FAQs

Is the library free?

Yes, the library is free to use and is open sourced on GitHub for anyone to use.

Do we have to share our data with Google to use the library?

Google won't have access to your input data, model, or results (apart from Google media data supplied through our MMM Data Platform). If you request data from the Google MMM Data Platform, that is the only data that Google has access to. But Google won't know whether you actually include that data in your model. Your actual model inputs and outputs are entirely private, unless you choose to share it with your Google representatives.

What is Meridian GeoX and how is it different from Meridian?

Meridian GeoX is our upcoming open source geo-based incrementality solution that bridges experimentation and MMM. Run Meridian GeoX to calibrate your Meridian model or as a standalone geo experiment to measure the true causal impact of your marketing performance.

Why is running GeoX experiments important for Marketing Mix Modeling (MMM) users?

Observational causal inference methods like MMM rely on assumptions to estimate the true incremental impact of specific channels. Meridian GeoX addresses this by providing causal priors based on experiments. It offers cross-publisher capabilities, ensuring a seamless integration that more accurately calibrate your measurement results.

As a current LightweightMMM user, is extra work needed to build data input for Meridian?

To take full advantage of the new Meridian innovations, you will need to add more data dimensions such as:

• Reach and frequency
• Experiments
• Google Query Volume (GQV)

You can still run Meridian without these dimensions, although you will miss out on the new innovations. For more information, see Migrate from LightweightMMM.

Can I collect all data types simultaneously (performance, YouTube reach and frequency, Google Query Volume) in the MMM Data Platform interface?

Performance data and YouTube reach and frequency data must be requested separately. The request workflow is detailed in the User Guide that is included with your MMM Data Platform access email.

What is the scope of the GQV data that I can request?

The Google Query Volume, the output includes:

• QueryLabel - Brand or generic
• ReportDate
• TimeGranularity (You can request Daily, Weekly_Sunday, or Weekly_Monday data.)
• GeoCriteriaId
• GeoName
• GeoType
• IndexedQueryVolume - All query volume data is indexed. Raw numbers aren't provided for Query Volume.

Can I apply the GQV methodology for non-Google search data?

Organic query volume from non-Google search engines is often unavailable. Some alternative options are described in Understanding query volume as a confounder for search ads.

Why should I get my GQV data from the MMM Data Platform?

The GQV report from the MMM Data Platform was purpose-built for MMM, providing the most holistic view of GQV performance at the right granularity for your model. The MMM Data Platform uses a consistent index value for each advertiser<>country pair, so you can easily update GQV data for every model refresh and accurately compare performance overtime. Since your GQV and performance data requests are linked in a single platform, we are able to ensure a consistent mapping of branded keywords with the brand delivered in your MMM data request.

Search query data from other sources may be used as a confounder in Meridian, but there may be limitations for the MMM use case. For example, Google Trends normalizes search data, aggregates queries by topic, and uses a sample of searches in reporting. While this is great for producing trend insights quickly, we recommend sourcing GQV from the MMM Data Platform for Meridian to ensure the data is brand-specific, consistent, and provided at the right granularity for your model.

For a given media lever, how can I set different priors associated with different time periods?

The closest thing to this would be the roi_calibration_period argument. Based on section 3.4 of the MMM calibration whitepaper, we suggest calculating a spend-weighted average ROI for the experiments and passing roi_calibration_period to match the four quarters of the experiments. If the experiments have very different standard errors, you might want to further weight the experiments accordingly. For more information, see Set the ROI calibration period.

Can I put a temporal prior for the knot values?

Meridian does not support time varying priors for knot values.

How does Meridian allow for time-varying media effects?

Meridian allows for the modeling of time-varying media effects through data partitioning. One must manually divide the channel's data into distinct periods and model them as separate channels. For example, if you suspect a channel performed differently in 2024 than in 2025, you would create two new channels in your dataset: one containing only the 2024 data and another with the 2025 data.

We caution against creating too many temporal breakdowns, as this can lead to a rapid increase in the number of channels. Too many channels increases the number of parameters Meridian needs to estimate, making it hard for the model to distinguish signal from noise with limited data (see Amount of data needed to learn more). In the extreme case, too many channels can lead to model identifiability and convergence issues.

How can I get detailed decomposition information of the regression, such as getting dataframes for the posterior draws?

Posterior samples are in the inference_data object, and you can turn this array into any dataframe you need. To access the data samples using the docstring, see meridian.model.model.Meridian.

Can I measure synergies between channels in Meridian?

Meridian doesn't support this kind of analysis.

