Experiment summary
- Proper Cloth’s Snowplow tracking made YouTube look expensive, while Google’s view through reporting made it look productive.
- Measurement vendor BlueAlpha ran a five week geo holdout. YouTube Demand Gen ran in Massachusetts, California and Pennsylvania, with a Bayesian synthetic control built from more than 40 other states.
- It reported a 12.7% cumulative sales lift, about $180,000 in incremental revenue (95% credible interval roughly $98,000 to $255,000) and a 2.43 return on ad spend.
- No lift appeared in week one, and the effect became statistically significant by week three.
- BlueAlpha published the case itself, and spend, state selection, pre period fit and the full model were not disclosed.
Proper Cloth’s tracking systems disagreed about YouTube. A five-week geographic test offered a revenue estimate and a clearer next experiment.
Two dashboards, one campaign, two verdicts. Proper Cloth’s Snowplow tracking could see sessions and purchases, and by that measure YouTube looked expensive. Google Ads counted view through conversions instead, which credit a sale to anyone who watched an ad and bought later without clicking, and by that measure the channel looked productive. Neither view could answer the question that mattered most. How much revenue would disappear if the ads simply stopped running?
The custom menswear company needed that answer before committing more budget to upper funnel video, the kind of advertising that builds awareness well before anyone is ready to buy. According to a case study published by measurement vendor BlueAlpha on December 15, 2025, the team used a five week regional test to estimate the incremental revenue caused by a YouTube Demand Gen campaign.
The Proper Cloth geo experiment is useful because it shows what changes when attribution becomes a counterfactual problem. A counterfactual is simply the sales you would have seen if the campaign had never run. The team stopped asking which tracking system deserved more trust and asked what sales would probably have looked like without the campaign.
Two attribution systems gave two different answers
Proper Cloth was founded in 2008 to make custom fit clothing easier to buy online. The company sells made to order shirts, suits and other menswear through its website and New York showrooms. The mix creates a familiar measurement problem. A customer may see a video, return through another channel and buy later, leaving different systems to claim different parts of the journey.
BlueAlpha says Proper Cloth’s click based analytics suggested poor YouTube performance, while Google’s view through reporting suggested meaningful value. Post purchase surveys added another signal but still could not isolate the channel’s incremental effect. The disagreement was far bigger than a small reporting discrepancy. It left the team without a defensible basis for scaling or cutting the spend.
The documented question was therefore narrower than whether video advertising “worked.” Proper Cloth wanted to know whether the campaign produced sales that would have stayed unsold without it. Marketers call that incrementality, and measuring it requires a comparison with an untreated condition.
Three states became the test group
BlueAlpha designed a geo holdout setup with Massachusetts, California and Pennsylvania as treatment markets. The YouTube campaign ran in those three states. More than 40 other states supplied data for a Bayesian synthetic control, a modelled comparison built to estimate what sales in the treatment markets would have been without the campaign. Picture a blended stand in made from the quiet states, weighted so its sales history tracks the three test states before the ads start.
The media setup was deliberately bounded. Proper Cloth ran 15 second in stream creative without a product feed and optimised the campaign for purchases. Targeting included in market, custom intent, lookalike and remarketing audiences. The primary outcome was total sales revenue from Proper Cloth’s transaction data, rather than conversions reported by the advertising platform.
This distinction is the core of the YouTube incrementality test. The ad platform still delivered the intervention, and it no longer graded its effect. The experiment compared observed sales in the treatment states with a modelled counterfactual assembled from untreated states.
The architecture also makes the test harder to copy than a simple regional split. A useful Bayesian synthetic control depends on stable historical relationships between markets, enough untreated regions, clean transaction data and no large concurrent change that affects treatment markets differently. BlueAlpha does not disclose the state selection process, pre period fit, campaign spend, traffic allocation or full model specification.
The signal took three weeks to appear
The weekly pattern helps explain the final number. BlueAlpha reports no measurable lift in week one. An early signal appeared in week two. By week three, the vendor says the effect had become statistically significant, and it persisted through weeks four and five.
At the end of five weeks, BlueAlpha reported a cumulative sales lift of 12.7 percent, corresponding to about $180,000 in incremental revenue. Its 95 percent credible interval ran from roughly $98,000 to $255,000, which is the range the true figure most likely sits in under the model’s assumptions. The vendor also reported a 2.43 return on ad spend, meaning about $2.43 of incremental revenue per dollar spent, with a range from 1.3 to 3.47, and a 0.36 percent modelled probability of a loss.
Those figures are vendor published estimates, and nobody outside the project has verified them. The wide revenue interval is particularly important. The model supports a positive effect under its assumptions, yet it leaves the outcome short of a single precise dollar value.
