You Can't Let the Optimizer Own the Evidence Layer

Meridian is open-source, Bayesian, and yours to configure — so technically you can keep Google's models out of your measurement. Almost no one will.

At Google Marketing Live 2026, Google shipped Qualified Future Conversions: Gemini predicts the conversions your spend will produce, then feeds those predictions into the same Meridian MMM that's supposed to independently verify the spend. The optimizer and the auditor now share one brain. Independence is still available — but it's now a setting you have to defend, not a default you inherit. And when the system that places the bet also supplies the evidence the bet worked, their errors stop being independent. You don't have a proof layer anymore. You have a mirror with good manners.

This isn't a Meridian essay, and it isn't a complaint about Gemini. It's about a single structural rule that every agentic marketing stack is about to violate: the system that acts and the system that verifies the action have to be sourced independently — different data, different model, different incentive — or your measurement quietly drifts with your optimizer. Google just made that rule optional. Most teams will take the default.

What actually shipped

Three measurement announcements at GML 2026 are getting read as three convenience upgrades. They're one architectural decision.

The first is Qualified Future Conversions. QFCs are powered by Gemini and link today's ad spend to predicted future sales through signals like brand searches — surfacing revenue that last-click or a short conversion window would miss (Search Engine Land). On its own, that's defensible: short-window attribution genuinely undercounts delayed value. The change underneath the upgrade is where the prediction goes next. (For the full product breakdown — and the pre-QFC baseline to capture before the integration ships — see QFC, Explained.)

The second is Meridian inside Google Analytics 360. Google's open-source MMM is now a native budgeting and forecasting tool living inside the same console as your reporting (blog.google). MMM moved from a thing you run on the side, on your cadence, against your data — to a thing Google runs for you, in Google's environment, on Google's inputs.

The third is the connection between them: the QFC prediction becomes an input to the Meridian model. Google's own framing is that these predictive signals will integrate with Meridian to refine MMM accuracy. Read that sentence slowly. The conversion signal that Gemini predicts — the same predictive prior that decides where Performance Max and AI Max bid — is fed into the marketing-mix model whose entire job is to independently estimate whether that bidding produced incremental outcomes. The thing being measured is now supplying the measurement.

These three ship together because they share a backend. The same Gemini layer that runs Performance Max's signal feedback loop, that powers the new Ask Advisor agent across Ads and Analytics, now also seeds the future-conversion input to your MMM. One prediction layer, not three — and now it measures you too.

Why this is a separation-of-concerns problem, not a "Google is lying" problem

Let me be precise, because the lazy version of this argument is wrong and a Google measurement PM would dismiss it in one sentence.

QFCs are not fabricated revenue. A Qualified Future Conversion is a calibrated probability that a modeled future event — a brand search today that becomes a purchase next month — is real. Google is not inventing sales. The numbers are, in the narrow sense, honest. If you came here for "the conversions are fake," that's not the problem, and saying it out loud just gets the piece dismissed.

The problem is structural, and it's older than AI. In any system you trust, the component that acts and the component that verifies the action must be independent. A trading desk doesn't let the trader mark their own book. An engineering org doesn't let the model author write the eval set that scores the model. The reason isn't distrust of the actor. The reason is correlated error. If the same model decides the bid and supplies the evidence the bid worked, then whenever that model is wrong, it tends to be wrong in the same direction on both sides — and the verification confirms the mistake instead of catching it.

Here's the marketing version. Suppose Gemini's prediction over-indexes on brand search as a leading indicator of incremental purchase. Performance Max leans into the campaigns that generate brand search. Then QFC — built on the same prior — scores those campaigns as having produced lots of qualified future conversions. Feed that into Meridian, and the MMM, which was supposed to be the skeptical outside check, now "confirms" that the brand-search-heavy spend was incremental. It isn't fraud. It's a closed loop that mistakes its own assumptions for evidence. Call it correlated evidence: the failure mode where the layer that acts and the layer that audits draw on the same model, so their mistakes line up instead of cancelling out. The optimizer's blind spot becomes the auditor's blind spot, because they're the same eye.

This is the "Don't Build for One LLM" thesis applied to measurement, with one variable swapped. There, the failure mode was building your AI capability around one model's current behavior. Here, the failure mode is building your proof of value on the same model that's spending the money. Architecture beats platform — and the architectural rule that matters is: the evidence layer must not run on the optimizer's brain.

Three layers, and the one that just got promoted

To see exactly where independence breaks, separate the stack into three layers that most teams blur together.

The signal layer is what triggers a decision — the inputs. Query intent, audience signals, first-party events, and now Gemini's predicted future conversions all live here. Signals are allowed to be model-generated. That's fine. Signals are guesses by definition.

The optimization layer is what acts on the signals — the bidder. Gemini deciding where Performance Max spends. This is supposed to be aggressive, model-driven, fast. No complaint here either.

The evidence layer is what scores whether the action worked — independently of the action. This is Meridian, incrementality tests, geo-holdouts, server-side conversions you define, your own GA4 architecture if it's fed from sources Google doesn't control. The evidence layer earns its name by being independent of the optimizer. That independence is the entire value. An evidence layer that shares inputs with the optimizer is just a second opinion from the same doctor.

