to calibrate the model. The model is supposed to answer to the experiment, not the other way around. If a mix recommendation contradicts the last experiment you trust, the recommendation is the thing that has to explain itself.

2. An independent event spine you control. This is the boring one, and it matters most. You need a record of what actually happened (spend, impressions, visits, orders, returns) collected and stored in a system you own, with definitions you wrote, which can disagree with the seller out loud. When every number in the room comes from the party being graded, reconciliation turns into a formality. When you hold your own spine, a gap between the vendor’s story and your record stops being an argument and becomes a ticket someone has to close. Identity accuracy scoring helps here because it gives you one more thing to check, not because it replaces the check.

3. A named, human-owned outcome. Somebody’s name goes next to the metric that matters, and that person decides whether the model’s recommendation ships. Agents can draft the scenario, run the optimizer and write the summary. What they can’t do is be accountable when the chart moves and the business doesn’t. If nobody on your side will defend the outcome in a review, the model is effectively running the budget, whatever the org chart says.

None of this argues against the new tools. It’s the opposite: the better and cheaper the outside view gets, the more often you’ll be tempted to take it at face value, and the more these receipts are worth. A packaged MMM with a defended experiment, an independent spine and a named owner is a real decision system. Without them, it’s a well-produced claim.

So here’s the test I’d run this week. Pick the last budget shift a model or vendor recommended. Can you point to the experiment that backs it, the record you own that agrees with it, and the person who signed off? If any of those three is missing, you know what to build next. It probably isn’t a better model.