Is it possible to get a temporal read-out of ROI with Meridian?

You can use the selected_times option in Analyzer().roi() to choose the time periods of interest.

Important: When tracking ROI over time, consider that even though the coefficients in the model are not time-varying, the ROI can still change over time because it is dependent on additional factors that might vary across time. For example, the Hill curves model the non-linear, diminishing returns of media execution, and therefore the amount of media execution at a given time can impact the ROI. Addtionally, media allocation can vary across geos over time with different effectiveness and the cost of media execution can vary across time.

Can Meridian's hierarchical model structure be used for a categorical variable besides geo?

Meridian's hierarchical modeling structure was designed for geo-level modeling, which has important advantages. Although it is possible to use another categorical variable in place of geographic region as the hierarchical variable, this is generally not recommended. If another hierarchical variable is to be used, it must be possible to stratify both the KPI and the media units by this variable.

Some examples of hierarchical variables that typically won't work include the following.

• Product (with halo effects). If you think that an advertising channel can affect the KPI of multiple products, then "product" shouldn't be used as a hierarchical variable. A channel can affect multiple products if it is targeted at the entire brand as a whole, rather than being product-specific. A product-specific ad campaign can also affect the KPI of other products due to so-called "halo effects".
• Sales channel (for example, online versus retail). Media units are usually not specific to a single sales channel.
• Ad campaign. Although media units can often be stratified by campaign, it is not possible to stratify the KPI by campaign.

Even if both KPI and media units can be stratified by a non-geo variable, it is important to consider whether the following hierarchical model assumptions are applicable.

• The hierarchical model partially pools information across categories to estimate the model coefficients for the treatment variables and for the control variables. Other model parameters such as time effect (trend and seasonality), adstock, and diminishing returns parameters are assumed to be common across categories (complete pooling). These assumptions are typically reasonable for geo hierarchy because the KPI and media units have essentially the same interpretation across geos.
• The assumption of common time effect, adstock, and diminishing returns parameters generally makes more sense if both the KPI and media units are roughly on the same scale for each category. For geo models, this is accomplished by population scaling both the KPI and media units. For other variables, it is less clear what scaling variable to use in place of population, and whether it should be applied to the KPI, the media units, or both.
• The stable unit treatment value assumption (SUTVA) requires that the treatment (advertising) exposure in one unit does not affect any other unit. This assumption is usually a reasonable for geographic treatment units (although never perfectly met). For non-geo categories, there may be more cross-category contamination of the media effects.

If you are considering using a non-geo hierarchical variable, keep in mind the following considerations.

• A control variable can take the same values across categories, as long as the control variable has variation within each category. However, convergence issues may arise if full or close to full knots are used to model time effects.
• Consider setting unique_sigma_for_each_geo=True in the ModelSpec, since residual variance may not be the same across non-geo categories.
• Because geo hierarchical modeling is the most common use case, Meridian's naming convention labels the hierarchical variable as "geo". If you are using the CSV data loader, be sure to label your hierarchical variable as the "geo" column in load.CoordToColumns.
• Meridian requires a "population" variable. This can be set to one for all category levels if the hierarchical variable is not geo, but it might make sense to scale by something else. Meridian automatically scales both the KPI and media units by population. If you want a variable NOT to be scaled, you must multiply the input data column by the population so that it takes the original value after population scaling.

If I use a subset of previously selected knots as required knots, will AKS produce identical results?

Not necessarily. Automatic Knot Selection is an iterative algorithm that utilizes a penalized regression solver. Providing required_knots influences the penalty weights and the starting conditions at the very first step of the algorithm. Because of this different starting condition, the optimization path that the solver takes can diverge, which may lead to a different final set of selected knots.

However, this shouldn't be a cause for concern. Both sets of selected knots are valid, data-driven representations of the underlying trend. The algorithm finds an alternative, equally valid optimal solution given the constraints.

Can I measure the ROI of bidding strategies based on the bid targets set?

Google's MMM data feed provides bid strategy type (such as Maximize Conversions and Target ROAS) by campaign, but the feed does not include the bid target itself. To view this specific dimension, advertisers can source bid strategy reports directly from Google Ads or work with their Google Account Representative on a custom data solution.

Can Meridian be used for forecasting?

Meridian can be used for future scenario planning, which includes budget optimization and estimating incremental outcome under hypothetical future budget scenarios. Meridian is designed for causal inference, not prediction. Consequently, Meridian cannot be used to forecast future raw outcome. See Why Meridian doesn't forecast outcome for more details.