Taylor Holiday

IROAS Uses Dual Factors, Campaign Ranks Vary

Thanks @PhilKiel for engaging seriously with the CPMr report. A thoughtful critique is beneficial to all research. But, two pieces of the thread rest on methodology misreads, and most of the "limitations you surface" are limitations we published ourselves. Here's how I see it. The central claim, that every campaign inside a brand gets the same multiplier so iROAS rank = ROAS rank, isn't how the analysis worked. Our Blended iROAS applies two factors, not one: an ACQ factor to new customer revenue and an RTN factor to returning customer revenue. Every campaign has a different new vs returning mix, so every campaign gets a unique effective factor. A 95% new customer campaign and a 70% new customer campaign in the same brand do not share a multiplier, and their iROAS ranks can, and do, diverge from ROAS ranks. Section 01 walks through this explicitly. The Section 08 line you quote ("iROAS conclusions are identical to ROAS") refers to a sensitivity cut, not the headline methodology. Second: the thread treats ρ(CPMr, incrementality factor) and ρ(CPMr, iROAS) as the same signal. They aren't. The cross-brand ρ = −0.62 says low-CPMr brands have higher factors, meaning Meta under-reports their revenue more. It does not say they have higher iROAS. A higher factor on a lower attributed ROAS can net flat against a lower factor on a higher attributed ROAS. Calling the cross-brand finding "the opposite direction from the headline" collapses two different dependent variables into one. Third, the limitations you surface are, with one exception, limitations we published: • p = 0.35 on the within-brand finding, Section 02 • Flat quartile dose-response, Section 04 • CPMr ↔ Frequency = +0.86, Section 07 • Win-rate collapse under spend thresholds, Section 08 • Account-level factors cannot resolve campaign-level incrementality, Sections 05 and 09 The report is self-critical by design. Reframing its own caveats as hidden flaws is a misread of the document, not a discovery about it. What I'd concede: the headline is built for clicks not nuance, and that compression makes the cross-brand finding easier to talk past. The Section 09 "Bottom Line" is the honest reading and I'd stand behind it. What's missing from the critique is data. Taikun's Q1 observation, 30.5% of spend beyond the 2nd view and 2 to 3x incremental CPA in high-frequency accounts, is directionally interesting, but it's an assertion, not a study. No sample, no method, no release. A critique of a dataset is sharpest when it brings one. I'd love to see yours, although I fear you already have identity wrapped into a default position. Where we agree, loudly: the next step is campaign tier lift testing. If you have completed this study I would also love to see it. Section 09 names it, and we're designing the holdouts now. I actually have no dog in this fight. Happy to go wherever the data leads.