Somewhere this week, a marketing director is handing an AI agent write access to a month of live ad spend. The pitch writes itself: the models are cheaper than ever, they can act unsupervised, so let one run the budgets.
The problem is what these systems are. They predict, they do not calculate. A model produces the most probable next step, not the correct one, and the two diverge fast once real money is moving every hour. Give a probabilistic system write access to spend and you import errors nobody can predict, least of all the system itself.
What decides whether a budget performs is not a smarter guesser. It is the deterministic maths of how the money is released, over time and across campaigns. That maths is pacing. It has a right answer, and you cannot chatbot your way to it.
Pacing gets treated as an art. A marketer glances at spend on a Friday afternoon, decides it looks about right, and nudges a number. It is nothing of the sort. It is a science with a provably correct answer, the core of it settled by economists in the 1850s, and almost nobody computes it.
Two questions hide inside the innocent "are we on budget?". The first runs across time: will the month land on target in a smooth glide, rather than blowing the budget by the twentieth and going dark, or dribbling it out and handing money back at the close? The second runs across campaigns: is the next pound going where it earns the most, right now? Most teams eyeball the first and never touch the second.
Pacing across time is not budget divided by days left
Naive pacing is the budget remaining divided by the days remaining. It fails because days are not interchangeable. Check the account at six in the evening and most of today's spend has already happened; treat the whole day as still ahead and you over-correct, starving tomorrow to fix a problem that does not exist. You weight by the shape of a real spending day instead: quiet overnight, climbing through the morning, a long midday-to-evening plateau, tapering late.
The calendar lies too. In a B2B account, weekends barely convert, so you pace across the working days that remain, not the raw days on the wall. The daily figure that lands the month exactly on budget is a different, truer number than the arithmetic hands you.
None of this is about tidiness. A budget that lurches between feast and famine keeps resetting the algorithm's learning and drives cost per acquisition up with it. Smoothness is performance.
The next pound has a right home
Every campaign sits somewhere on a saturation curve. The first pounds buy a lot. Each extra pound buys a little less, until you reach a ceiling where more money buys almost nothing, because rank, not budget, is now the limit. You can estimate where a campaign sits from what it spends and what it loses to a constrained budget; on search, the impression share lost to budget reads the headroom off directly.
Splitting a fixed budget across campaigns is where most accounts leak, and it is the part with a theorem. Fund every campaign to the point where the last pound spent on each one buys the same value. If campaign A's next pound buys more than campaign B's, move money from B to A, and keep moving it until their last pounds are worth the same. Economists call this the equimarginal principle, and it provably maximises total conversions, or conversion value, for the money you have.
A picture helps. Pour the fixed budget in and let it find its level, filling every campaign to the same waterline of marginal return.
Make it concrete. One campaign's next ten pounds would buy two leads. Another's next ten pounds would buy five. The money obviously moves to the second. Doing this properly means making that call across dozens of campaigns at once, every day, and no Friday glance ever will.
The raw maths needs two disciplines or it misbehaves. The first is an exploration floor: keep funding a new or unproven campaign enough to learn what it can do, because information has value too. The second is a coverage floor: do not strip a proven, efficient campaign that is capped by rank just because the numbers momentarily prefer a neighbour. A genuinely wasteful campaign still gets cut hard. It just never gets cut blindly to zero.
Targets come from last month's actuals
All of the above rests on knowing what each campaign is worth, which means cost-per-acquisition and return-on-ad-spend targets set from the last full month's actuals, not a figure someone typed into the platform a year ago and forgot. This gets sharper from 17 August, when Google's change to budget-limited bidding starts spending harder against whatever target it finds. A stale, over-loose target stops being untidy and starts being expensive.
The honest ending is a measurement, not a number
The convention here is to close with a number. This approach lifted return on ad spend by some flattering percentage. I refuse, and the refusal is the point of the piece.
The equimarginal maths is not in doubt. Given the curves, that split is optimal and it beats a Friday glance every time. You do not have to believe me; you can derive it. That is a stronger thing to be able to say than any case study, which is exactly why a percentage would cheapen it: a testimonial asks you to believe, where the maths lets you check.
What honesty does demand is measuring the right input. Wherever a brand reports its own orders, the engine scores each campaign on that real revenue, not on the conversions the ad platforms award themselves, because a platform grading its own homework is how budgets get optimised toward numbers that do not survive contact with the bank. Get the input honest and the allocation is settled.
That restraint is the argument. Pacing is one of the few places in marketing where you can know rather than believe, so the right thing to build is the apparatus for knowing and the discipline to use it over a Friday glance. So we built it. The Crane Engine™ is the spine of the data and automation under every paid media account. Every morning it recomputes the whole answer: the intraday-weighted glide, the equimarginal split, each target reset from last month's actuals. The right number is already waiting before anyone opens the account, so no one has to remember to.
If you want to test your own account today, it takes one question. For each campaign you run, what did the last pound buy yesterday? If the answers differ and you cannot say by how much, the money is sitting in the wrong places, and the Friday glance will never show you where.
