What we did with your model, and what it taught us.

i.Executive summary

We took your draft May MMM, validated it against our own numbers, and combined it with an independent cash-payback model to build a sequenced, gated reallocation of the paid account. Every planned change is live and verified. The system then did the most useful thing it could: it refused to over-scale. As cohorts matured, the early-June "room to ramp" read turned out to be a maturity artifact. Blended predicted CAC now sits right at the six-month line, but that aggregate is carried by mobile and the cheap build-out geos, while the big desktop lanes are over their ceilings, so the disciplined gates correctly held those lanes flat rather than scaling into them.

Every payback and CAC figure below is drawn from the canonical warehouse dashboard and labeled by lens (matured cohort vs snapshot). Every account change is reconciled against the live Google Ads account as of July 13.

ii.The two-lens operating model

Your MMM answers one question precisely: which channels are incremental within a market. It does not, on its own, tell us which geographies we can afford to buy. So we run it beside a second, independent lens, a per-geo cash-payback model built from realized subscription revenue. A dollar moves only when both lenses agree, and a channel goes to test when they disagree.

The reason both lenses are load-bearing is that platform last-click and MMM incrementality invert each other in this account. Branded Search looks best on last-click and worst on incrementality (people would have converted anyway). Demand Gen / YouTube looks worst on last-click and carries real view-through demand that leaks into the organic bucket. Judging spend on last-click alone would scale exactly the wrong channels. That inversion is the single most valuable thing your model surfaced, and it is the backbone of everything that follows.

iii.What the May MMM told us

You delivered the draft model as four market cuts (US-CA, EU-UK, India, ROW), each showing cost per incremental New MRR by channel. We reconciled it against the source workbook cell by cell, it matched, and we adopted it for relative channel ranking. We deliberately did not repeat the model's absolute 86%-paid-incremental figure until the baseline is defended (see the agenda in section 8).

Figures are cost per incremental New MRR from the draft May model. NonBranded is the most efficient channel in every market; Demand Gen carries almost no incrementality outside US-CA while absorbing the largest single budget line. That contrast set the entire reallocation.

iv.The plan we set

We wrote it up as an eight-part action plan (one master, seven channel and geo sub-plans) with owners, gates, and kill switches. The posture: rebuild momentum by ramping paid back up, but gate every step on a six-month blended cash payback and, on the contested channels, on a real incrementality experiment before any large cut.

v.What we implemented, confirmed live in the account

Every row below was reconciled against the live Google Ads account on July 13. Meta changes were executed in Ads Manager (Eric) and Apple in the Search Ads UI (no API); those are noted. Nothing here is aspirational, it is the current state of the account.

Owners: Matt (Google Search, exec gating), Shanik (PMax + Demand Gen, in-house), Eric (Meta). Google account changes were executed through a gated dry-run then commit workflow on the production account.

vi.What we learned

1. Cohort maturity moved the answer by ~1.55x

The plan's headroom map was built in mid-June on a predicted-CAC read that imputes subscriptions for young cohorts. As those cohorts matured, the read re-scored upward, uniformly, by about 1.55x. The blended desktop CAC that looked like $72.87 (a ~3.5-month payback, room to ramp) matured to about $111 to $113 (a six-to-ten-month payback, at or over the ceiling). This is not a data error, it is the difference between an immature snapshot and a matured cohort. The lesson we now enforce: gate only on cohorts with at least three full days of maturity, and read the cumulative matured window, never a single fresh day.

2. Blended sits at the line, but the desktop majors are over

Payback runs on blended predicted CAC (total spend over total predicted subs, all platforms), which is $86.57 on the Jul 1-11 cohort, sitting right at the ~$92.6 six-month ceiling. That number is healthy only in aggregate: it is held down by mobile ($48.48, which monetizes far worse and cannot yet be priced on this model) and by the cheap build-out geos. On the desktop cut that carries the value, the lanes we prioritized to scale (US, EU ex-UK, UK) are all above their ceilings once cohorts mature. This is the most important thing to reconcile against your June model, and the room is in the build-outs, not the majors.

3. Incrementality tests: we re-sequenced away from CTV

Your first proposed test was CTV. We put a POV to you that the first measurement dollars should go to the two video channels we already spend the most on and can least see, Google Demand Gen / YouTube (~$29K/day pre-cut) and Meta video prospecting (~$1.16M/mo run-rate), with CTV moved to a later wave and given a faster instrument (logins and downloads as the primary read, matched pairs, predicted payers as the money metric, rather than a four-week New MRR read on ~$85K of net-new spend). Demand Gen is the highest-value test because its view-through demand is the most likely driver of the rising organic share in the same window we scaled video.

4. The attribution gap was the quiet win

Unattributed desktop spend fell from about $442K in June to about $79 in July. This was primarily a warehouse fix: the Fivetran pipeline began resolving country-level ad spend, so spend that used to fall into an "Unknown" geo bucket now carries a region. The geo-scoped campaign restructure (moving budget out of broad multi-geo campaigns into region-specific ones) helped secondarily. Last month roughly 16% of desktop spend could not be tied to a geography, which meant every regional decision, including several in this plan, was made on 84% of the money; that blind spot is now essentially closed, so this month's regional reads are trustworthy in a way June's were not. One caution, and it matters: this is a reclassification, not a spend cut, so the apparent blended-CAC improvement it produces is mostly an artifact, not a real efficiency gain. We are not banking it as progress.

vii.Results in the account

The reallocation is real and visible: Google spend pulled back to a June 20 trough, then re-ramped, EU-PMax-led, to a July plateau. Money moved out of the Demand Gen leak and trimmed NonBranded, into EU PMax and the new build-outs. Overall paid is flat, by design, because the gates refused to add budget to over-ceiling lanes.

viii.Reconciliation agenda for your June model

We have not seen the June refresh yet. These are the specific questions we want to hold the two models up against, live. The goal is one agreed source of truth where the MMM's channel view and our matured geo-payback view currently diverge.

1. Do the channel verdicts still hold?
   Does the June model still rank NonBranded as the universal hero and Demand Gen / YouTube as the near-zero-incrementality leak outside US-CA? Any lane that flipped is a real signal we need to act on.
2. Does June reflect what we changed?
   The June spend it trained on contains the Canada cut, the EU PMax ramp from zero, the Demand Gen pull-back, and the Branded cap-down. Does the model see those structural breaks, or does it read them as noise?
3. Reconcile MMM incrementality against our matured geo-payback.
   Our matured read puts US, EU, and UK over their six-month ceilings and the build-outs (ROW, AU/NZ, India) under. Where does the June model agree, and where does it say we are wrong to hold?
4. Defend the 86%-paid-incremental baseline.
   Your model attributes ~86% of conversions to paid as incremental; our word-of-mouth and organic read is closer to the inverse. We held off repeating the figure upward until the baseline is defended. What is the June model's organic baseline?
5. Does the model reflect the closed attribution gap?
   June desktop spend was ~16% geo-blind; July is under 0.1%. Regional MMM cuts trained on June inherit that blindness. How should we treat the June geo splits given the gap?
6. CTV: status and instrument.
   Given the re-sequencing, is CTV still on the roadmap, and if so, do we agree on the faster instrument (logins/downloads primary, matched pairs, predicted payers as the money read) rather than a four-week New MRR read?

ix.Recommended forward plan

The matured read rewrites the priority order. This is the plan we are implementing next, and it is deliberately the inverse of the June-era EU-first ramp.