Buy against closed revenue instead of platform conversions

Add up the conversions each platform claims and the total routinely exceeds the client's actual order count. Every platform counts a touch it saw, so the same sale is claimed several times over.

Reconciling against one closed-sales export gives a single number that cannot be double-claimed. The per-platform gap it exposes can be the most valuable thing in the report — and the most uncomfortable.

Bidding algorithms optimize toward whatever you feed them. Feed them platform conversions and they chase the cheapest claimed conversion, not the most profitable sale.

A closed-revenue figure per platform tells you where the inflation actually is, which changes both the split and what you report at the end of the month.

Every serious media buyer eventually proposes an incrementality test, and most of them die in the measurement. A region is held out, a campaign is paused, an audience is excluded — and then the result has to be read against something, which in practice means platform-reported conversions in the treated cells and nothing comparable in the held-out one. The design is sound and the readout is contaminated by exactly the self-reporting the test was meant to escape. Clients notice, which is why holdout results are argued about rather than acted on.

Closed sales from the client's own system are the independent count that was missing, and they can be split by the dimension the test used, provided the revenue file carries the column — a region or store mapped as Location, or an audience segment or product line mapped as Product / service. The comparison is then between two groups measured the same way, from a source neither the agency nor the platform controls. It does not make a badly designed test valid and nothing can. What it removes is the objection that finished most previous tests: that both halves were counted by interested parties.

There is a practical limit worth being honest about. A holdout needs enough closed revenue in each cell to be readable, and in a business with long cycles or low deal counts the test may simply be underpowered however carefully it is run. The report shows that plainly, as small numbers with obvious volatility, which is more useful than a confident result built on eleven sales. Where the volume is there, a geographic holdout read on closed revenue is the most persuasive artefact a media agency can put in front of a finance function.

Creative decisions in paid social are made within days, because that is the rhythm the platform reports in and the rhythm the budget moves in. In categories with a purchase cycle measured in weeks, that loop is faster than the outcome it is trying to measure, so creative gets judged on the leading indicator it produces rather than on the customers it produced. The creative generating the most cheap conversions is not reliably the one generating the most revenue, and the gap is widest in the categories most worth buying for.

A monthly closed-revenue read gives a second, slower loop underneath the fast one. Day-to-day optimization continues as it must, and once a month the reconciliation says which creative cohorts actually produced revenue and at what value per sale. That does not replace rapid iteration and should not; it corrects the direction the iteration travels in, which is a different job entirely. Run alongside daily optimization, the monthly read is what can stop a quarter of testing converging on a local optimum nobody wanted, at the cost of an export and an upload.

It works best where the lead file carries one campaign label per lead — a campaign, an ad set or a creative identifier, whichever level of detail the naming convention supports. The match itself only needs a person's contact detail, but the slicing is only ever as good as the columns supplied. Agencies with disciplined naming get creative-level readings; agencies without get channel-level ones and often decide to tidy their conventions. That tidying has value well beyond this report, which is a reasonable secondary argument for doing it.

View-through conversions are the most contested number in programmatic and paid social, and the dispute is structurally unresolvable inside the platforms. A platform can see that an impression was served and that a purchase later happened; it cannot see whether the impression mattered, and it has no incentive to be conservative about the question. Clients discount view-through credit heavily and arbitrarily, agencies defend it on principle, and neither position is informed by anything external. The argument recurs at every quarterly review and is never settled.

A closed-sales reconciliation does not adjudicate view-through and should not be sold as doing so. What it provides is a fixed total: the revenue the client actually closed in the period, which every claimed conversion of every type has to fit inside. When the platforms' combined claims exceed that total, the excess is not a matter of opinion, and its size tells you roughly how much inflation is in the system. That is a boundary rather than an allocation, and a boundary is enough to move a budget conversation from belief to arithmetic.

Upper-funnel activity remains genuinely hard, and this does not solve it. Where a campaign's job is to create awareness that surfaces weeks later as a branded search or a direct visit, no amount of reconciliation will attribute it, and the revenue sits in Direct / Unknown with everything else that arrived without a traceable source. Presenting that honestly is more durable than claiming a view-through figure the client already disbelieves. The unattributed bucket growing during heavy upper-funnel investment is weak evidence, and weak evidence labeled as weak beats strong evidence nobody can check.