-1.png?width=1132&height=292&name=Asset%201@2x%20(12)-1.png "Odins.ai Logo")
By the Odins team · Updated October 2026 · About 8 minutes
Radio and podcast advertising have no click. The usual workarounds (a promo code read out by the host, a vanity URL, a "how did you hear about us" survey) catch a fraction of the people the ad moved, and none of the ones who searched for the brand a week later. Platform lift studies are surveys of recall, not sales. Marketing mix modelling measures audio the same way it measures everything else: weekly spend against weekly sales, with the rest of the mix in the model at the same time, so the ROI of radio and podcasts reads on the same scale as Meta and Google. This guide covers the setup, the data, and the one test that settles the question when the data cannot.
The short answer
Put radio in the model as its own channel, with spend dated by air date from the agency's spot lists. Put podcasts in as their own channel when the spend is big enough and runs often enough to vary; until then, group them with radio as one audio channel and say so. Read promo codes and vanity URLs as a floor on the effect, never as the measure. And when the model's range on audio stays wide, run a pulse test: audio switches off cleanly, which makes it one of the cheapest channels to test.
Why audio needs a model rather than a code
A promo code measures the people who heard the ad, remembered the code and typed it in. Everyone else who heard it and bought through a search, the app or a store is credited to whatever they touched last, usually brand search. The code understates audio and overstates search, in the same report, and the error grows with the brand: the better known you are, the fewer people need the code.
An MMM does not need the code. It looks at the weeks when audio ran at different levels and asks whether total sales moved with it, after seasonality, price and the other channels are accounted for. The answer is a response curve for audio with a confidence range, on the same yardstick as every other channel. The full guide to marketing mix modelling covers how that works; this guide is about audio specifically.
Get the data in
Audio data is held by the agency and the platforms, and the request is worth making properly: spend per station or show, dated by when the spots aired, going back as far as the sales history, with gross versus net, agency fee and production made explicit. Impressions and listener numbers help set the channel's curve before training; the model itself runs on spend.
| Channel | Where the spend comes from | Date it by | Exposure data | In the model as |
|---|---|---|---|---|
| Radio | Agency reports and spot lists, per station | Air date, weekly | Impressions or listener reach when the stations report it | Its own channel |
| Podcasts, host-read | Sponsorship agreements and invoices | Episode publish dates, spread over the sponsorship period | Downloads per episode when the show reports them | Own channel if big enough; otherwise joins audio |
| Podcasts and streaming audio, inserted | Platform exports, weekly | Air date | Impressions | Joins the podcast or audio channel |
Host-read sponsorships need one extra step. The invoice is one lump; the exposure is an episode a week for eight or twelve weeks. Spread the fee over the episodes before it goes into the model, so the effect is read in the weeks people heard it rather than in the week finance paid for it.
At Odins, audio comes in through a pipeline from the agency, with the digital channels through platform connections, and everything lands as weekly spend in one structure. The data guide sets out the full agency request.
Podcasts: small and lumpy
Podcast spend has a shape the model finds hard: a few campaigns a year, nothing in between, and often less than two percent of the media budget. A channel like that rarely earns a response curve of its own. Model it alone and the result is mostly the prior you gave it, dressed up as an estimate.
The working answer is to group podcasts with radio as one audio channel until the podcast spend is big enough and regular enough to carry its own curve. Put the grouping in the model notes so nobody reads the audio result as radio only. The day podcasts grow, split them out and retrain. If the business wants a podcast number sooner than the data allows, a concentrated test is the way to get one.
When the data cannot settle it: a pulse test
Some audio histories have never varied enough to read. The station buy has been the same most weeks for years, so the model cannot tell audio from the baseline. The fix is not more years of the same data. It is a planned change.
A pulse test changes one channel on purpose and only that channel: a fixed window of four to eight weeks, some weeks off, some at double, with everything else held steady. Before it runs, the model says how big the pulses have to be to show up in sales and what the test costs against the plan. Afterwards, the result calibrates the model's estimate for audio, and the comparison with every other channel stays on one scale. Radio and inserted audio are good candidates because they switch off cleanly and come back without a relearning period, which paid search and Performance Max do not.
Codes and vanity URLs: a floor, not the measure
Keep the promo code. It is free evidence, and it is useful as a lower bound: the model's estimate for audio should never fall below what the codes alone brought in. If it does, something is wrong with the setup. But the gap between the code count and the model's estimate is the whole point of modelling audio, so do not let the code be the number in the budget meeting.
Reading audio ROI next to digital
The model gives two numbers per channel, and for audio the difference matters. The average return says what each unit spent on radio earned over the period. The marginal return says what the next unit would earn at today's level. Radio often has a good average and plenty of headroom in markets where it has been bought at the same level for years, which makes it a candidate for the next unit of budget rather than the first cut.
Audio also pays back over a longer horizon than search. Part of what a radio campaign does is build brand, and those sales arrive over months. A model that counts only the next three weeks will undercount audio against search every time. Odins splits each channel's effect into sales within weeks and a share that builds brand and pays out later, so audio and search can be compared on the horizon the business plans on. The method is explained in brand effects in the model.
Questions to ask a vendor
- How does radio spend get into the model, who does it every week, and is it dated by air date?
- At what spend does a podcast channel earn its own curve, and what happens below that?
- How do you handle a host-read sponsorship paid in one invoice?
- Can the model design a pulse test for audio and use the result to calibrate itself?
- Over what horizon is audio compared with search, and does the longer payback count?
Odins builds and runs the model, collects offline and digital spend every week, and delivers a monthly recommendation with a confidence range per channel. Companies like CDON, Nettbil, Hyre and Megaflis run their marketing budgets on it.
Frequently asked questions
Can MMM measure podcast advertising with a small budget?
Only as part of a bigger audio channel, until the spend is large and regular enough to vary. A podcast line at one percent of the budget, run three times a year, cannot carry its own estimate. Group it, say so, and split it out when it grows. A concentrated burst designed as a test is the fastest route to a podcast-specific number.
How do we get radio spend into the model?
From the agency's spot lists and invoices, per station and air date, as far back as the sales history allows. Ask for gross versus net and agency fee explicitly. Odins takes those files through a pipeline so the weekly feed runs without the marketing team touching a spreadsheet.
Do we need listener numbers?
No. The model runs on spend. Impressions and reach help set the audio curve before training and are worth collecting when the stations report them, but a model on spend alone is a sound model.
How quickly does radio show an effect in sales?
The direct effect lands within days; the model estimates how long it carries over rather than assuming it. The brand part of the effect plays out over months, which is why the horizon matters when audio is compared with search.
Is a promo code enough to measure a podcast campaign?
It is enough to prove the campaign did something, and it is a useful floor. It is not enough to size the effect or compare it with other channels, because most of the people the campaign moved never use the code.
Start with the complete guide to marketing mix modelling, see how offline media data gets into the model, or book a demo and bring your radio and podcast plans.