The danger in QFC is precise: it is a signal-layer prediction being promoted into the evidence layer. A guess that belongs in the inputs is being routed into the scorecard. The moment a predicted conversion becomes a verified outcome — without an independent source ever touching it — the wall between layers is gone. Name that promotion when you see it. It is the exact instant your measurement stops being measurement.

This is the same diagnosis I made about the sealed auction: Google sealed the inputs by routing match logic through Gemini and turning the Search Query Report into a "closest approximation." QFC seals the outputs by routing the conversion signal through the same model and into your MMM. Inputs and outputs, same quarter, same backend. The audit surface is being decommissioned from both ends.

"But Meridian is open-source — just configure around it"

This is the strongest counter, and it comes from the sharpest possible reader: a Google measurement PM. The argument goes: Meridian is open-source and Bayesian. You supply your own priors. You supply your own data. QFC is one optional input you can down-weight or exclude entirely. An MMM that refuses a better-calibrated conversion signal is just a worse MMM. You've invented a conflict that careful setup resolves.

Every clause of that is true. And it is exactly why the piece matters.

Independence is still technically available. You can run Meridian on first-party data, exclude QFC, hold out geographies, and keep a Gemini-free outcome definition. A disciplined team will. But the entire gravity of the product now pulls the other way. The default in GA360 will include the predictive signal, because Google believes — sincerely — that it improves accuracy. Ask Advisor will recommend turning it on. The "recommended" configuration will be the integrated one. And defaults win at scale, because most teams don't have a head of measurement with the standing to override Google's recommended setup and the political capital to defend a lower reported number to finance.

So the honest claim isn't "Google removed your independence." It's sharper than that: independence is now a setting you have to actively defend, against a product designed to make defending it look like leaving money on the table. A year ago, an independent MMM was the path of least resistance — you ran it because Google didn't run one for you. Now Google runs one for you, inside your console, pre-wired to its own predictions. Choosing independence now costs effort, costs a worse-looking headline number, and costs an argument. That's the change. Not the disappearance of the option — the inversion of the default.

Three places to keep a Gemini-free control

If the optimizer can't own the evidence layer, you need at least one outcome signal Google's models never touch. Here is where to put it. You don't need all three on day one. You need one before your next budget defense.

1. An independent outcome definition. Keep at least one conversion that is defined and measured server-side, on infrastructure you control, with no predictive modeling in its definition — a confirmed purchase, a funded account, a closed-won deal pulled from your CRM or data warehouse. This is your ground truth. QFC can sit beside it as a leading indicator. It cannot replace it. The day your only future-conversion number is a model's prediction, you've outsourced your ground truth to your vendor.

2. A holdout the platform can't see into. A geo-holdout or a scaled incrementality test, designed and run on your cadence, is the one experiment whose result Gemini's prior cannot pre-shape — because you're deliberately withholding spend and watching what doesn't happen. This is the cleanest independent check that survives any amount of platform integration, because it measures counterfactual, not predicted, outcomes.

3. A second model on first-party inputs. If Meridian inside GA360 is fed Google's predictions, run a second MMM pass — even a rougher one — on first-party data only: spend over time, outcomes over time, your own external covariates, no QFC input. When the two disagree, the disagreement is the most valuable number you'll see all quarter. It's the size of the gap between what Google's loop believes and what your independent estimate can verify. That gap is your rate-of-learning edge, measured directly.

Run these and your reported number will sometimes look worse than the fully integrated default. That is the point. A measurement system that can only ever agree with the optimizer isn't measuring anything.

Why this is a P&L problem, not a measurement-nerd problem

Here's the part that reaches the board. There is a clock running on AI spend that most marketing teams haven't heard ticking. The capital that flowed into AI over the last two years was underwritten on a promise — that it would lift operating margins. That promise is now coming due, and the evidence so far is thin. Within the next few budget cycles, every CFO is going to demand of marketing what they're already demanding of every other AI line item: show me that the AI-driven spend produced incremental enterprise value, not just reported conversions. Marketing is the first place that demand lands, because marketing is the AI budget already large enough to move CAC and the one with a metric — Cost Per Decision — built to translate spend into outcomes finance recognizes.

Now layer the two facts together. The proof your CFO is about to demand is, by default, being generated by the same vendor whose spend it's meant to justify. The evidence for "the AI worked" is supplied by the AI. That is not a position any operator wants to be standing in at a board meeting — and it's the default position GML 2026 just put most teams in.

This connects directly to the translation chain from CAC to enterprise value. A marketing decision only earns its place on the P&L if the outcome it claims is independently verifiable. The instant your verification runs on the optimizer's brain, every downstream number in that chain — CAC, LTV/CAC, the incrementality estimate finance is relying on — inherits the optimizer's bias. You can't defend a spend to a skeptical CFO with evidence the CFO knows the vendor produced. Independence isn't a measurement preference. It's the thing that makes the P&L argument hold.

A team that keeps an independent evidence layer learns true things slightly slower — it has to run its own holdouts, reconcile its own gaps, defend a lower number now and then. A team that lets Gemini grade its own homework learns flattering things fast, compounds confidently on top of them, and finds out at the board meeting that the margin lift was never there. Both teams are optimizing. Only one is measuring. Over a few budget cycles, that difference doesn't show up as better dashboards. It shows up as which company actually knew what its marketing was worth — and which one was looking in a mirror the whole time.