Ridgeline Skincare: when the attribution model and the mix model disagree

They disagree because they answer different questions with different failure modes, and the current mix has pushed both toward their weak spots.

MTA sees only conversions it can observe and credit. A 30% loss of measured social conversions does not remove the conversions; it moves them into modeled estimates that favor the channel with the most signal - paid social. Meanwhile CTV has no click, so its contribution arrives as direct and organic traffic that MTA assigns elsewhere. MTA is systematically upper-funnel-blind, and its 0.6 on CTV is close to an assumption, not a measurement. Its 4.2 on social is inflated by the same mechanism.

MMM handles carryover and cross-channel effects but is only as good as the variation in the data. If brand spend has been cut steadily for 18 months, there is little independent variation left, and the model may be reading trend and seasonality as CTV effect - or, conversely, understating it. Saturation estimates above $700,000 a week are more trustworthy because spend has actually varied there.

Which governs which decision. Use MTA for tactical in-channel decisions - bid, creative, audience, budget pacing inside a channel where it can see outcomes. Use MMM for cross-channel allocation and for anything with carryover. Neither should decide the CTV question, because that is precisely the decision both handle badly.

One signal the CFO should notice: blended CAC rose from $41 to $58, above the $46 first-order margin, so Ridgeline is now paying more to acquire than the first order returns - a mix problem, not a channel-ROI problem.

Design: a matched-market geo holdout with an intent-to-treat read.

• Cells: 24 matched DMA pairs, paired on pre-period new-customer volume, category penetration and prior CTV weight. Within each pair, one market goes dark on CTV for 10 weeks; the other continues at current weight. Randomize which side of each pair is dark.
• Why 24: 12 pairs gives adequate power for the size of effect that matters. With weekly new-customer counts and pre-period covariance matching, this design detects roughly a 5–7% change in new-customer volume at 80% power. That threshold is chosen deliberately: a CTV effect smaller than 5% does not justify the $4.8 million annual spend.
• Duration: 10 weeks - two weeks for the dark-out to take effect, four weeks of carryover decay, four weeks of clean read. Shorter and the four-week carryover contaminates the result.
• Primary metric: incremental new customers per market per week, not reported conversions, so the measurement does not depend on platform signal. Secondary: branded search volume, direct traffic, and 90-day repeat rate - because CTV may work by improving customer quality rather than volume.
• Read: difference-in-differences against the paired control, with the pre-period used to check parallel trends. If the dark markets show no detectable loss, the CFO is right and CTV should be cut. If they show a loss above 5%, the MMM's 1.9 is credible and MTA was misattributing.

Run a parallel spend-down test in the two most saturated social audiences to establish the true marginal return there, since that is where the reallocated money would go.

Do not cut CTV by 70% before the test reports. A 70% cut is a decision the test exists to make, and it is irreversible in the sense that restarting a CTV buy loses the frequency built over months.

Recommended shape for the quarter, holding total spend near $6 million:

• CTV: keep at about $1.2 million (the current 20% run-rate). Cutting now would also invalidate the holdout by changing national weight while the test runs; keep national weight flat and let the dark markets carry the variation.
• Paid social: hold, do not add. MMM shows saturation in specific audiences, and social is already 80% of spend. Additional dollars at the margin are the lowest-return dollars in the plan, which is the opposite of the CFO's assumption.
• Carve out $600,000 for the measurement program - geo holdout instrumentation, incrementality vendor, and a clean-room conversion-lift study to recover the 30% of signal lost to privacy changes. This is not overhead; without it, every future allocation argument repeats this one.
• Search: protect. It is the closest thing to a demand-capture channel and should not be raided.

Guardrails: (1) blended CAC must fall below $50 by quarter end or the mix discussion escalates regardless of test results; (2) no channel's budget moves more than 25% in a single quarter; (3) the CFO pre-commits to the test's decision rule before it starts, so the result is not relitigated.

The core argument: the CFO is right that CAC is broken. The evidence does not yet say which channel broke it, and the cheapest way to find out is one quarter of disciplined testing rather than a permanent reallocation based on a model that cannot see the top of the funnel.