The Two Architectures, Part Three: The Effectiveness Reckoning

July 28th, 2026

Part one argued that media services is splitting into two businesses with distinct architectures. One is pooled, platform-driven decisioning for the median advertiser. The other is per-advertiser decisioning for the enterprise, whose competitive position depends on private valuation the pooled stack cannot accommodate. Part two measured the gap that split leaves on the table. Across roughly 1.4 billion impressions of open-web inventory, clearing price and independent page quality are statistically unrelated. One enterprise advertiser captured a 20% gain for 66% less cost by decisioning against its own value function rather than the pooled consensus.

This piece is predictive. It argues that the same split is now arriving one layer up, in measurement. An external shock will accelerate it. The specific event does not matter. What matters is that the shock will compress a decade of latent knowledge into a single decisive cycle. The claim is not that attribution disappears. Most of the market will keep it. The claim is that the enterprise tier stops using it to decide, and that the next downturn is what converts that from a private position held by a few sophisticated marketers into action taken across the whole tier at once. The trend is already visible. The largest platforms have begun to concede it.

The MMM Concession

Between 2021 and 2025, the three dominant advertising platforms gave away measurement. Meta released its open-source marketing-mix model, Robyn, in 2021. Amazon followed with a self-service model in 2022. Google released Meridian in January 2025. Each tool lets an advertiser build and run aggregate effectiveness measurement on its own data, off the platform’s books.

The common explanation is signal loss. Cookies are deprecating. Apple’s App Tracking Transparency cut the user-level signal that fed deterministic attribution. Aggregate modeling is durable under privacy rules, so the platforms adopted it. This explanation is mostly wrong. Google did not lose its signal. It is logged in across Search, YouTube, Android, Chrome, and Maps. If signal loss were the motive, Google would be the least motivated of the three. Instead Google built the most ambitious of the three tools. A platform’s own signal position does not explain why it released the tool.

The releases are better read as a concession. Enterprise advertisers have spent several years building demand for independent research. They increasingly want to measure effectiveness themselves, in a frame the seller does not control. The platforms released MMM tools to meet that demand on their own terms, rather than cede the measurement relationship to independent parties altogether. When the largest incumbents concede ground to a demand, the demand is real and capital is moving toward it. This is the same signal transmitted by Publicis acquiring LiveRamp. The incumbent reveals where value is going by what it is willing to accommodate.

Three forces built that enterprise demand:

• The scale of the investment. Consumer advertising budgets at large public companies are now big enough to draw direct financial scrutiny. The CMO Survey reported that pressure on marketing from the CFO rose 52% between 2023 and 2025, and pressure from the board rose 21%. Gartner’s 2025 CMO Spend Survey found budgets flat at 7.7% of revenue for two straight years. It also found that more than 40% of CMOs who push for larger budgets lose standing, because they cannot clearly demonstrate return. NielsenIQ reported that 84% of marketing leaders now treat ROI as their primary basis for allocating budget. All advertising budgets north of $100M will soon be measured by the company that spends them.

• Distrust of the biggest sellers. The platforms have a long record of grading their own homework. Facebook settled advertiser claims that it overstated video-viewing time for $40M, after plaintiffs alleged the inflation ran as high as 900%. Evidence from US vs. Google, meanwhile, included a letter from IAC’s Barry Diller saying his price for 500M website visits from Google search increased from $21M to $300M over five years.

The agency channel fares no better. The ANA’s programmatic transparency study found that of every dollar entering a demand-side platform, only about 36 cents reaches a valid, viewable, measurable impression. Roughly a quarter of open-web spend is lost to waste. A CFO who has read or even heard about those findings cannot take the seller’s number on faith.

• Pressure to adopt AI and run leaner. Machine learning has made aggregate measurement fast and cheap to operate. Work that once took a consulting engagement 6–12 months now runs in weeks, and refreshes continuously rather than quarterly. The same boards demanding return are demanding AI adoption. Self-operated measurement is one of the clearer places to show both at once.

The walled gardens’ concession is bounded. The boundary is the strategy. They will give ground on measurement. They will not give ground on decisioning. The MMM they ship cannot be neutral. An honest aggregate model does one thing above all others: it separates demand the advertising created from demand the advertising merely harvested. The optimizer’s credit depends on never drawing that line. So the tool that reaches the market is calibrated to protect the credit. Meridian’s defaults favor search. Robyn favors Meta impressions. A measurement system built by the party being measured cannot be built to indict its own product. The platforms concede the measurement and keep the decision. PMax and Advantage+ preclude private input to valuation. The bargain reads plainly once you see it: “You may measure. We will make the decisions.”

