Time granularity: This refers to the frequency at which you collect data, such as daily, weekly, or monthly.

Media granularity: This pertains to the depth of media data you collect, such as at the media level (e.g., TV), media channel level (e.g., individual TV channel), or campaign level (e.g., ad or ad campaign on TV).

Sales granularity: This indicates the level at which you measure the impact of marketing, such as on total sales (e.g., clothes), product category sales (e.g., sports clothes), or individual product level (e.g., a specific sports t-shirt).

Result accuracy: Increasing time and media granularity generally increases the number of data points in the model, bringing several benefits:

Optimization opportunities: Improving time, media, and sales granularity results in a more detailed MMM, thereby expanding optimization opportunities. For instance, if you have MMM results at a product level instead of a product category level, it allows you to optimize marketing for individual products by focusing on the best performing product instead of just shifting budgets between categories. Similarly, weekly MMM results enable marketing optimization on a weekly level, allowing more precise timing of campaigns than when using monthly level data. Moreover, if you measure the impact of individual campaigns instead of the media as a whole, you can optimize marketing at the campaign level by repeating the most impactful campaigns.

Using aggregated data often leads to the "tyranny of the average", where only the average for the entire measurement group is visible, not the best and worst performing parts. For instance, you won't see the best and worst-performing products in a product group if you only measure marketing on the group level. However, if you measure marketing at a lower level, you'll typically find more components to optimize, and the range of performance differences usually expands. This means you'll have more optimization opportunities. (For further discussion, refer to Heliste (2019).)

The example below illustrates this phenomenon by showing the distribution of ROI between product categories, product groups, and individual products:

Equal granularity requirement: In Marketing Mix Modelling, it is typically required to have data at the same level of aggregation across all sources. Achieving this can be challenging, leading to the common practice of aggregating data to match the source with the least granularity. An alternative solution involves disaggregating low granularity data, such as dividing monthly spend evenly across weeks. However, this method often requires assumptions that may not be accurate. (For further discussion, refer to Chan and Perry (2017).)

Lack of data: Paradoxically, a higher granularity can lead to a lack of data. This depends on how granularity is being increased. Take sales granularity as an example. Usually, this entails dividing the model into multiple sub-models (i.e., one model for each sales category to be measured). If you're attempting to analyze the impact of marketing on a specific t-shirt, but you have only advertised it once, the model's accuracy will likely be low due to insufficient data points. This issue can also surface when enhancing media granularity; if certain media are seldom used, you may only have a few data points for each of them, resulting in lower quality estimates for these media.

Issues with data collection: Obtaining reliable, detailed data can pose challenges, particularly if the collection process is not adequately automated. Poor data quality can, in turn, lead to unreliable results.

Impact of higher-level activities and cross-category influences: Promoting your brand or a product category often benefits individual products too. Similarly, promoting specific products can increase sales of other products. When measuring marketing effectiveness of individual products, it's essential to consider these effects, for example, by allocating a portion of the spending from higher-level activities or from promoting other products.

Adstock and carryover effects: When using daily data, the impact of marketing often affects multiple days in the time series. However, with monthly data, most of the impact likely falls within the same month. Therefore, considering adstock and carryover effects is crucial in high granularity modeling.

More complicated models: Greater granularity usually results in more data points or/and parallel models, increasing the demand for computing power. Similarly, factoring in adstock, high-level activities, and cross-category effects will make the model more complex.

Business needs: Are you looking to optimize marketing at the product group level versus the product level, or at the media level versus the campaign level? If you do not intend to make optimizations at lower levels of sales and media dimensions, then it may be unnecessary to increase granularity for those dimensions. However, enhancing time granularity is generally advantageous as it simply expands the data points for the model.

Optimization capabilities: Can you make the hundreds or thousands of different optimization decisions required to properly optimize marketing at a low level? Generally, the more decisions you have to make, the more automation is needed in decision-making. If high granularity Marketing Mix Modeling (MMM) is not combined with automated decision-making, optimizing marketing can become challenging. In some instances, it may be sensible to use different tools at various levels of granularity. For example, an MMM tool at a high level for high-level budget allocation and a promotion optimization tool at lower levels for optimizing promotions.

Data collection capabilities: Do you have the resources to obtain the necessary data for the modeling? If you desire a higher level of granularity, you will need to undertake more data collection and ensure the data is of high quality. Therefore, it is important to invest in automating data collection.

Modelling capabilities: Do you have the capabilities and resources to build and run more complex and heavier models? If not, it might make sense to resort to simpler higher-level modelling.