The short version

The five things that matter, then the workings.

Four independent reads on September

The client goal is £260,511, split £63,761 subscription and £196,750 everything else. Here is that number against three reads built from data rather than from a target.

The business is growing fast. Comparing the same window in both years, 1 to 24 August, revenue went from £83,113 to £151,980. That is 1.83x year on year. Carry that rate onto last September and you land near £193,877. The client goal is 34% above the rate the business is currently compounding at, and 2.46x last September. It is not a fantasy number, but it does not happen by carrying on as we are.

September, built from the bottom up

Revenue splits into three pieces that behave completely differently. Only one of them responds to ad spend, and sizing the budget against the other two is the mistake that wastes money.

The subscription line, and how it has moved

In August the client's subscription target was the single biggest source of error. The September picture has changed twice in five days, and both changes are worth recording.

Until 21 August the subscription forecast was inferred from billing history, because the Shopify API refused access to subscription contracts. That block is gone. Pott's Recharge account is directly readable, so we now have the actual forward charge ledger: 3,082 active subscriptions, every one with its next charge date and amount. This is a schedule, not a model.

changed since 21 Aug On 21 August this page said September would beat her subscription target by about £5,310. Two things have happened since. She raised her own subscription number from £58,897 to £63,761, moving it £4,864 toward our ledger. And our ledger came down, because 86% of September is now genuinely queued in Recharge with a date and an amount (£54,754), leaving only £8,897 still resting on our roll-forward assumption. As real queued charges have replaced the roll-forward, the number has settled lower than the roll-forward predicted. September now lands about £4,565 short of her subscription target, not ahead of it.

That is a useful lesson about the method rather than a problem with the business: rolling a renewal forward at the customer's billing interval systematically over-predicts, because it assumes no cancellation, no skip and no failed payment between now and then. The closer we get to the month, the more of it is real and the more the estimate firms up. Expect the same drift in the October figure, which is currently £81,965 and still mostly rolled forward.

Range on the September number is £57,286 to £61,742 depending on how many charges skip or fail. We have used £59,196. Still worth confirming with Ailis that her £63,761 means recurring charges only, and does not include first-time subscription orders, which land in new-customer revenue and are running £20,593 across the month to date on top.

How much does more spend actually buy?

Every number below turns on one question the last version of this page never asked: when we spend more, does new-customer revenue rise in step, or does each extra pound buy less than the last?

The answer is expressed as an elasticity, b. If b = 1.00, new-customer ROAS is the same at every budget and scaling is free. If b = 0.80, a 10% budget rise buys 8% more revenue and ROAS slips as you climb. The old model implicitly assumed b = 1.00 by holding NC ROAS at 1.65 in every scenario.

The best evidence available is the budget cut on 19-20 August, because it is the only deliberate, sustained budget change in the data, which is exactly the kind of move a September ramp would be. Comparing matched weekdays either side of it:

Cutting spend by 20% only cost 17% of new-customer revenue. Run backwards, that is the warning: putting the money back will not buy proportionally more.

The regressions do not agree, and the disagreement is the finding

Fitting b across 269 days of daily data gives different answers depending on which days are included. That is worth showing rather than hiding behind a single number.

They split cleanly into two camps. Anything that includes peak season lands near b = 1.00, meaning no decay at all. Anything that excludes it lands near b = 0.80. Both are real. In November there is enough demand in the market that £3,427/day still converts at 1.91 NC ROAS. In April to August there is not. September is not November, so this plan uses b = 0.82, which is what the August budget cut measured and what the off-peak and in-season regressions independently agree on.

What each budget level actually delivers

Priced against the measured response curve (b = 0.82), calibrated on the 6 days since the consolidation, when the account ran £1,443/day at 1.73 NC ROAS. Profit is contribution at 56.5% less ad spend less £28,896 of monthly overhead.

For comparison, the same budgets under the old constant-ROAS assumption. The gap between the two tables is the cost of that assumption: at goal-chasing spend it overstates September revenue by £17,429.

The number that should drive the decision is the marginal one

Average NC ROAS is what a dashboard reports. It is not what decides whether the next pound is worth spending, because the average is dragged up by cheap early spend we would make anyway.