The Distrust Is Already Funded

Attribution means an ad preceded the conversion. This is useful for small advertisers, whose product is unknown to consumers who have not seen an ad. For large advertisers, many sales stem from preexisting demand: repeat buyers, and from intent generated where the attribution system cannot see it — by television, by reputation, by years of physical and ambient presence.

Tying machine learning to attributed signal, despite the wrinkle of pre-existing demand, was a mismatch of standardized tooling to a nonstandard use case. The platforms’ AI systems are trained to achieve the lowest possible cost per attributed sale. But lowest attributed cost per sale for the enterprise does not come from creating demand. It comes from taking credit for demand. An optimizer told to minimize attributed CPA is, in structure, an instrument for harvesting demand created elsewhere. This is not a defect in any platform’s execution. It is what minimizing cost per attributed conversion means. The better the system performs against the objective it was given, the more completely it claims credit for sales that would have happened anyway.

Attribution does not measure effectiveness. It measures the ability to predict sales from preexisting demand. For most advertisers the two are close enough to be the same. For the enterprise advertiser whose franchise is built on the demand it generates, they are opposites.

Enterprise demand for independent measurement is not talk. It is already spending. If enterprises believed the seller’s number, they would manage to it and stop checking. They do not. They pay, now, to re-measure it. They run third-party verification alongside platform-reported conversions. They commission log-level audits of their own supply chains. They run geo holdouts and incrementality tests in parallel to the dashboards they nominally optimize against. In most large advertisers there is already a budget line for not trusting the seller’s measurement.

Some have gone past checking. Firms tracked by Adweek and eMarketer report enterprise clients moving decisions onto their own first-party data and independent measurement, and away from the walled gardens’ reported results. That spend is the market revealing a belief it has not yet fully acted on. No one triangulates a number they trust. The conviction is established and funded. What is missing is not belief. It is a forcing function.

The Shock and the Lag

Nascent trends do not begin in crises. They accelerate in them. The Nasdaq collapse, the financial crisis, and the pandemic each took a behavior already underway and stripped the slack that had let firms postpone it. The crisis is not the cause. It is the removal of the excuse.

In advertising, the slack is abundance. Inside a growing budget, an advertiser can fund the harvesting the attribution system rewards and the creation it ignores at the same time. It never has to choose between them. Platform measurement’s invalidity costs nothing when nothing is being cut. This is why a decade of advertisers knowing the number is wrong has changed almost no behavior. Knowing has been free.

A downturn ends that. A real cut forces the question the dashboard answers backwards. The measurement-optimal cut protects the lowest attributed CPA and removes the highest. That means protecting the harvest and severing the creation. For an enterprise whose position depends on the demand it generates, that cut turns a soft quarter into a lasting loss of share. The sophisticated ones have watched it happen to competitors across prior cycles. Under pressure, they cannot act on a number they already disbelieve. The shock teaches them nothing new. It ends the affordability of pretending they do not know it.

Each time financial markets were shocked, institutions with capability to act on their own valuation took share. The ones still managing to the broker’s number got burned. That capability is now being built in advertising. The next shock is when it will be used.

The Split That Does Not Reverse

Attributed CPA will not disappear. For most advertisers it is good signal. The SMB whose private information is product and creative, rather than per-customer valuation, benefits from pooled data on big brands’ repeat sales. Platforms serve them well. They lack both the means and the stake to build an alternative. They conquest the same attributed sales big advertisers can only harvest. Below the sophistication threshold, platform advertising and attributed CPA persist and grow.

Above that threshold, there will be rapid change. The enterprise tier decisioning against its own value function, measuring against its own definition of effectiveness, will accelerate. Their measurement will be incrementality-anchored. It will be private. It will be unavailable to any counterparty that profits from the harvest. The measurement layer will stratify to support what’s newly possible at the decisioning layer. It will split along the same line, for the same reason. A pooled proxy cannot serve an advertiser whose value is private. It was not built to.

Once an enterprise has invested in measurement it controls and validated it against ground truth, the cost of returning to the seller’s number is higher than the cost of keeping its own. Today, independent measurement exists in tension with platform attribution and now with platform MMM. After the shock, enterprise preference will be clear and irreversible.

The prediction is not that platforms decline. Platform advertising and attributed CPA are permanent for most of the market. The prediction is the largest advertisers will walk away from them. It is perhaps only slightly easier to imagine than platforms declining.

In fact, this change is already in motion, slowed only by inertia. In the context of a forcing function, it will happen across the tier all at once, rather than as a slow drift, because shock removes the option to wait and see “how the whole AI thing shakes out.”