Google Meridian: What you need to know

Google has recently announced the launch of its new open-source media mix model (MMM) platform called Meridian. Google’s Meridian is currently in limited access and is likely to remain this way for the next 4-5 months.

Based on a recent visit to Google’s Think Measurement event, access will remain limited until the second half of 2024. Google’s current focus appears to be to make Meridian available to priority advertisers first and agencies later, so as a result, we like many peers have not yet had access to the modelling software here at Impression. Once we have more updates on the package and get access to the source code we will update this blog post.

MMM is a cookieless form of measurement which utilises statistical techniques to help you answer three main questions: 

1. To what extent are my advertising channels contributing to KPIs?
2. What was my return on advertising spend (ROAS)?
3. How do I optimise my budget allocation such that my ROAS is maximised?

For a comprehensive understanding of MMM and its role in media effectiveness measurement, we recommend diving into our beginners’ guide to incrementality testing and media mix modelling. This guide delves deep into the intricate details of MMM in general, exploring its principles, methodologies, and practical applications.

Key out-of-the-box features Meridian has to offer

Utilises Bayesian Regression: By using Bayesian methods, Meridian can quantify any uncertainty associated with advertising contributions more effectively and make informed decisions based on these intervals of uncertainty. This approach allows for flexible modelling in cases where data is limited and the complexities of your business make the incorporation of prior knowledge crucial. Ultimately Bayesian modelling can provide richer insights into the underlying processes.

Geo-based hierarchical modelling: Meridian allows for the incorporation of geo-level advertising data, which likely contains much more information about your marketing effectiveness than national-level data. By using a geo-based hierarchy, geographies can share information with each other. This improves uncertainty when modelling geographies with limited data or very little prior knowledge. Meridian will support models ranging from 50+ geographies to the standard national-level approach if geo-level data isn’t available.

Uses ROI priors: The Meridian model structure can be re-parameterized such that the ROI of each channel is a model parameter. This enables insights from incrementality tests or industry benchmarks to be used for calibration. Google has announced a slight revamping of their incrementality tests such that user-based and geo-based tests are now going to be referred to as “conversion lift” in Google ads. As these new “conversion lift” tests are going to use user-level data, it’ll be important to ensure that advanced MarTech tracking is in place such as Enhanced Conversions and Server Side Tagging, this will ensure your results are as accurate as possible as cookies are depreciated.

Incorporates reach and frequency: Reach refers to the number of users that see an ad within each period. Frequency refers to the corresponding average number of times a given user sees an ad. Meridian provides the option to use reach and frequency data as model inputs to provide additional insights into the performance of particular media channels. Google claims that currently, MMM is more budget-focused as opposed to outcomes/ROI-focused, incorporation of reach and frequency data, along with ROI priors should help to tackle this.

Time-variant intercept: The intercept in an MMM describes what your baseline KPIs would be in the absence of any contributing variables such as media activity, non-marketing factors or seasonality. In traditional MMM the intercept is modelled as constant, meaning that the idea that baseline sales can fluctuate over time is neglected. Meridian allows a time-variant intercept, hence allowing for a more flexible way to capture brand awareness or general business growth.

Dashboard output: Based on our recent visit to Google’s Think Measurement event, we know that once Meridian becomes available for public use we’ll see model outputs in terms of a dashboard. We believe that this will be similar to the 1-pager outputs from Meta’s Robyn. 

Things to be cautious about regarding Meridian

Does not allow for time-variant media parameters: Meridian neglects the idea that media contributions can fluctuate over time. Admittedly making every media channel time variant may be unnecessary and cause some serious overfitting problems. However not having the ability to experiment with this or to apply a time-variant contribution to a particular channel is a major drawback.

Limited prior customisation: Based on recent documentation, it isn’t yet obvious if Meridian will allow a wide range of customisable prior distributions. Capturing the complexities of a specific business using a “one prior fits all” approach might not work for all business types. Meridian has hinted that in the future a wider range of priors might be added, but no timeline was attached to this claim.

Geo-based hierarchy with limited geos: Not all businesses will advertise in a large number of countries. Some businesses might only advertise in two or three countries. If these types of businesses were to model using a geo-based hierarchy, then they might run into some sampling problems. This is because hierarchical models typically require a larger number of levels to achieve reliable estimates.

Lightweight MMM discontinued

Before the release of Meridian, Google’s unofficial software solution for media effectiveness measurement was LightWeight MMM. Lightweight MMM is an MMM library which also utilises Bayesian statistics and adopts a geography-based hierarchical structure. Here, Google has announced that once Meridian reaches public availability, lightweight MMM will be discontinued.

The discontinuation of Google’s unofficial product, Lightweight MMM, in light of the forthcoming release of Meridian, heralds a notable transition for users of the former. This shift signals a significant change, as Meridian, leveraging TensorFlow Probability over Numpyro in Lightweight MMM, presents an alternative approach to MMM. This development suggests that users familiar with Lightweight MMM will need to adjust to a new system that diverges considerably from its predecessor in methodology and user experience.

Google Meridian’s four pillars

Google Meridan claims to be sold on the following four pillars:

Innovation:

• Combines video planning with outcomes – YouTube reach and frequency input data.
• Calibration with experiments – feeds causality into the MMM.
• Measuring search – incorporation of Google search volume (GSV).

Transparency:

• OSS – increased efficiency and reduced manual effort.
• Customisable – recognises that a one-model-fits-all isn’t always the best approach to MMM.

Actionability:

• Enables actionable decisions – optimal frequency such that ROI is maximised.
• Provides the most impactful media mix under constrained optimisation.
• Full technical documentation on academia & how to use and implement

Education:

• Full technical documentation on academia and implementation. For those with early access, more documentation is available.

Alternative MMM platforms to consider

Currently, there are a handful of different open-source MMM platforms out there for use. The most popular include Meta’s Robyn, PyMC and PyMC marketing.

See some of the main comparisons tabulated below:

At Impression, we advocate for a tailored approach to measurement, understanding that a one-size-fits-all model may not capture the intricacies of diverse marketing strategies. That’s why we’ve developed our bespoke custom media mix model using cutting-edge tools from PyMC labs, ensuring adaptability to the dynamic digital marketing landscape while delivering precise insights tailored to each client’s unique needs. 

As you can probably tell, our preference is PyMC as we feel it allows us to construct, implement and analyse complex probabilistic models with ease. Although it requires some statistical and computational knowledge to fully understand how to implement MMM in PyMC, the extensive documentation and notebooks that PyMC has to offer make it accessible to both beginners and seasoned practitioners in the field. With its robust capabilities for time-variant parameters, uncertainty quantification, and out-of-sample prediction, we feel PyMC is the optimal available tool that facilitates sophisticated analysis and decision-making in the digital marketing domain.

As we eagerly await the broader release of Google’s Meridian, the true practicality of it remains uncertain. While documentation hints at statistically robust foundations of a similar nature to other established MMM libraries, the full extent of its capabilities and ease of integration will only become clear with wider adoption and access to its source code.

If you are looking for support with conducting media mix modelling for your business, don’t hesitate to get in touch with our data science team.