What Hourly API Scaling is actually worth

Two brands running paid social through Meta were put on the same operating process: budgets are moved programmatically and many times a day, driven by what the account's own delivery is doing that hour, rather than adjusted by hand once a day. We call this Hourly API Scaling, and each individual move a scaling action.

How those moves are decided, sized and executed is our own method and is not described here. What follows is only the measurement — what the actions produced, and how much of it would have happened without them.

The counterfactual we are trying to price is simple to state and awkward to measure: if nobody had made the change, what would the account have spent? Everything below is an attempt to answer that without borrowing the answer from the result.

A second question sits underneath the first. Both brands received the same budget process, but only one received the Industry CAC work — the analysis that establishes what an acquisition should cost in a category and therefore how far an account can profitably be pushed. If the budget mechanism is worth the same on both and the outcomes differ by an order of magnitude, the difference is attributable to the baseline, not the mechanism.

Delivery comes from the Meta Ads API at campaign × hour granularity. Order and revenue truth comes from the Shopify Admin API and from Statlas MCP, which holds the modelled and reconciled daily figures used for efficiency benchmarking.

The record of what was changed, and when, comes from the account's own activity log, read at day-level granularity across every day of both measurement windows.

Every measured day is scored from the hour of its first scaling action through to midnight. The actual spend in that window is compared against two different counterfactuals:

• Isolated action. The baseline is the same hours on the most recent day with no scaling action on it at all. This asks: how much more did this change produce than the state immediately before it? It is a strict floor, and it deliberately gives no credit for the fact that the day before was itself elevated by an earlier change.
• Compounded. The baseline is the last clean day before the programme began, held fixed. This asks: how far has the account travelled from where it started? It captures the ratchet — a floor that is raised tends to stay raised — which is the entire point of running budgets this way.

Neither is the true number on its own. The isolated reading understates because it treats an elevated starting point as though it were free. The compounded reading overstates because it credits the budget mechanism with everything that happened after the programme started, including creative and offer work. The honest answer is bracketed by the two, and we report both throughout.

There is no clean holdout here and we are not going to pretend otherwise. A held-out set of days was considered and rejected: on Brand B the process ran from the first day of the engagement, so no untouched days exist to hold out. The comparison is therefore within-account and within-campaign — the same campaign, the same hours of the day, against its own recent history.

Held constant. Campaign identity, hour-of-day window, account, currency, and attribution setting. A campaign is only ever compared against itself.

Not held constant, and material:

• Creative volume. Both accounts launched new creative throughout. On Brand B the relationship runs the wrong way for a creative explanation — 84 new assets in April moved spend barely at all, 43 in May moved it more than fourfold — which is addressed directly in section 06.
• Offer and landing page. Brand B changed its offer page mid-period. Brand A changed feeds on 13 July.
• Auction conditions. Uncontrollable and unobservable at this granularity.
• Efficiency. Brand B's cost per purchase fell 22% over the window. That is the Industry CAC layer working, and it is a confound for the budget layer — which is precisely why the two brands are reported separately.

The consequence is stated plainly: the isolated figure is defensible as a causal floor, because a same-campaign, same-hour, recent-day comparison absorbs most of these. The compounded figure is a programme-level result and should be read as the value of the whole operating process, not of any single scaling action in isolation.

The comparison is deliberately narrow. The account is only ever measured against itself, over the same hours of the day, on the most recent day that carried no scaling action at all. Nothing is compared across accounts or against a forecast, so the differences below cannot be produced by mix or by a model assumption.

Across the 20 pairs the account spent €37,610 in those hours on its comparison days and €55,068 on the days a scaling action ran — €17,458 more, a 46.4% difference. 5 of the 20 pairs are negative, which is the point of showing all of them rather than a summary statistic.

The account had two regimes, and the boundary is a single day

Scoring each day's spend against the budget actually available to it separates the period cleanly. For nine consecutive weeks the account spent between 89% and 115% of everything it was allowed to spend. Budget was the binding constraint, and every euro released was taken.

