The third week of July was a bad week for numbers. The Wall Street Journal printed a publisher's search-traffic decline as 85 percent and corrected it to 43 within a day. LiveRamp showed that one percent of bad identity data is enough to reverse a channel ranking. And while the industry's independent measurement was failing in public, a new kind of number arrived in every ad account: a precise, causally attributed figure for the growth you didn't get.
Stay with the correction for a moment, because it deserved more attention than it got. Same publisher, same period, half the collapse. The wrong number circulated for a day and reached tens of thousands of readers; the fix followed slower and quieter, the way fixes travel. An industry allocating capital in the hundreds of millions was working from measurement that moved by a factor of two in twenty-four hours. LiveRamp's study made the small version of the same point: in their simulations, one percent contamination in the identity layer didn't dent the ROI reading. It reversed which channel looked best.
So here is the thesis, and it has nothing to do with any single platform: in agentic systems, whoever generates the counterfactual owns the budget conversation. "Here's what happened" is accounting. "Here's what would have happened" is power. It decides which comparison gets made, and comparisons are where budgets move. Every organization delegating decisions to autonomous systems is about to learn that the scarcest asset in the building isn't intelligence. It's a counterfactual you can check.
Now Google can enter the frame. Over the past few weeks, a feature that lived quietly in Google Ads Labs graduated into the Recommendations tab as a beta. It estimates the clicks, conversions, and conversion value your account left behind because budgets were capped or bids ran low. It attributes the cause, budget or bids, and it prescribes the remedy, which you can apply from the same screen. The trade press covered it as a visibility win, and to be fair, advertisers have asked for years what their constraints were costing them. Now there's a number.
The accurate description is different. The estimate is a counterfactual: a simulation of auctions you didn't enter, at bids you didn't set, spending budget you didn't have. It is computed by the only party in the market who can run that simulation, and delivered by the party that gets paid when you act on it.
A quote rendered as telemetry
Be precise about what the number is, because the objection is not the one you've probably heard. I spent 18 years inside this ecosystem and I'd guess the estimate is good. It's built from your account's real auction participation; Google's simulation of your headroom is the best-informed simulation on earth. If your instinct is to attack the accuracy, you will lose that argument, and you'll deserve to.
The problem is a different property. There is no holdout. You cannot re-run last month's auctions without the constraint and compare. No analyst, no auditor, no rival vendor can replay the simulation and check the claim. The number is unfalsifiable by construction, which means it can be accurate, but it can never be evidence. In any other procurement process in your company, a figure with those two properties, generated by the seller and checkable by no one, has a name. It's a quote, and it gets a competing bid.
A quote can be accurate. It's still a quote.
Where the number lives matters almost as much as what it is. In Labs, it sat in sandbox culture: interesting, ignorable, three clicks off the main path. The Recommendations tab is different territory. Recommendations feed optimization score, optimization score feeds the auto-apply reflex, and the whole surface borrows the visual grammar of reporting (figures, attribution, causality) for what is functionally a proposal. In systems terms, something specific happened when that feature moved: a sales channel was reclassified as an instrument.
Two surfaces, one negotiation
Now put the date next to it. On August 17, as I wrote last week, the targets in budget-limited campaigns stop being aspirations: the tCPA someone typed in 2023 becomes the price the system actually delivers at. The Bid Target Adjustment Tool has been sitting in accounts since July 6, offering to help you re-price before enforcement begins. Assemble the sequence in order. A tool arrives to help you rewrite your numbers. An estimate arrives showing what your current numbers are costing you. Then enforcement begins on whatever you agreed to.
One surface generates the case for more spend. The other enforces the number you accept. The account now contains a complete negotiation: evidence, proposal, contract. Every element of it is supplied by the counterparty. Aspiration Debt, in the last essay, was the stock of stale numbers written when numbers didn't bind. The missed-growth estimate is the flow that re-prices them, one recommendation at a time.
The two people who should push back
The strongest defense of the feature would come from the PM who shipped it, and it deserves a full hearing: these estimates are computed from your account's real auction data. The beta label is honest. Advertisers begged for exactly this visibility, and we built it. Concede every word. The provenance is real, the label is honest, the demand existed. That is precisely what makes this worth an essay instead of a complaint: the number doesn't have to be wrong to be dangerous, and a critique that depends on Google fudging the math is a critique that dies the first time the math checks out. The issue survives perfect accuracy. The party generating the counterfactual is paid on its acceptance, and the buyer has no mechanism to check it. Everything else is implementation detail.
The second pushback comes from the best practitioners I know: we ignore the tab. Everyone competent ignores the tab. Also true, and for a decade it was a sufficient answer. I don't think it survives this year, because the number no longer stays in the tab. It travels. It compounds into the optimization score at the top of the account. It arrives in the deck your Google rep brings to the QBR. It surfaces in the CFO's inbox as a question: why are we leaving four hundred grand on the table? The practitioner's ignore-habit worked when the counterfactual was a suggestion; a quantified counterfactual with causal attribution is a different kind of traveler. It shows up in rooms the practitioner isn't in, carrying more precision than the org's own numbers can muster that day. "We ignore it" is not an answer a CMO can give a board. The org needs a standing answer, and standing answers are policy, not habit.
