← All case studies
Case Study

How incremental are podcast ads? A 3-month DTC case study

Brenden DelaRua · October 10, 2025
How incremental are podcast ads? A 3-month DTC case study

Probably more than the promo codes showed. Across three windows from February to April 2025, our models estimated that a DTC brand's podcast sponsorships drove $175,890 in incremental revenue on $75,520 of spend. That's a blended incremental ROAS (iROAS) of 2.33. The brand's promo-code reporting put podcast ROAS at 1.0 to 1.5. February and March were stronger. April was close to break-even.

iROAS here is model-estimated incremental revenue divided by spend: $175,890 / $75,520 = 2.33. Incremental revenue is what the models estimate would not have happened without podcasts. Nobody counted those sales one by one.

So the honest headline is narrower than "podcasts work." The models suggest podcasts were doing more than promo codes credited them for. For this brand, that made a stronger test worth running before any big budget move.

Table of contents

Were promo codes enough to judge podcasts?

Probably not. Promo codes only count buyers who remember the code and type it in. A listener who hears an ad, then searches the brand or buys weeks later, gets credited to search, direct, or nothing. Codes can also pick up buyers who would have purchased anyway. So 1.0 to 1.5 tells you what the codes captured. It doesn't tell you what podcasts caused.

Our client is an ecommerce brand that runs most of its paid media on Meta and Google. In 2024 and 2025 it added podcast sponsorships on shows including Meidas Touch, Some More News, Steve-O's Wild Ride, and The H3 Podcast. Its promo-code reporting put podcast ROAS at 1.0 to 1.5, which made the channel look marginal.

We wanted to know whether that number was missing something, and how sure we could be about it.

How do you measure podcast incrementality without a holdout?

You model it. These were national host-read placements, so there was no clean way to hold out a region. Instead, models learned how daily revenue moved with Meta, Google, and TikTok spend, promotions, seasonality, and organic demand. Then they forecast revenue without the podcast effect. The gap between that forecast and actual revenue is the estimated incremental revenue.

This wasn't a randomized experiment. We used observational time-series models to estimate what revenue might have looked like without podcasts. That estimate depends on the quality of the counterfactual, and on whether the models adequately account for everything else affecting sales. Strong historical fit helps, but it doesn't establish causality on its own.

Several shows also aired on the same days, so we measured all podcast spend together, as one program.

The analysis modeled daily revenue and daily orders, using 120 days of training data before the test windows. Podcasts went in as daily spend plus a yes-or-no flag for air days. We compared results across Bayesian structural time series (the approach behind Brodersen et al., 2015), temporal regression, and synthetic control approaches.

The analysis also reported these diagnostics:

  • Fit error (MAPE). Mean absolute percentage error is how far, on an average day, the forecast missed actual revenue in the training period. Lower is better. It describes daily fit in training, which is a different thing from how uncertain a cumulative lift estimate is.
  • Posterior predictive checks. These test whether the model can reproduce the revenue data it learned from. A model that can't isn't worth forecasting with.
  • Placebo checks. We shifted podcast activity to alternative dates and re-ran the models, and reported no comparable lift.
  • Significance level. The model output marks each window as significant or not at 90%, 95%, and 99%. We used 90%, a looser bar than the 95% many analysts default to. We report all three so you can hold the results to the stricter one.

These checks can help flag model sensitivity and implausible results. They don't eliminate the possibility of unobserved confounding. The original analysis also didn't record enough detail for an independent review: no per-model estimates, no placebo distribution, and no description of the test behind the significance flags. So read the methods as what we ran, and the screenshots below as what we can show.

What did the podcast incrementality results show?

The model flagged all three windows as significant at 90%. None reached 95%. February came in at 2.89 iROAS, March at 2.64, and April at 1.21. Combined, $175,890 of estimated incremental revenue on $75,520 of spend works out to 2.33. In each window, podcasts added 2% to 10% over what the forecast expected.

WindowSpendIncremental revenueiROASIncremental ordersLift vs forecastPre-period error (MAPE)
February 2025$27,160$78,4202.8960010.3%16.57%
March 2025$27,110$71,6802.645436.19%8.85%
April 2025$21,250$25,7901.211952.35%6.38%
Total$75,520$175,8902.331,338n/an/a

Lift vs forecast is incremental revenue as a share of what the model expected without podcasts. Each monthly figure comes from the model outputs shown in the screenshots below. The total row adds them up.

Podcast incremental ROAS by measurement windowEstimated iROAS was 2.89 in February, 2.64 in March, 1.21 in April and 2.33 combined. The brand's promo-code reporting put podcast ROAS at 1.0 to 1.5.0.0x1.0x2.0x3.0x4.0xPromo-code ROAS range (1.0x to 1.5x)2.89xFebruary2.64xMarch1.21xApril2.33xCombined
Model-estimated incremental ROAS per window. The combined bar is total incremental revenue divided by total spend across all three windows. Source: Stella model outputs for one anonymized DTC client, February to April 2025. Promo-code ROAS as reported by the client.

2.33 against 1.0 to 1.5 is a big gap. It's also built from three windows that don't deserve equal weight.

WindowWhat looks goodWhat a skeptic should push onHow we read it
FebruaryBiggest lift, 2.89 iROASHighest pre-period error (16.57%)Encouraging, precision unknown
MarchRunning total climbed steadily, 8.85% pre-period errorOnly flagged at 90%, not 95%Most consistent modeled signal
AprilBest fit (6.38%)Running total below zero until the final weekNot a stable read either way

What are the limits of these results?

