← All case studies
Case Study

Tushy held linear TV out of its markets for 28 days. Revenue dropped where it went dark.

Brenden DelaRua · September 24, 2026
Tushy held linear TV out of its markets for 28 days. Revenue dropped where it went dark.

Tushy and Stella · Linear TV · Inverse geo holdout · 28-day test

One inverse geo holdout, read across two storefronts. With TV removed, Amazon revenue came in an estimated 9.34% below its modeled counterfactual, about $47K over the test. Shopify showed a 9.17% estimated reduction in the same markets and window. Both effects met Stella's 90% significance threshold.

Table of contents

The Customer

Tushy is a direct to consumer bidet brand known for blunt marketing, selling modern bidet attachments, electric bidet seats, and related bathroom products through its own Shopify store and Amazon. Alongside its digital channels it runs linear TV, a channel with no click to trace.

The Problem

With no click to trace, TV falls out of last click attribution. When a Tushy ad sends someone to Amazon or Shopify days later, the storefront logs the sale and search takes the credit, while the ad that started the purchase leaves nothing either platform can read. That makes TV impossible to judge on the same terms as channels with tracked clicks and conversions. So Tushy asked the blunt version: what does revenue do when TV comes out?

The Test

Tushy ran a 28 day inverse geo holdout in Stella from February 3 to March 3, 2026. Linear TV kept running across the broader market and came out of a set of holdout regions chosen by a location analysis that matched them to that market. Stella modeled what those regions would have earned with TV still on, then compared it to what they actually did once TV was off. The difference is the estimated causal effect.

The one intervention was read against two outcomes, Amazon revenue and Shopify revenue, in the same markets and window. They are two reads of one experiment, not two experiments. The models tracked each revenue series closely before the holdout, R² 0.99 and 7.55% MAPE on Amazon, R² 0.98 and 9.11% on Shopify. Those are fit diagnostics; the causal read comes from pulling TV, not from the fit.

The Result

Amazon

Amazon: Test vs Control
Linear TV inverse geo holdout · Amazon revenue · February 3 to March 3, 2026
i
-47.25K
iRevenue
The estimated additional revenue generated by your campaign, beyond what would have occurred without it.
i
-9.34%
Percent Lift
The percentage increase in your target metric attributable to your campaign, calculated as incremental impact divided by the predicted baseline.
i
15.05K
Spend
Total advertising spend during the test period. For inverse holdout tests, this shows synthetic spend, the estimated amount that would have been spent in test regions if ads were kept live, calculated from historical spend ratios.
i
-501.9
iOrders
The estimated number of additional orders generated by your campaign, beyond what would have occurred without it.
i
29.98
iCPO
Incremental Cost Per Order, how much you spent for each additional order generated by your campaign.
i
-3.14 [-5.95, -0.36]
iROAS
Incremental Return On Ad Spend, the additional revenue generated for every dollar spent on your campaign.
i
7.55
MAPE (%)
Mean Absolute Percentage Error, measures model prediction accuracy. Lower is better. Under 10% is excellent, 10-20% is good, above 20% may need more data.
i
90%
Stat. Sig.
Statistical Significance, the confidence level that your results are real and not due to chance. 99% is very strong, 95% is strong, 90% is moderate.
i
0.99
R²
R-Squared, measures how well the model fits your historical data. Closer to 1.0 is better. Above 0.8 is good, above 0.9 is excellent.
Amazon holdout scorecard. Because this is an inverse test, the values read negative: TV was removed and revenue fell with it. The magnitude is the incremental revenue TV was producing, about $47K over 28 days at 9.34% lift, significant at 90%. iROAS is measured against Stella's $15.05K computational test basis, not Tushy's aired TV spend, so it is not a media ROAS. Tile figures are exact; figures in the write up are rounded.

Pulling TV cut Amazon revenue an estimated 9.34%, about $47K and 502 orders over the window, at 90% significance. Stella puts that at a 3.14x return on its $15K test basis, with a 90% interval of 0.36x to 5.95x. It is clear that TV drove incremental Amazon revenue, and much less clear how large the return was. The 3.14x is not a media ROAS: the $15K is Stella's computational test basis, not Tushy's aired TV spend.