The delay before a detectable signal also illustrates a practical risk in high tempo experimentation. A team reviewing only the first week could have stopped the test before the response emerged. A team watching platform conversions alone could have scaled before it had causal evidence. The experiment needed enough time for purchases to accumulate and for the counterfactual comparison to separate from normal variation.
The result changed the test roadmap
BlueAlpha reports that Proper Cloth decided to expand YouTube into its highest revenue markets with a budget two to three times larger. The company also planned a product feed Demand Gen test, a Performance Max holdout and a connected TV experiment. The sequence follows from the fact that the Proper Cloth geo experiment left some questions about video open. It reduced uncertainty enough to justify a larger test and a new set of comparisons.
Jeffrey Lai, Proper Cloth’s head of analytics and advertising, is the named practitioner in the case. His LinkedIn profile describes responsibility for new customer acquisition at the company. BlueAlpha also reports that Proper Cloth had set a goal for new customers to generate 42 percent of 2026 revenue. The case does not disclose how the five week experiment affected that target, customer quality, margin or repeat purchase behaviour.
Synthetic controls put the assumptions on the table
The lesson that travels is about design. Incompatible attribution systems are often a sign that the decision needs a different kind of test. Platform reporting and session analytics answer questions about observed journeys. Geo incrementality asks what would have happened without the intervention.
A Bayesian synthetic control can be useful when user level randomisation is unavailable or would cause contamination across channels. It can also create false confidence if the control markets fail to mirror the treated markets, if regional shocks break the historical relationship or if other changes occur during the test. The model leaves every assumption in place. It concentrates them in places an operator can inspect.
For a technical growth team, that suggests a more disciplined sequence. Define the business outcome in first party data. Choose a counterfactual that matches the way the intervention is delivered. Decide in advance how long the test must run and what uncertainty is acceptable. Then use the result to choose the next experiment and resist turning one estimate into a permanent channel verdict.
Conflicting dashboards also appear when consent rules thin out the data. RS Group’s AI attribution pilot in Italy shows how a rebuilt conversion signal can change what a paid search bidder learns.
LinkedIn’s causal targeting experiment tackled the next measurement question, identifying which audiences an ad could persuade rather than only reconciling which dashboard claimed the conversion.
The CAC diagnostic helps examine acquisition economics after the experiment: what the incremental customers cost, and how the answer changes with conversion and margin assumptions.
| Experiment detail | What the reporting shows |
|---|---|
| Change | BlueAlpha ran a five-week geo holdout with YouTube Demand Gen live in Massachusetts, California and Pennsylvania, and a Bayesian synthetic control built from more than 40 other states. |
| Reported result | Estimated 12.7% incremental sales lift, about $180,000 (95% credible interval roughly $98,000 to $255,000), with a 2.43 return on ad spend. |
| Evidence | Measurement vendor case study (BlueAlpha) naming a Proper Cloth practitioner, no independent audit |
| Limits | Vendor-published model estimate. Spend, state-selection method, pre-period fit, full model specification and concurrent regional changes were not disclosed. |
FAQ
How did a geo experiment help Proper Cloth?
Its click based analytics and Google’s view through reporting disagreed about YouTube, and neither could show how much revenue would disappear if the ads stopped. A geo holdout supplies a counterfactual instead of another attribution claim.
How was the test set up?
The campaign ran in Massachusetts, California and Pennsylvania for five weeks. More than 40 other states fed a Bayesian synthetic control, and the outcome was total sales revenue from Proper Cloth’s transaction data, with no reliance on platform reported conversions.
How big was the estimated lift?
BlueAlpha reported a 12.7% cumulative sales lift, about $180,000 in incremental revenue with a 95% credible interval of roughly $98,000 to $255,000, and a 2.43 return on ad spend with a range of 1.3 to 3.47.
How reliable is the estimate?
It is a vendor published model estimate with a wide interval. The case study does not disclose spend, state selection, pre period fit, the full Bayesian specification or concurrent regional changes, and nobody has independently audited it.
Sources and limits
- BlueAlpha’s Proper Cloth case study, published December 15, 2025
- Proper Cloth’s company history and business description
- Jeffrey Lai’s LinkedIn profile
BlueAlpha designed, ran and published the case. The report identifies only three treatment states and does not provide the spend, state selection method, pre period fit, complete Bayesian specification, traffic allocation or a record of concurrent regional changes. The reported lift and uncertainty should therefore be read as the vendor’s model based estimate.
Spotted an error? Contact the desk. Our corrections policy explains how we handle corrections and updates.
Read next
What changes when AI influences marketing economics? Read the wider analysis of attribution and AI-driven acquisition.
Find the reported method, result and limitations of other experiments in the Growth Signal Index in IC Lab.
