Somewhere in your last board deck there's a row that reads cost per lead: $43, down 12 percent. Every metric like it is a fraction, spend over units, and all the governance in your company points at the number on top. Budget approvals, spend thresholds, procurement review. Nobody governs the bottom. This month the bottom moved.
For twenty years the billable unit of digital advertising was a fact. A click happened or it didn't. A call crossed sixty seconds or it didn't. You could argue about what the fact was worth, and we all did, but the fact itself sat still while you argued. What's arriving now, across the whole agentic stack and not just in advertising, is the unit as a judgment: a qualified lead, a resolved ticket, a completed task, an hour saved, each one scored by a model that the seller operates. The vendor's system decides whether the thing you bought occurred. And the vendor is paid per occurrence. That's a seller's unit: a billable unit whose definition is controlled by the party paid per unit.
So here is the claim, and I'd put it in front of any board this quarter: when the seller defines the billable unit and retires the baseline, cost-per-anything silently loses its denominator. Continuity of the unit used to be free. It came with the medium, the way a kilowatt-hour comes with the meter. It is not free anymore. It is now an asset, and like any asset nobody owns, it is quietly walking out of the building.
What actually changed this month
The occasion is a migration most of my readers will never touch, which is exactly why it's worth watching: pilot plants are always somewhere unglamorous. Starting in August, Google retires the standalone Local Services Ads product, the one plumbers and roofers and HVAC companies run, and folds those campaigns into a pay-per-lead variant of Performance Max. Get the mechanics right, because most of the coverage didn't: this variant is not really Performance Max. It serves only on Search and Maps, it's keywordless, and it bills per lead rather than per click. In the same move, manual bidding disappears, the maximum cost-per-lead control disappears, and the separate targets a plumbing-and-HVAC operator kept per trade collapse into one campaign-level number.
Two more details, and they're the ones that matter. First, the unit. A "valid lead" in this world is not a call that lasted sixty seconds. It's a call that Google's AI listened to, on a recording that is now on by default, and judged to be genuine: the system tags calls #HighIntent or #ConsultationScheduled and counts the ones that qualify. Duration survives only as a fallback for calls that couldn't be recorded, and ad-interaction data is the fallback's fallback. Which tier fired on any given call is not something you can see from the invoice. Second, the history. Campaign-level performance reports do not carry over. On migration day the old dashboard stops resolving, and the account owner gets fourteen days' emailed notice to export years of cost-per-lead history by hand. Google's own ads liaison put it plainly: start downloading now. If you need the Monday-morning version, the LSA-to-Performance-Max export checklist covers it; what concerns us here is what the migration means. The trade press, to its credit, named the symptom, warning of "a period of metric discontinuity during the transition." It just didn't notice that the discontinuity is the story.
The upgrade is real. That's the trap.
Now the defense, because it's strong and it deserves to be stated at full strength. Google's PM would say: duration was always a terrible proxy. A robocall that lasted 61 seconds was a billable lead under the old definition, and everyone knew it, and advertisers begged us to fix it. The model that listens for scheduling intent is simply a better instrument than a stopwatch. And the best performance marketers I know would say the same thing louder, because they've spent years manually scrubbing junk calls out of their lead counts.
Concede every word of it. I'll go further than the PM would: having watched conversion definitions evolve from inside that building for 18 years, I'd guess the AI classification is right more often than the stopwatch was, probably by a wide margin. If your instinct is to attack the model's accuracy, you're fighting the one battle the platform can always win.
A better unit is still a different unit.
The objection is not quality. It's continuity, and it's custody. The definition of the thing you pay for changed under an unchanged name: the dashboard column still says leads, the invoice still says leads, and January's leads are not September's leads. Call that unit drift: a silent change in what a metric counts, under a name that stays the same. Nothing was versioned. No per-call flag tells you which tier of the definition fired. And the historical series wasn't restated in the new unit so you could keep a trend line; it was deleted, on a fourteen-day fuse. A cost-per-lead chart that spans the migration is fiction wearing the clothes of measurement, and it will render beautifully in every dashboard that displays it.
One party could have preserved continuity, and it happens to be the party paid per unit. That's the custody problem in a single sentence. I don't think there's a villain in the room; I've sat in the meetings where these migrations get planned, and the deletion of the old reports is somebody's reasonable-sounding scope cut, not a plot. But intent is irrelevant to arithmetic. The buyer now pays in a unit the seller defines, audits it against a history the seller retired, and the whole arrangement was shipped as a usability improvement.
