Marketing Mix Modelling and Attribution Modelling are two popular methods of evaluating marketing activities and have the same basic goal: measure the business impact of marketing channels and help marketers invest their budget more effectively. The key difference between the two methods is that MMM analyzes the effectiveness of marketing budget distribution from a top-down perspective, while attribution models are calculated at the level of a specific user (bottom-up approach). However, relying on user-level data is becoming more problematic - gaps in user data tracking are becoming bigger due to missing consents, privacy regulations and technical changes in iOS or web browsers related to cookies. In this article we take a look at both of these methods and their pros and cons.
Marketing Mix Modeling is a method based on statistical modelling with the aim of finding the correlation between marketing cost (and/or media exposure) and target metrics such as revenue, number of new customers or number of app installs on a “macro level”. In this context the macro level means “top-down” approach working with aggregated data – such as sales and investment breakdown by days or weeks – at the level of marketing channels and campaign types (both online and offline).
MMM provides two main results:
Attribution modelling is a different and more traditional approach to measuring digital marketing effectiveness. Every attribution model essentially works this way: for each conversion (order, app install,…) it tries to reconstruct the touchpoints (these may be ad clicks, direct visits, display ad views etc.) that preceded it for the given user and then using some set of rules it distributes the credit for that conversion to some or all of the touchpoints.
Marketing platforms often have their own attribution models which define how conversions will be attributed to channels or campaigns within them – whenever you see revenue associated with a campaign in Google Ads or Facebook or Google Analytics UIs, some attribution model had been applied.
Each platform has different attribution model – both in terms of the rules applied and the user touchpoints taken into account, e.g. Google Analytics is unable to see impressions on Facebook for a specific user. On the other hand, Facebook attribution model does not take into account clicks from Google Ads campaigns or other non-Facebook touchpoints in the user journey.
That’s why the number of conversions from Google (Universal) Analytics and Facebook won’t ever match and marketing platforms will often credit their own touchpoints with many more conversions – as they only see “their own clicks and impressions” and ignore or do not see anything else in the user journey. This often creates confusion for marketing teams who have to cope with multiple “versions of truth” and results in endless discussions on what attribution model is correct or should be used.
MMM is a proven method that has been used by large sophisticated advertisers to evaluate the effectiveness of TV or other offline media channels – historically it was reserved for advertisers with massive budgets as setting up an MMM solution required a significant cost and effort. However with recent developments in technologies (such as machine learning) that make MMM easily available to all types of advertisers and considering increased problems with user data reliability for attribution, MMM is getting a lot of traction in the digital world too. A study conducted by Accenture shows that companies using modern MMM methods have achieved 14–38% improvement in marketing ROI.
Furthermore the technological changes mentioned above also helped to move from one-off analyses to continuous and heavily automated MMM solutions. Modern MMM solutions are usually provided as a software solution, not as an analysis in Powerpoint. They automatically run, update their models on new data and provide new insights on an ongoing basis – giving marketers an “always-on” solution for marketing measurement.
To sum up – an attribution model distributes the credit for a conversion to the touchpoints it can see in the user journey according to some rules. It thus works in a bottom-up way: for each conversion and each user it tries to estimate what influenced the conversion and allocate credit accordingly. This way it is then possible to calculate the ROI of specific channels, campaigns or even creatives.
Attribution models differ in how smart they are about distributing the credit. Some models use simplistic rules such as “100% of the credit goes to the last touchpoint” (the infamous last-click model), some have more sophisticated but still kind of arbitrary rules (eg. all touchpoints get the same share – why?, U-shaped models etc.), others use advanced mathematics and algorithms to arrive at a better approximation of true incremental value of each touchpoint (eg Shapley models, Markov-chain based models) – these are often called data-driven attribution models.
