Linear TV, connected TV and YouTube are usually bought by the same team, often through the same agency, and reported in three different currencies: ratings for linear, impressions and completed views for connected TV, views and view-through conversions for YouTube. None of those is sales, and none is comparable with the others. Marketing mix modelling puts the three on one yardstick: incremental sales per unit of spend, and what the next unit would add. This guide covers how to set the three channels up so the model can tell them apart, which data each needs, and how to read the comparison once it is there.
Model linear TV, connected TV and YouTube as three separate channels wherever each has enough spend, with weekly spend dated by air date, carry-over for the weeks after a flight, and a saturation curve per channel. Then compare them on the same measure: the return on the last unit spent, not the average. Three things make the comparison trustworthy:
YouTube reports view-through conversions: people who saw an ad and later bought, whether or not the ad made a difference. Connected TV platforms report households reached and, with the right setup, matched conversions on the same device graph. Linear TV reports ratings and reach, and no sales at all. Each number is correct inside its own wall, and none of them answers the question the budget meeting asks, which is what each channel added to sales that would not have happened anyway.
MMM answers that question from the outside. It relates weekly sales to weekly spend across every channel at once, after separating out seasonality, price, promotions and whatever else moves sales. It does not need a click or a matched household, which is why it is the method that can hold linear TV and YouTube to the same standard. The full guide to marketing mix modelling explains the mechanics; this guide is about the three video channels specifically.
The table is the setup we start from. The common thread is that spend is dated by when the ad aired, not when the invoice arrived, and that exposure data is collected where it exists.
| Channel | Spend data | Exposure data | Carry-over | What to watch |
|---|---|---|---|---|
| Linear TV | Broadcaster and agency reports, per spot or campaign, dated by air date | Ratings or gross rating points per week | Weeks; the longest of the three | Flights that always coincide with out-of-home or radio; a strong brand share of the effect |
| Connected TV | Platform or agency reports, weekly | Impressions and completed views | Shorter than linear | Small spend in the first year; may need to join linear TV until it is big enough for its own curve |
| YouTube | Google Ads, by campaign type | Impressions, views | Short; the shortest of the three | Separate from search and Performance Max; brand formats and performance formats behave differently |
Two of the three need a decision about grouping. Connected TV spend is often small in the first year and runs in a few bursts. A channel under about two percent of spend rarely earns a curve of its own, so it joins linear TV until the volume justifies splitting it out, with a note in the model so nobody reads the TV result as linear only. YouTube is the opposite problem: it comes through the Google Ads connection mixed with search and Performance Max, and has to be split out by campaign type before it can be read as video.
Odins collects linear TV through a pipeline from the agency or broadcaster, and the digital channels through the platform connections, so all three land as weekly spend in the same structure. Our guide to what data marketing mix modelling needs covers the rest of the data side.
A model separates two channels by how they move relative to each other. If connected TV only ever runs in the weeks linear TV runs, at a fixed share of the budget, the model sees one channel and reports one combined effect. The data cannot fix that; only the plan can.
In practice three things give the model what it needs:
When a test has run, its result calibrates the model rather than replacing it. The model's estimate for that channel is pulled toward what the test showed, and the comparison across all three channels stays on one scale.
The question "how do we compare linear TV, connected TV and YouTube" has two answers, and they are often different.
Where the long tail goes
Linear TV's sales do not all arrive in the weeks after the flight. A model that only counts three weeks will favour YouTube, because YouTube's sales arrive fast. Odins splits each channel's effect into sales within weeks and a share that builds brand and pays out over a longer horizon, so TV and YouTube can be compared on the horizon the business plans on. How that works, and the numbers behind it, are in brand effects in the model.
The practical output is a response curve per channel with a confidence range, a recommended split for the next period, and a list of what is still uncertain. Moving budget toward the channel with headroom typically gives 5 to 15 percent more effect from the same budget. Where the model is unsure which of two video channels has the headroom, the recommendation is a test, not a guess.
Odins builds and runs the model, collects the offline and digital data 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.
Yes. The model runs on weekly spend. Ratings sharpen the curve, because two weeks with the same spend can buy very different audiences, but they are an improvement rather than a requirement. Start with spend by air date and add ratings when the agency can deliver them.
A working rule is about two percent of total media spend, running often enough to vary. Below that, group it with linear TV and split it out when the volume grows. The model notes should say which channels are grouped.
It belongs on its own. Grouped with TV, it inherits TV's long carry-over and brand share, which it does not have. Grouped with paid social, it is lost among channels that behave differently. Separate it by campaign type from the rest of the Google Ads account.
With a prior and a wide range. The model starts from what linear TV does, adjusted for what connected TV is expected to do, and narrows the range as the weeks accumulate. If the first year's data cannot settle it, a structured test can. The honest answer in month three is a range, not a point.
A geo experiment gives a clean answer for one channel in one period, if you can hold out regions large enough to read. In small markets you often cannot. Either way, the experiment's result should calibrate the model, so one model answers the budget question across all channels, rather than stitching test results together by hand.
Read the complete guide to marketing mix modelling for the full method, see how offline media data gets into the model, or book a demo and bring your TV, connected TV and YouTube plans.