How much history and data does a media mix model need?
Our answer
Plan on at least two years of weekly sales and spend data, about 10 data points per channel or factor, and spend that has moved. With less, use daily or regional data, merge channels or wait.
- Confidence
- Medium
- Strongest proof
- 1 controlled test
- 5 official rules, 3 practitioners, 6 opinions or claims
- Sources
- 13 sources
- 15 statements
G
New here? Take the 1-minute tour
Why this is the answer
3 reasons and our judgmentMeta's own mix model sets two years as the floor. Its Robyn guide asks for at least two years of weekly data, or 4 to 5 years of monthly data, and about 10 observations for each variable in the model.
Two years can still be too few for many channels. Google's Meridian guide says two years of weekly data, 104 points, is too few for 12 channels and six controls. NIQ asks for three years when you model national totals only.
Vendors and agencies treat two years or more as normal. Recast, a marketing mix modeling firm, asks for 27 months of data to model two years. Flighted, Cometly and PPC Land describe two to three years of history as the norm.
Our judgment
We set the two-year floor and the fallbacks below it from several sources. Meta's Robyn guide sets two years as its minimum, Google calls two years too few for many channels, and Recast, Flighted, Cometly, PPC Land and Improvado want about two years or more. We read the shorter floors from Balistro (12 to 18 months) and Elevarus (18 months) as fitting a model with few channels, since 18 months is about 78 weeks, enough for about seven variables at Robyn's 10 observations each.
Our judgment, built from Meta, Google, Recast, Flighted, Cometly, PPC Land, Improvado, Balistro and Elevarus: our reading of these sources, not a finding from any one of them.
What the sources say
15 statements from 13 sourcesMeta's Robyn guide says a mix model needs at least two years of weekly data (4 to 5 years if monthly), about 10 observations per variable, and spend that varies enough to learn from.
“In order to get robust results, an MMM will need a minimum of two years of historical weekly data.”
Google's Meridian docs ask for a KPI, media exposure and spend by channel, week and ideally region, plus controls that affect both sales and media. Spend can stand in for impressions, with caveats.
“If physical impressions or click data is completely unavailable, you can use spend as a proxy for the media exposure metric.”
Recast, which sells mix models, asks for 27 months of data to model a two-year window, and does not look further back because marketing performance changes over time.
“Two years gives us a chance to observe seasonal patterns and holidays more than once, so we can distinguish between anomalies and strong seasonal behavior.”
Recast's data guide asks for a daily table of spend by channel and the KPI going back 27 months, with search split into branded and non-brand, plus a promotions calendar, re-shared every week.
“We recommend sharing 27 months - and this will need to be re-shared every week.”
Boathouse's strategy chief says that for one client an off-the-shelf mix model credited paid media with about 75% of new customers, while a custom model on the same data said under 1%. Check models with lift tests.
“A result below 1% deserves as much scrutiny as a result of 75%.”
PPC Land says most modelers want two to three years of weekly history for a mix model, and that spend must vary or the model has nothing to learn from.
“Spend must also vary, or the model has nothing to learn from.”
Flighted describes a mix model as a regression on two to three years of weekly or daily data, and tells early-stage brands to skip it until they spend across three or more channels.
“If you're an early-stage brand still finding your first profitable channel, skip MMM entirely for now”
Balistro, an agency, says to add up the conversions every platform claims and compare with real orders; for most DTC brands it audits, the sum is 1.4x to 2x real orders.
“For most D2C brands we audit, the platform sum is 1.4x to 2x real orders.”
Google says a mix model with 26 parameters on two years of weekly national data is too little to estimate reliably; fewer channels, more years or geo-level data help. It calls the guidance rough.
“This sample size is too low to estimate the model reliably.”
Eightx puts the mix model adoption floor near $20M+ revenue and $3M to $5M+ a year in media across 3 or more channels; under $1M, use platform reports plus quarterly incrementality tests.
“The adoption floor for DTC is roughly $20M+ in revenue and $3M-$5M+ in annual media spend”
Elevarus says most accounts need 18 months of weekly spend and revenue for a mix model, single-channel advertisers don't need one, and a small geo test should come first. No data.
“Single-channel advertisers do not need this. Multi-channel advertisers do.”
A PyCon DE talk on synthetic data shows hierarchical regional mix models narrow uncertainty for data-poor countries by borrowing from rich ones, but fail at about 8 data points per country.
Improvado, which sells marketing data tools, says a stable mix model needs 80 to 100 observations, roughly 18 to 24 months of weekly data, and about 10 observations per variable.
“if budgets stayed flat for 18 months, you won't have enough signal regardless of time span.”
NIQ's Meridian playbook asks for at least 104 weeks of weekly data for regional models and 156 for national ones, with about 15 observations per variable, so 104 national weeks support about 7 variables.
Cometly, which sells attribution, says a mix model needs two to three years of data and at least four channels; under $100,000 a year across fewer than four channels, focus on attribution first.
“The practical threshold for most small businesses: if you're spending less than $100,000 annually on marketing across fewer than four channels, focus on attribution first.”
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
- What the research saysWhat academic research adds, kept apart from the answer
- How sure we areWhat its confidence rests on, and what would change the answer