The Metric That Was Never About You

The cold open

Here’s the poll exactly as it ran:

Most marketing metrics were designed for human decision-making. What happens when software starts making more of the decisions than people? Which metric breaks first?

Options: Attention. Reach & Frequency. Attribution. ROAS. Core thesis at launch: human measurement frameworks may not survive agentic decisioning.

Read fast, this sounds like a question about precision — which of these four numbers gets noisier once a machine is doing the choosing. That reading is wrong, and it’s wrong in a specific way worth naming before the vote. All four of these are not measurements. They are inferences — statistical stand-ins for something that has never been directly observable: what happened inside somebody else’s head. The question isn’t which inference gets less accurate. It’s which inference has nothing left to infer once the head in question isn’t a head at all.

The vote

The poll closed with 43 votes. (Shares are LinkedIn’s final rounding, which is why they sum to 101.)

The plurality landed on Attribution, with Attention and Reach & Frequency close behind it, forming what was effectively a three-way cluster of “this breaks” — and ROAS sitting off by itself, undisturbed, in “this survives” territory. The order among the three shuffled all week: Attention led the early count, Attribution overtook it, and at the wire Attention closed the gap — 33 against 35 — but Attribution held. That instability is itself a small piece of evidence for the argument below: the room was confident these three are fragile and kept re-litigating only the ranking among them, while ROAS never moved at all.

The reflex read is simple: Attention is the softest, most vibes-based number on the sheet, so of course it dies first; ROAS is cold arithmetic, so naturally it’s safe. That’s directionally right, and it’s also incomplete, because it treats Attention, Reach & Frequency, and Attribution as three flavors of the same softness instead of asking what each one was actually built to do — and it treats ROAS’s survival as a merit badge, as if ROAS earned its outlier-low share through superior rigor. It didn’t earn anything. It was just never fragile to begin with.

The reframe

Here’s the actual mechanism. Attention, Reach & Frequency, and Attribution were never measurements of the market. They were techniques for inferring an event nobody could directly observe: what happened inside another mind.

Reach & Frequency infers whether a message registered and stuck, because you cannot watch memory encode. Count the exposures, assume some threshold produces recall — “effective frequency” is a number invented to answer a question you have no direct instrument for.

Attribution infers a causal story linking a touchpoint three weeks ago to a purchase today, because the actual decision process was never observable even in principle. Attribution modeling doesn’t recover what happened in someone’s head. It manufactures a plausible narrative and assigns credit along it. The story is the product; the causality underneath was always a guess wearing a model’s clothing.

Attention infers whether registration happened at all — whether a message crossed the threshold into awareness. It is the most direct of the three inferences, because it isn’t trying to reconstruct a memory or a journey, only a single binary: did this get noticed. It became the leading indicator for everything downstream precisely because it was the last rung before the unobservable part started.

Now put an agent in the loop instead of a person. It isn’t degradation. It’s the disappearance of the thing being inferred. Whatever an agent weighted, retrieved, or ranked on its way to a decision isn’t hidden behind a skull — it can be logged, replayed, and read directly. There is no encoding event to infer, because there’s no encoding event to protect. There is no hidden journey to reconstruct, because the decision path was written down as it happened. There is no gap between noticing and acting for attention to occupy, because for software, evaluating an input and acting on it are the same instruction. Three inference techniques, and in each case the target of the inference simply stops existing.

The Instrumentation Test

The Instrumentation Test: a metric survives agentic decisioning only if the event it measures remains unobservable in the system now making the decision — inference has no work left to do once you can just read the record.

Walk it through mechanically.

Reach & Frequency fails. There is no unobserved encoding event in an agent to infer toward. The number doesn’t go wrong; it goes inapplicable — a unit for a phenomenon that isn’t occurring.

Attribution fails. There is no hidden causal chain to reconstruct, because the agent’s decision path is already a matter of record. The entire inferential exercise attribution exists to perform has nothing left to infer.

Attention fails hardest and fastest. It was the most direct inference of the three — a stand-in for a single unobservable event, registration — and once there’s no boundary between perceiving and acting, that event has no referent to point at. Not degraded. Void.

ROAS passes. Not because it’s clever, but because it was never an inference about anyone’s mental state to begin with. It’s a ratio of dollars returned to dollars spent, and it doesn’t care whether a human or an agent did the deciding in between. There was never a hidden event on the other side of it, so instrumenting the decision-maker changes nothing about what it measures.

Which lands the real point: ROAS didn’t win this argument. It won by default. It’s the only one of the four that was never trying to infer a mind, so it’s the only one with nothing to lose when the mind left the loop.

Why Attention specifically

Reach & Frequency and Attribution are each one inferential layer further removed from the vanished event than Attention is. Frequency isn’t memory — it’s a proxy for a memory-encoding event. Attribution isn’t the journey — it’s a proxy for a story about a journey. Because they’re proxies for a proxy, they degrade rather than vanish outright: a frequency count still counts something, an attribution model still assigns credit to something, even after the meaning underneath has left. They retain machine-legible residue.

Attention has none. It isn’t a proxy for the registration event — it names the registration event directly, with nothing standing between the metric and the thing it’s clocking. So it doesn’t degrade gracefully the way the other two do. Fewest inferential steps between the metric and the vanished target: zero. That’s the mechanical reason Attention is the cleanest kill of the three, however the vote finally sorted them — not “softest.” Least insulated.

The record that replaces the guess

Name the old dynamic honestly first: prior weeks in this series already established that an accountability demand doesn’t vanish when its object disappears — it relocates. This week’s poll was another instance of that pattern, not a new discovery, and it’s worth saying so plainly rather than dressing it up as something else.