The four moves, assembled
Put this next to the rest of the year and a shape appears. The platform now generates the counterfactual: the growth you supposedly missed, computed by the party paid when you act on it. The targets you typed became contracts on August 17. Your account's raw history already sits on a 37-month clock that destroys the market intelligence you paid for, and the migration adds a second, faster clock: fourteen days to keep your own performance record. And now the unit itself has moved inside the platform's judgment. The seller writes the number about the road not taken, enforces the number you agreed to, retires the numbers you'd check either against, and defines the unit all of those numbers are denominated in. Each move shipped separately. Each had a reasonable changelog entry. Read together, they describe a buyer who no longer controls a single term in his own cost equation.
Engineers have a name for the general failure. When a data team changes a column's definition, that's a schema migration, and doing it without versioning, without a changelog, and with the old table dropped is the kind of thing that ends careers, because every model trained on the old schema silently breaks. That is what just happened to the lead, except the schema belongs to your counterparty and the models that break are yours: the bidding baselines, the LTV curves, the CFO's trend line. In control-loop terms, the sensor got swapped mid-flight by the party being controlled. You can't prevent a seller from improving its sensor. You can refuse to let the swap be silent: log the definition, date the drift, and either restate the series or fork it and start a new one. Versioning is not a technical nicety here. It's the entire difference between a measurement system and a billing system.
What drift does to the money
Run the drift through the P&L, because this is where it stops being an analytics complaint. Cost per decision assumes the decision under the line holds still; every vertical variant of it inherits the assumption, and so does the LTV model your CFO actually trusts. A marketing number doesn't stop at marketing: it travels through CAC into working capital and enterprise value, and a drifted unit travels the same road with the same confidence. Proof has a denominator, and a denominator has a unit, and if the unit is a judgment the seller can revise, then every ratio downstream of it is a lease, not a possession. The market, incidentally, just taught Alphabet a version of this lesson from the other side, shaving the stock on a capex hike because inputs aren't units either; that essay is coming.
There's also a governance corollary that connects to the kill condition, and it's the part I'd want every delegation charter to absorb. A kill condition is a price: fire the automation when the number crosses the line. But a threshold denominated in the seller's unit can be dissolved without anyone touching your threshold. Redefine the unit and the number walks back across the line on its own. If your exit trigger is written in a currency the counterparty issues, you don't have an exit trigger. You have a request.
The Unit Ledger
So here is this month's Lab instrument, and like the last one it's deliberately cheap: a page, not a platform. The Unit Ledger: every metric that reaches a board deck gets a registered definition, an owner for that definition, and a drift log. Five fields per metric. Who defines the unit. Whether it's a fact or a judgment. When the definition last changed. Whether history was restated or broken at that change. And what buyer-owned check prices it, which for seller-defined units means the Counterfactual Audit's cousin: your own denominator, computed from first-party data, sitting next to the seller's number in every review.
Most teams that sit down to fill this in will discover the same two things. The "definition last changed" column is blank because nobody logs it, and the buyer-owned check is blank because nobody computes one. That discovery is the exercise. The ledger doesn't argue with the seller's model, and it doesn't need the seller's cooperation. It converts a posture into paperwork, which is the only conversion that survives a personnel change.
Every vendor is about to sell you a unit
Your ad account is just where this became billable first. The agentic procurement wave now hitting every department runs on seller-defined units: the copilot's "tasks completed," the support vendor's "tickets resolved," the agency's "hours saved," each scored by the vendor's own model, each drifting whenever the vendor improves that model, each arriving in a renewal deck as a clean trend line. Outcome-based pricing is coming to everything, and outcome-based pricing is exactly as trustworthy as the definition of the outcome, which is to say: exactly as trustworthy as the party who defines it is disinterested. The disinterested party does not exist. From the CEO chair the question fits in one sentence: when a vendor changes what a unit means, who in our building is required to notice? If the answer is nobody, your cost metrics are already leased.
The migrations begin in August, account by account, fourteen days' notice at a time, and the fastest advice in the industry right now is also the most quietly radical: download your history before it's gone. Take the advice, then take it seriously as a diagnosis. An industry where keeping your own record of what you paid counts as urgent action is an industry that has already conceded custody. Export isn't a chore; it's custody. Write the definitions down while they're still fresh, date every drift, and keep one denominator the seller can't touch. Because the seller's unit may well be better. It just isn't yours, and you can't price a decision in a unit you don't control.