Under diminishing returns the marginal return is b times the average. At today's 1.73 average, the marginal pound is really earning 1.42. That reframes the whole question:

Read the last two columns. Inside September, every marginal pound comes back as less than a pound of contribution, and has done for some time. Over three years, every marginal pound comes back as several. Average NC ROAS would need to reach 2.15 for extra spend to wash its face inside the month, and no month in the 9-month history has reached it (the best was 2025-12 at 1.97). So the old conclusion holds and is now properly founded: scaling is a cash and lifetime-value decision, not a profit decision. What has changed is that it is no longer free, and the price rises the harder we push.

How wrong could the response curve be?

Everything above rests on b = 0.82. Since the regressions genuinely disagree, here is the goal priced at each plausible value rather than only at our best guess.

The spread is the point. Depending on which reading is right, hitting her number costs anywhere from £87,993 to £119,237 of ad spend. No one can close that gap from a spreadsheet. It closes by stepping the budget up and reading the result, which is what the plan below does.

A correction to the 21 August evidence

That version argued saturation was not a risk, on the grounds that cost per new customer was flat across spend quartiles. The quartile table was right; the conclusion drawn from it was not.

Pooling every day of the year and sorting by spend fills the top quartile with peak-season days. It reads as high spend still performs when what it actually shows is November performs. Split by season and the top quartile drops to £1,584/day for 2026, nowhere near the £3,419/day the goal needs. The flat-CAC evidence was real but it only covers the range we have already operated in. It says nothing about the range the goal requires.

What still stands from that version. In November 2025 the account did spend £102,810 in a single month at 1.91 NC ROAS, so it has absorbed goal-level budget before without breaking. The caution is unchanged and now better quantified: last September the account spent £29,782 in total, and the goal asks the same seasonal window to absorb 3.4x that.

Recommendation

Do not jump to £3,419/day on 1 September, and do not sit at £1,443/day either. Step it up weekly and read the result at each rung. The ramp is not a compromise between the two, it is the experiment that tells us which of the readings above is true.

That path spends £69,650 and lands near £218,876, which is 84% of the goal. It gives up £6,310 of September profit against holding today's budget, and buys about £50,447 of extra contribution over three years. Under the old constant-ROAS assumption the same spend would have been forecast at £223,618, so expect this plan to look like it is underperforming if anyone scores it against the old numbers.

The trigger to stop stepping up is a sustained 7-day CAC above £50, equivalent to about 1.45 NC ROAS. That is tighter than the £55 used on 21 August, which was set when the model assumed no decay. It is not a falling blended ROAS. Blended ROAS will slide throughout the ramp because spend is growing against a £108,696 base of subscription and returning revenue that ads do not touch. That slide is arithmetic, not a leak, and reacting to it would cut the wrong thing.

Three things to settle before 1 September. One, confirm with Ailis that £63,761 means recurring charges only. Two, the cash question is real and now larger than we said: reaching her number costs £102,570 of September ad spend against Q4 inventory commitments, and the return arrives in month four rather than in September. Three, her sheet carries a £14,000 online day on 12 September for the Autumn Launch, which is 7% of the month's online target resting on one day. If that launch is not built, the plan is short before the month starts, whatever we spend.

Rest of August, restated

The budget change on 19 and 20 August, and what August now lands at.

Her plan for 21 to 31 August assumes £66,000 of non-subscription revenue, against a run rate of £50,373. That is a 31% lift, resting on the bank holiday push on the 25th and the Tomato Vine dates. Possible, but it is the optimistic end.

Google: in the numbers, but not a lever

Worth being explicit, because this was the original question.

Every campaign is spending well under budget with no impressions lost to budget. Google is finding as much demand as it can at current bids and is limited by the auction, not by us. It is forecast in this plan at £240/day, treated as a floating input we predict rather than a number we set. That is the opposite of Meta, where the budget is a decision.

Assumptions, so they can be argued with

Not modelled: a September promotional calendar (none supplied), any October pull-forward, and any Google scaling. All three are upside to the numbers above rather than risk.