2 February 2026 · 9 min

When modeled conversions mislead

The model was not “wrong” in a textbook sense. It was trained on a quieter media mix than the one you ran in week three of January.

Laptop showing financial figures

A grocery app we later taught in Cohort Studio had a modeled conversion line that hugged SKAN through December. In January they added a brand burst on connected TV. The model kept assigning mobile “credit” as if the burst were a small paid-social test. The holdout in two northern cities told a different story: incremental first orders barely moved.

Nobody had lied. The vendor’s model did what it was asked: impute missing identifiers from last quarter’s mix. Privacy-Safe Mobile Measurement includes the discipline of saying when that imputation is no longer the number you print in a GB board pack.

What we now ask teams to publish

Three columns: SKAN (or Android aggregate) counts, modeled estimates, and the holdout or geo split. If column two and column three disagree by more than the band you wrote down in week one, column two is annotated or withheld. The annotation is a sentence, not a colour in a chart.

The uncomfortable meeting

Paid media argued that withholding the model hid their work. Finance argued that printing it hid the burst. The compromise was a page titled “What we will not forecast this month.” It looked incomplete. It was the first pack the CFO signed without a side conversation.

Helen Rowe uses this episode in module four of SKAN Signal Craft. It is not a morality tale. It is a reminder that models have a media-mix expiry date.