3. MMM, Incrementality
3.1. Intro
- Shortcomings of click-based attribution
- Digital campaigns w/o clicks, e.g. video campaigns
- Dark social
- Cross-device attribution, since click attribution can be made only in the same session
- Long sales journeys, conversions happening outside of lookback windows
3.2. Incrementality
- Click incrementality
- It’s not just sales or conversions, it can also be paid brand search clicks per se
- This is interesting also to understand cannibalization btw Paid and Organic
- There’s a study by Google showing how many clicks were incremental and how many were cannibalized to Organic Search
- Here’s another case history by Barbara Galiza Measuring incremental impressions, clicks and conversions for Paid Search (Methodology + Google Sheet Template)
- Conversion incrementality
- There are cases where clicks are not incremental but conversions are
- This tries to estimate conversions that wouldn’t have happened without a specific mktg initiative
- How incrementality is measured
- Holdout tests are performed turning off a campaign to all or certain audiences
- Geofencing is a technique similar to the previous one but in this case campaigns are turned off based on geographical areas. GeoLift is an R package made by Meta helping with this GeoLift Walkthrough | GeoLift
- Causal inference, is a statistical model measuring correlation btw campaigns and sales.
- Design a test
- Control targeting, be sure you can select audiences precisely
- Experiment control, start and stop the experiment as needed
- Conversion event tracking, it’s important to measure frequency of conversions within the test group and overall
- Campaign metrics, analyze spend and impressions for both control and test group
3.3. MMM
- MMM in layman’s terms way
- There are a bunch of inputs and outputs
- MMM tries to find the correlation btw inputs and outputs
- Understanding Bayesian models
- Prior knowledge, Bayesian models need historical data, for instance for daily data you need at least 1 year or even more
- Data updates, the model gets updated when new data arrives
- Probabilistic approach, probabilities are assigned to different outcomes based on input data
- MMM and Bayesian models
- Simulations: MMM runs simulations
- Statistical analysis: Correlation is calculated for each simulation
- Channel impact: Finally MMM isolates the impact of each mktg channel
- Requirements for MMM
- Date-level datasets, data granularity should be daily or weekly depending on mktg activities involved and the type of business
- Mktg activity data, all media activities should be included
- Target metric, of course this depends especially on the business
- Historical data, at least 1 or 2 years of data are needed
- Sample dataset generator by Timo Dechau https://replit.com/@TimoDechau/Marketing-Mix-Model-Playground
- When to use MMM and limitations
- There’s a need for extensive enough datasets
- Time and understanding to fine tune MMM is important
- It’s good for high-level channel decisions
- Barbara recommends 6 figures per month is the minimum recommended spend for those who want to use MMM