Each window has its own problem. January was confounded and thrown out. February had the weakest training fit. April only turned positive in its last week. And this is one brand over three months, so it says little about podcasts in general.

February: highest return, weakest fit

February produced the highest estimated return at 2.89 iROAS, alongside the highest reported pre-period MAPE at 16.57%. That raises questions about forecast reliability. But training-period error alone doesn't tell us how uncertain the cumulative incremental revenue estimate is. February's result is encouraging, though the available diagnostics aren't enough to establish how precise it is.

February 2025 podcast incrementality model output: $78.42K incremental revenue, 2.89 iROAS, 16.57% MAPE, significant at 90% but not 95%
February model summary.
February 2025 actual versus predicted daily revenue and cumulative incremental revenue from podcast ads, rising to about $78K
February actual vs predicted revenue, and cumulative incremental revenue.

March: a more consistent modeled signal

March's model estimated $71,680 in incremental revenue, equivalent to 2.64 iROAS. Its cumulative estimate rose fairly steadily through the month, with a reported pre-period MAPE of 8.85%. The result was flagged as significant at 90%, but not 95%. The pattern is encouraging, though neither the trajectory nor the training fit establishes the causal effect on its own.

March 2025 podcast incrementality model output: $71.68K incremental revenue, 2.64 iROAS, 8.85% MAPE, significant at 90%
March model summary.
March 2025 model fit, actual versus predicted daily revenue, and cumulative incremental revenue from podcast ads
March model fit, actual vs predicted revenue, and cumulative incremental revenue.

April: unstable, and we don't know why

April's modeled return fell to 1.21 iROAS, compared with 2.89 in February and 2.64 in March. The running total of incremental revenue sat below zero for most of the month and only turned positive in the final week, as the chart below shows. The reported significance applies to the complete measurement window, not to each day or week. Because the estimate changed direction late in the period, it shouldn't be read as evidence of consistently positive performance through April.

The available analysis can't establish why the effect differed across months. That's exactly why one month of modeled performance should not be treated as a stable estimate of what a channel does.

April 2025 podcast incrementality model output: $25.79K incremental revenue, 1.21 iROAS, 6.38% MAPE, significant at 90%
April model summary.
April 2025 model fit and cumulative incremental revenue from podcast ads, below zero for most of the month before turning positive in the last week
April model fit, actual vs predicted revenue, and cumulative incremental revenue.

January: confounded, so excluded

January's model estimated an 8.42 iROAS ($109,290 of incremental revenue on $12,980 of spend). But a major promotion and overlapping sponsorships made it hard to isolate the podcast contribution. We excluded this window from the reported results rather than treat a potentially confounded estimate as evidence of exceptional performance. It's not in any total on this page.

January 2025 podcast model output, excluded from results because a large promotion overlapped the window
January model summary. Excluded from all totals.

What should a brand do with a 2.33 iROAS?

Treat it as a reason to measure better before spending more. The analysis suggested podcasts may be contributing more revenue than promo-code attribution alone indicated. For this brand, that made a stronger measurement design worth pursuing before any major budget decision.

A few things the 2.33 doesn't tell you. It's the average return on all podcast spend in those windows. The next dollar usually earns less, which is why average and marginal iROAS differ. It's revenue, so whether podcasts made money depends on gross margin. And since shows overlapped, it can't rank individual shows.

The next step we recommended was a design with sharper contrast than steady spend: after a sustained run, go dark for a month, or pause for three months and then run one heavy month. Bigger swings in spend give the models a stronger signal to work with. Those designs still are not randomized, so they would sharpen the estimate without settling it. Where a randomized holdout is possible, it's the stronger design, and Stella runs both kinds of incrementality tests.

Frequently asked questions

Can you measure podcast ad incrementality without a geo holdout?

Yes, with more uncertainty. When a host-read ad airs nationally there's no region to hold out, so you model what revenue might have looked like without podcasts and compare it to what happened. The result depends on how well the models separate podcasts from everything else driving sales.

Why not just measure podcasts with promo codes?

Promo codes only count buyers who remember the code and type it in. Listeners who hear an ad, then search the brand or buy weeks later, never show up. Codes can also pick up buyers who would have purchased anyway. In this case codes credited podcasts with 1.0 to 1.5 ROAS, while the models estimated 2.33.

Is 90% confidence enough to make a budget decision?

It depends on the decision. A result flagged at 90% is weaker statistical evidence than one that meets a 95% threshold, and neither threshold rules out model bias. We'd not use these results alone to justify a major budget increase. The right next step depends on contribution margin, the cost of further testing, and how much uncertainty the brand can tolerate.

Why was April's podcast result so much lower?

The analysis can't say. April's running total of incremental revenue sat below zero for most of the month and only turned positive in the final week. On its own, one month like that doesn't mean podcasts stopped working.

Does a 2.33 iROAS mean podcasts are profitable?

Not on its own. iROAS is incremental revenue divided by spend, so it ignores margin. Whether 2.33 makes money depends on gross margin and repeat purchase. It's also an average across all podcast spend, so it doesn't tell you the return on the next dollar or on any single show.

Which podcast show drove the most incremental revenue?

This study can't say. Several shows often aired on the same day, so their effects overlap in the data. The models measured the combined effect of all podcast spend. Comparing shows needs either separated air dates or a per-show signal such as promo codes, used for ranking rather than sizing.

Promo codes miss a lot of what sells outside a direct-response path. One month of modeling can mislead too. If either is the main evidence for or against your podcast budget, book a demo and we'll show you how we'd measure it on your data.