The only number in the room
Why does this month matter more than any of the other months Google shipped a recommendation? Because the same week the seller's number got sharper, the buyer's numbers got visibly worse. The correction. The one-percent flip. Match rates decaying to the point where the trade press asks whether attribution is an illusion. Proof has a denominator, and the industry's denominators are wobbling in public.
Meanwhile the platform's self-measurement is scaling with total confidence. Half a million advertisers on AI Max, at a self-reported 15 percent more conversions at similar ROAS, a figure measured by the system being praised, while the independent tests that found the opposite go unanswered. Three-quarters of ads support queries now resolved by Google's own agents. And on the Alphabet earnings call two days ago, covering the largest advertising business in history, analysts asked zero questions about advertising. Nobody is even auditing the auditor's press release anymore.
The evidence-layer essay argued that the system that spends and the system that verifies cannot share a brain. That was about grading the past. The missed-growth estimate escalates the problem by one tense: the platform now also supplies the subjunctive: what would have happened, what you should feel about it, what to do next. And the infrastructure is consolidating the same direction. The month's other big push, Tag Gateway, asks you to move tag serving inside Google's own stack. Signal in, verdict out, and now the counterfactual in between, all on infrastructure the counterparty operates. None of these is sinister on its own. Together they describe an organization that has outsourced its entire evidence supply chain to its largest vendor, and whoever holds the only stable-looking number in a room full of broken ones wins the meeting by default.
The oldest pattern in the discipline
Every mature engineering field separates the actor from the evaluator. The service does not grade its own uptime; the trading algorithm does not mark its own book; the model does not score its own eval. And within that separation there's a hierarchy, because the most privileged sentence in any control loop is here's what would have happened. The component allowed to emit counterfactuals owns the postmortem, and the roadmap follows the postmortem, and the budget follows the roadmap. Serious architectures therefore keep the counterfactual generator outside the policy; it's the whole reason holdouts, control groups, and shadow deployments exist. Google merged the two in the account UI, and the industry filed it under helpful.
There's a governance corollary that connects to the kill condition, and it's uncomfortable. A counterfactual you can't audit is a decision class you can't take back. Fire the automation, and the seller will always have a number for what the firing cost you, a number with the same unfalsifiable authority as the one that justified the delegation. If the take-back can never be proven right, it will rarely be attempted. Unauditable counterfactuals don't just win budget conversations. They make delegation a one-way door while everyone still believes they hold the handle.
The Counterfactual Audit
So here's this month's Lab instrument, and it's deliberately cheap. The Counterfactual Audit: no seller-generated counterfactual enters a budget conversation until it has been priced against your own denominator and its stability is on the record. In practice: pick one budget-limited campaign (mine would be a Performance Max campaign, because that's where the signal loop concentrates) and log the missed-growth claim weekly for six weeks. Two columns. The claim, and the claim divided by your own cost per decision, computed from first-party data.
Then watch the claim itself. If Google's counterfactual for the same campaign holds steady across six weeks, you've learned something real: the constraint probably exists, and you can price the decision to lift it on your own timeline rather than the seller's. If the claim oscillates week to week, you have established, with the seller's own instrument, that the number is a quote. I'd take that bet in most accounts, though I'm not certain and would love to be wrong. Notice that the audit never argues with the estimate. It can't; the simulation is unrepeatable. It just establishes what kind of object the number is, and either finding is worth more than the number itself.
The card does the same job the kill-condition card did: it converts a posture into paperwork. And most teams that run it will hit the same wall, which is that the denominator column is blank because nobody computes cost per decision from their own data. That discovery is the exercise. You can't out-argue a seller's counterfactual with a feeling.
Every vendor is about to hand you one of these
Zoom out from the ad account, because the ad account is only where this became enforceable first. Every agentic vendor is shipping seller's counterfactuals right now. Your agent saved 350 hours. Your copilot sourced 30 percent of pipeline. Adopters see 15 percent more conversions at similar ROAS. Procurement in the agentic era is going to be conducted almost entirely in subjunctive numbers, generated by sellers and checkable by no one, and the finance function that treats those numbers as telemetry rather than quotes will allocate capital exactly as well as the seller intends it to. The placement rule is one sentence long: the counterfactual generator sits in your evidence layer, or it sits with the seller. There is no third place.
From the CEO chair, the question fits in one sentence too: when a vendor tells us what we missed, what number of our own goes on the table next to it? Silence means every budget negotiation in the building is conducted with one side holding all the evidence, and the side holding the evidence is selling. Your organization already knows how to handle that situation. It just hasn't recognized it yet, because the quote arrived dressed as a report.
The estimates are in the tab now. By Q4 planning season they will be in the deck, and some of them will be right. That was never the question, because accuracy was never the property you could check. Before you act on what you missed, ask who counted. And if the answer is the seller, put a denominator of your own on the table before the number starts traveling, because it travels fast, and it travels alone, and a room with one number in it always agrees with the number. When the seller writes the counterfactual, the buyer owns the denominator, or the buyer owns nothing.