Why the creative counts do not explain any of this

The obvious rival explanation is that creative volume drove the growth. The data runs against it, and not marginally:

• April — 84 new assets, spend broadly flat at €856/day.
• May — 43 new assets, half as many, and spend rose more than fourfold to €4,402/day.
• July — 110 new assets, the largest batch of the period, for an increase of roughly €3,400/day.

The rank order is inverted, so more creative did not buy more spend. What does move with spend is order value. Across the period it rises from €58.81 to €89.88 — a 53% improvement — and its week-by-week correlation with daily spend is r = +0.67, against +0.46 for creative volume on far too few points to mean anything. Cost per acquisition falls 6% over the same stretch while spend rises more than thirteenfold.

That is the relationship the Industry CAC work exists to find. A higher order value raises what an acquisition is worth, which raises what the account can afford to bid, which is what allows it to clear at far greater volume without the economics degrading. The account did not scale because it was handed more assets to run. It scaled because each order became worth more, and the ceiling on what could profitably be paid moved up with it — the same assets simply became affordable to run harder.

Both series trend upward across the period, so the correlation is supporting evidence rather than proof of direction. It is reported because it is the one relationship in this account that holds across the whole window, where creative volume does not.

What holds up

• Two independent readings agree on Brand A. The strict pairing gives 15.4% of account spend and the compounded reading 30.8%. A third check that shares no assumptions with either — scoring each campaign-day against its own delivery pattern rather than a comparison day — lands inside the same band. Different failure modes, same conclusion.
• The per-action effect replicates across accounts. The size of the response to a single scaling action is close to identical on two unrelated advertisers in different categories. Account-level impact is therefore a question of dosage, not of effect size.
• The regime split is not a story imposed on the data. It falls out of a single ratio computed per day, and the boundary lands on one identifiable action.

What does not hold up

• The compounded figure cannot separate layers. On Brand B it credits the budget mechanism with growth that the offer, the creative and the Industry CAC work also produced. It is a programme result. Anyone quoting 19.0% as the value of one scaling action alone is misreading it.

Two ways this measurement goes wrong

• Account-level scoring. Measured across a whole account, the effect is diluted by campaigns that were never touched. Every figure here is scored per campaign.
• Multi-day windows. A window that spans midnight collapses to its start date unless it is expanded deliberately, which files the second day as an untouched comparison day. Left alone, the highest-spending day in the period ends up serving as a control for itself.

Read the two compounded intervals together and the pattern is the useful one: Brand A at 30.8% [23.2, 38.5] and Brand B at 19.0% [13.8, 25.4]. The intervals overlap. On two accounts of very different size, in different categories, at different stages, the budget programme lands in the same band — call it 19–31% of account spend.

The two instances above answer what happened where the process ran. This section asks the inverse: across a portfolio of brands that were never put on it, how much constrained spend was sitting unreleased?

The model has three parts, and each is separately falsifiable. A trigger identifies the campaign-hours where an account is demonstrably running into its own limit while performance is holding. The conditions are tuned per brand and are not published here. A value is attached to each trigger, calibrated so that a brand where the process genuinely ran reproduces the result it actually achieved — not assumed, and not chosen to flatter the outcome. Compounding is not a parameter at all: raises persist, so the raised floor becomes the base the next trigger fires from.

Where brands actually land

Mean 27.4%, median 28.0%, modal band 25–30%. The three agree closely, and the spread is tighter than the headline range suggests: half of all brands sit between 23% and 34%.

Run over a single month the same analysis looked bimodal — brands piling up at both ends with almost nothing in between. That turned out to be an artefact of the short window. Over thirty days a brand either happens to catch a constrained stretch or it does not, and the result swings on that accident. Over ninety the runs average out and a single-peaked distribution shows through. It is worth stating because the bimodal version was the more striking chart, and it was the wrong one.

How much to trust this

Read the portfolio total with more confidence than any individual row. The aggregate is built from 37 brands and is stable; a single brand's figure carries roughly ±50%, because it rests on that account's own trigger pattern over one window.

The counterfactual is also unobservable by construction. On a brand that never ran the process there is no version of history where the budget was released, so a trigger is evidence of unmet demand rather than proof of what would have been delivered.