How should connected TV be credited in a media mix model?
Our answer
Give connected TV its own line in your mix model, fed by flight with impressions and spend. Count the branded search and retail sales it drives, and anchor its estimate with a geo holdout.
- Confidence
- Medium
- Strongest proof
- 3 official rules
- 4 datasets, 2 practitioners, 8 opinions or claims
- Sources
- 8 sources
- 17 statements
GRPATVCP
New here? Take the 1-minute tour
Why this is the answer
3 reasons and our judgmentAn uncalibrated model can give TV almost any coefficient. TV spend is lumpy and moves with other upper-funnel spend, so geo tests are needed to anchor it (Measured).
Connected TV gets undervalued when it is grouped with linear TV or online video, or when it has too little weight or history to read. Roku and Vibe, which both sell connected TV ads, give this advice.
Much of connected TV's effect lands in other channels. It shows up in branded search, direct visits and retail sales that other channels get credit for (Prescient AI, mix modeling software maker).
Our judgment
- Give connected TV its own line, separate from linear TV and online video. This follows Roku's and Vibe's advice to model streaming as its own channel and Measured's finding that streaming and linear TV return differently.
- Vary connected TV spend in clear flights rather than one flat run. Vibe says a continuous run with no budget variation gives the model nothing to regress against, and Measured says TV spend that moves with other spend is hard to separate.
- Have the model credit connected TV for the sales it drives through branded search, direct visits and Amazon or retail. Keep branded search an outcome the model explains, not a driver fed in beside connected TV (consistent with Checking whether a mix model is right), going by Prescient AI's account of the halo and Recast's method for upper-funnel channels.
- Read an uncalibrated estimate as a wide range and let a geo holdout anchor it. We side with Measured, Roku and Google's Meridian calibration over Prescient AI's view that the model should judge the test, because the holdout is a controlled test and the model is not.
Our judgment, built from Roku, Vibe, Measured, Prescient AI, Recast and Google: our reading of these sources, not a finding from any one of them.
What the sources say
17 statements from 8 sourcesGoogle's Meridian mix model docs say any channel with reach and frequency data can be modeled that way, with data at the same geo and time level as sales, since impressions alone ignore repeat exposure.
“Typically, MMMs rely on impressions as input, neglecting the fact that individuals can be exposed to ads multiple times, and the impact can vary with exposure frequency.”
Recast, which sells mix models, says some sales an MMM calls baseline may come from brand spend months earlier, which the model struggles to separate. Its add-on uses awareness surveys. No results shown.
“some of the sales an MMM calls “baseline” today may reflect brand investment made months earlier.”
Roku says mix models should treat streaming as its own channel with its own decay curve, and CTV looks undervalued when lumped with linear or given too little spend. Its Nielsen multiples have no method.
“a model can’t learn from spend it barely sees.”
Roku, which sells CTV ads, says the most common CTV measurement error is leaning too much on MMM or MTA, and advises quarterly incrementality tests fed back into those models. No data.
“The most common CTV measurement error we see is brands relying too heavily on MMM or MTA. Incrementality is essential to triangulate your MMM or MTA results.”
Tatari relays how Jones Road, at about $105 million in revenue, grew TV to about a quarter of its media, and branded search rose from about $1K to $4K a day after linear TV began. One brand.
“Branded search went from about $1K/day to $4K/day.”
Google's Meridian docs say its CalibrationBuilder turns lift experiments into ROI priors, discounting older tests with a fixed 52-week half-life. It suits tests against zero spend, not reduced spend.
“Recency calculations use a fixed half-life of 52 weeks (1 year).”
Measured's CTV panel, over $25M of spend, found streaming about 75% more incremental than linear TV and YouTube CTV and video at 1.6x the category benchmark. Small-budget CTV tests often showed no measurable lift.
“the platform is rarely the variable that decides the outcome”
Google's Meridian flags channels whose results look uncertain, biased or implausible (a score below 67.5) as priorities for experiments, putting higher-spend channels first.
“Spend-weighting also prioritizes calibration for higher-spend channels due to their greater business importance.”
Recast, which sells mix models, says upper-funnel channels also drive sales through brand search, so it now models brand search volume and judges those channels on search demand too. No figures given.
“the total impact of an upper funnel channel is the sum of its direct sales plus its assisted sales.”
Measured, which sells testing and mix models, says an uncalibrated mix model can give TV almost any effect because TV spend is lumpy and seasonal, so geo test results should calibrate it.
“An uncalibrated MMM can assign TV almost any coefficient the data will tolerate.”
Measured says that in one four-week geo holdout of a retailer's streaming TV, roughly half the conversions the platform claimed were incremental. It says TV tests typically run four to eight weeks.
“most streaming platforms support geographic targeting exclusions, so the holdout can be implemented directly in the buying platform.”
Cody Plofker's first independent check on TV is the post-purchase survey, scaled up by the response rate (if half of buyers answer, double it), then compared with other attribution models.
“never allow anyone to create their own homework”
Vibe, which sells streaming TV ads, says most mix models need at least 13 weeks of consistent CTV spend before it shows up as an independent variable. It gives no method or data for the threshold.
“most MMM engines need at least 13 weeks of consistent CTV spend before streaming TV shows up as a statistically independent variable.”
Michael Kaminsky, whose firm Recast sells mix models, says most businesses do not need one and it can mislead them; he cites studies where last-touch was about 85 percent as good as experiments in simpler mixes.
“most businesses do not need an MMM. In fact, an MMM will be net negative for them because it will actively mislead them.”
Pushes back on geo tests: Prescient AI says they suffer contamination between markets and show only one window, and that its own mix model shows Tatari TV campaigns reporting $13 where it credits $28,053.
“None of these methods works best in isolation. The strongest approach uses one to check the other.”
Tatari says TV attribution that does not subtract normal baseline traffic overcredits TV. In one client case, shared IP addresses got 28% of impressions but 57% of 1-day view-through responses.
“Use view-through to optimize; use incrementality to justify.”
Prescient AI, which sells mix models, says CTV's effect shows up in branded search, direct and retail, and that geo tests are snapshots that miss how CTV works with other media. It gives no data.
“Geo-based incrementality testing can establish revenue lift from a CTV campaign in isolation, but it can't show how that campaign interacted with the rest of your media mix.”
In the full answer
For readers who sign in- What to do this week3 steps
- The numbers behind it3 figures from cited sources
- Where each source standsA chart placing each source on the question
- The strongest disagreementThe best case against the answer, from a cited source
- How sure we areWhat its confidence rests on, and what would change the answer