What’s actually new is what the relocated demand can now ask for. The old proof was necessarily circumstantial, because the thing being audited — a human’s private reasoning — was never accessible from outside. Frequency curves, attribution models, attention scores: all of it was the best available evidence about a process nobody could open up and read directly. Marketing measurement, underneath its dashboards, has always been a discipline of educated guessing.

An agent’s reasoning isn’t private in the same way. It doesn’t have to be inferred from downstream traces, because it can be required to produce its own record: what it knew at the moment of decision, what alternatives it weighed, what outcome it predicted against the outcome it actually got.

The Delegate Audit Record: a structured trace of why an agent made a specific spend decision — its constraints, the alternatives it considered, and its predicted outcome measured against the actual one.

This is not the Dissent-Capacity apparatus from a few weeks back wearing a new label. That mechanism asked whether a human could still refuse a recommendation before it executed — an ex-ante veto with a clock on it. The Delegate Audit Record has no veto and no clock. It’s assembled after the decision has already run, and its job is retrospective judgment, not pre-emptive refusal. One tries to stop a bad call before it happens. The other explains a call after it already did.

It’s also a different kind of artifact than any of the four legacy metrics, not just a fifth option among them. Attention, Reach & Frequency, Attribution, and ROAS are all outcome-side: was it seen, was it remembered, was it credited, did it pay back. The Delegate Audit Record is process-side: given exactly what the agent knew, was the decision itself sound, independent of how the outcome happened to land. An agent can reason well and lose to an unpredictable market, or reason badly and win by luck — ROAS can’t tell those two apart, because it only reads the ratio at the end. Separating decision quality from decision luck has never had a standard instrument, because human decision-makers were never expected to produce a clean trace of their own reasoning on demand. An agent can be.

The objections

The comments did real work this week, and the sharpest pushback came from someone who builds these agents for a living. Kyle Dozeman’s read: agents will make ROAS and Reach & Frequency better — faster aggregation, fewer errors — and push Attribution toward more consistent modeling, even though “the core model is still built on human assumptions.” His verdict: he’s “not sure it breaks any of the frameworks.”

That’s the reflex read from “The vote,” restated by someone who’d know better than most — which makes it worth walking through explicitly. Faster and more consistent is an execution claim: a machine can run the calculation better than a person could. The Instrumentation Test asks a different question: does the number still refer to anything. An agent can compute a frequency count with perfect fidelity forever, and the count still measures nothing, because there’s no encoding event on the other end of it to count toward. Execution quality and referent validity are independent axes — conflating them is the exact reflex this essay opened by naming.

Except his own sentence on Attention gives the argument away: agents “aren’t good at questioning whether the proxy means anything.” Read literally, that’s not a limitation. It’s a diagnosis. There’s nothing left to question, because there’s no longer a proxy for anything. Arrived at independently, by someone arguing against it.

The thread surfaced a sharper, adjacent risk too: if an agent’s model can be tuned, what stops it from being tuned toward outcomes that flatter whoever’s tuning it? That’s not the referent disappearing — it’s a new gaming risk that arrives precisely because the modeler is now a party with something to optimize for. An unobservable event nobody can check is one failure mode. An observable one that someone has an incentive to misreport is a different one, and the Delegate Audit Record has to survive both, not just the first.

One write-in deserves a direct answer. Craig Tuck named Transparency — off the ballot, correctly, because it isn’t a fifth metric competing with the other four. It’s the precondition the Delegate Audit Record depends on to mean anything. A record nobody can inspect is just a fifth guess wearing better production values.

And Luis Segovia landed close to this essay’s own ending without borrowing its language: agents “aren’t going to care about catchy headlines… that changes the game from winning attention to being the most relevant, and machine-readable.” Relevance is the right instinct, and it’s still an outcome-side answer — was the right thing surfaced. The Delegate Audit Record goes one layer under that: not just whether the right choice got made, but why.

Five Fridays, one reservoir

Five weeks in, the throughline holds: every one of these essays lands on the same reservoir — trust, accountability, who actually answers for the outcome. This week doesn’t sit above the others or explain them away; it just finds a narrower, more mechanical piece of the same territory. A few weeks ago this series was about whether a human could still refuse a call before it executed. Then it was about which inputs an agent could manufacture for itself. Neither of those addressed the instruments directly. This one does: the actual tools built to make accountability legible — the dashboards, the frequency curves, the attribution models — were themselves built around one assumption, that the thing being measured was hidden and had to be guessed at. Agentic decisioning is what breaks that assumption. Not trust. Not scarcity. The educated guess.

My vote, and the corrected reason

My own vote was Attention, and the reason was arithmetic, not conviction. Of the three metrics that break, it had the fewest inferential steps to the vanished target — zero, not one, not two. It doesn’t stand for the registration event. It names it directly. That’s why its break is total and immediate rather than gradual — a mechanical reason, independent of which of the three the vote happened to rank first on any given day.

Next Friday

If the audit target moves from the consumer’s cognition to the agent’s decisioning, the Delegate Audit Record itself becomes the thing that has to be trusted — which means it needs to be produced honestly, read correctly, and not gamed by the very agent whose judgment it’s supposed to be checking. So the question that follows: who audits the audit. What does an agent-decision record actually have to contain, structurally, to satisfy a human who still holds the budget and still holds the liability — and what stops the auditor itself from becoming the next laundered layer of accountability.

Let’s see how this plays out. My 2c, as always — food for thought for the weekend.

(It played out the Friday after: Who Owns the Agent’s Decision? — half the room put one name on the outcome, and the answer ran through exactly this machinery: versioned operating rules plus an append-only decision ledger.)