Shopify

Shopify: Test vs Control
Linear TV inverse geo holdout · Shopify revenue · February 3 to March 3, 2026
i
-43.76K
iRevenue
The estimated additional revenue generated by your campaign, beyond what would have occurred without it.
i
-9.17%
Percent Lift
The percentage increase in your target metric attributable to your campaign, calculated as incremental impact divided by the predicted baseline.
i
15.05K
Spend
Total advertising spend during the test period. For inverse holdout tests, this shows synthetic spend, the estimated amount that would have been spent in test regions if ads were kept live, calculated from historical spend ratios.
i
-298.37
iOrders
The estimated number of additional orders generated by your campaign, beyond what would have occurred without it.
i
50.44
iCPO
Incremental Cost Per Order, how much you spent for each additional order generated by your campaign.
i
-2.91 [-5.82, -0.02]
iROAS
Incremental Return On Ad Spend, the additional revenue generated for every dollar spent on your campaign.
i
9.11
MAPE (%)
Mean Absolute Percentage Error, measures model prediction accuracy. Lower is better. Under 10% is excellent, 10-20% is good, above 20% may need more data.
i
90%
Stat. Sig.
Statistical Significance, the confidence level that your results are real and not due to chance. 99% is very strong, 95% is strong, 90% is moderate.
i
0.98
R²
R-Squared, measures how well the model fits your historical data. Closer to 1.0 is better. Above 0.8 is good, above 0.9 is excellent.
Shopify holdout scorecard, the same inverse read on the same $15.05K test basis: about $44K and 298 orders over the window at 90% significance. The 90% interval runs to a lower bound near zero, so read Shopify as support for the direction of the Amazon result, not a firm number of its own.

The same pull cut Shopify revenue an estimated 9.17%, about $44K and 298 orders, again at 90%. Stella reports a 2.91x effect on the same $15K basis, interval 0.02x to 5.82x. That lower bound almost on zero makes the Shopify number far less certain; the test reads direction more clearly than size. And because it came from the same holdout, Shopify moves with Amazon rather than confirming it on its own.

The Outcome

The holdout landed the same way on both storefronts. Revenue fell in the markets TV left, the revenue no storefront report would have traced back to TV. The test shows TV was incremental here. It does not fix the return on the next dollar, since the intervals are wide, Shopify especially, so a second holdout at a higher spend level is the way to map that.

Cumulative incremental Amazon revenue for the Tushy linear TV inverse geo holdout, falling steadily below zero from February 3 to March 2 2026 and ending near -$47K with a widening 90% confidence band
Cumulative effect of removing TV on Amazon revenue, with the 90% confidence band. The line runs negative because this is an inverse test: it is the revenue that came out with TV, building steadily to about $47K across the 28 days, a sustained effect rather than a one time spike. The band widens over the window, the honest read that the cumulative size is less certain the further it runs.
Cumulative incremental Shopify revenue for the Tushy linear TV inverse geo holdout, ending near -$44K with a wide 90% confidence band that reaches back above zero
Cumulative effect of removing TV on Shopify revenue, with the 90% confidence band, the same inverse read ending near $44K. The band is wide and reaches back above zero, the marginal interval on this channel made visible: the direction holds, but the size is uncertain enough that Shopify supports the Amazon result rather than standing on its own.

Working with the team at Stella, we set out to figure out two things: (1) how can we measure the true impact of a channel like Linear TV beyond limited, clicks-based models, and (2) how can we quantify the lift across key sales channels like DTC and Amazon. The team was an amazing partner in both the test design and analysis, providing us with the clarity and confidence to not only adjust our media mix, but also reshape the way we think about Linear TV’s impact on our business holistically.

Jerel Blades, Head of Growth at TUSHY

Method: 28 day inverse geo holdout in Stella, February 3 to March 3, 2026, using custom test groups. Linear TV came out of a set of holdout regions selected by location match while it kept running across the broader market. One intervention was read against two outcomes, Amazon and Shopify revenue, at a 90% confidence threshold. Model fit before the holdout: Amazon R² 0.99, MAPE 7.55%; Shopify R² 0.98, MAPE 9.11%. Revenue and order counts are rounded; lift, intervals, model fit, and significance are shown as reported. The test basis is a computational basis, not Tushy's aired linear TV spend, so return figures are not media ROAS.

A number you can move budget on beats a number you have to trust. See how Stella measures incrementality, or book a demo.