But no matter how simple or advanced the attribution model is, there is one key piece for it to work properly: you need to have complete and reliable data on user identity and the corresponding touchpoints: what were the individual impressions the user saw and where, what campaigns they clicked etc. across all relevant channels. This was always very very difficult but recent ecosystem changes – users not being tracked (often 40-50%) due to privacy regulations and settings, cookie limitations (Safari cookie lifetime shortening, iOS14 restrictions and planned limitation of 3rd party cookies in Google Chrome) etc – have rendered data available for attribution into a state where it’s just too incomplete and too wrong to be useful.
Marketing Mix Modelling advantages vs attribution modelling:
Compared to MMM, attribution modelling has the following advantages:
Generally speaking, MMM will give you more accurate insights on what actually works, but it does not deliver on a micro-granular level (individual creatives, keywords,…). Attribution modelling can give you answers on a very granular level but is often significantly off from the actual truth (gaps in data, ignores non-media factors etc) and often overvalues the effect of specific digital channels.
However – MMM vs attribution is not necessarily an “either/or” situation – many sophisticated advertisers use both MMM and attribution modelling in their measurement stack – each for different use cases. Let’s review the main ones:
PROS
MMM does not require user-level data at all, only aggregated data for channels, spends and total revenue (or other target metric). For attribution calculation you need user-level data. The more cookies are blocked or deleted, the more gaps are present in the attribution model dataset.
Unlike attribution, MMM is not influenced by restrictions of today’s online world such as cookie bars, limitations in browsers (shortening cookie lifetime), ad blockers, iOS restrictions and so on.
MMM works equally well for TV, OOH, radio and other offline marketing activities, even live events.
The press & PR efforts of your CEO, branding video campaigns or influencer cooperations? Those are typically activities that most click-based attribution models struggle with. MMM is generally much better at understanding the true impact of activities that do not necessarily result in immediate clicks.
MMM can work with the concept of channel saturation and diminishing returns. These effects are extremely hard to capture in attribution modelling.
MMM typically also delivers not only the measured (historical) effectiveness but can also recommend optimal budget allocation subject to user-defined constraints, this is far beyond attribution models’ capabilities that merely report channel ROI, leaving the marketer alone in the optimization.
Attribution generally ignores the impact of pricing changes, discount promotions, competitor activity, market trends, seasonality, holidays, or factors such as weather – all these certainly do have an impact on your revenue.
The basic requirements for a reliable model assume having at least 1-2 years of historical data available in a daily or weekly granularity.
Although we can simulate what-if scenarios, it is still not the same as randomized controlled experiments. MMM alone without calibration from running experimentation results may have difficulties to distiguish whether the result is causal or just correlated.
PROS
For MMM you would typically need 2 years of historical data. Attribution modelling has no such requirement.
For decisions on a creative- or keyword-level, attribution is more useful as it allows you to break down the results to a lower level of granularity compared to MMM. The typical use case for attribution modelling would be day-to-day optimization by online / performance marketing specialists.
If you invest less than 10k USD / month into advertising, using an MMM solution would probably be an overkill both in terms of effort and money spent.
Some marketers like to have multiple views of their channels to find out and understand which of them work as users’ purchase journey opener (e.g. video or display ads) and which are the closing ones (emailing, product listing ads etc.). While useful in some cases, these views however do not help to measure the channel effectiveness or give insight on how to optimize budget allocation. in our experience, this approach more often than not simply creates confusion and endless discussions rather than actual improvement or value generation.
Data needed for multi-touch attribution are becoming harder to get. Users not giving consent to be tracked, platforms and browsers implementing new barriers, walled gardens like Google or Facebook not providing user-level impression data etc – all this impacts attribution and all this means that even if you have some data (say clicks), the attribution results will be likely significantly off from the true incremental value of the investigated channel or campaign.
Attribution models can cover only digital media. You cannot attribute a billboard ad touch to a specific user. This wouldn’t work well for a company with a large spending on offline media.
Attribution models generally do not consider brand strength that generates your sales and as a result overvalue digital channels and especially performance marketing
Attribution models credit your media (and organic and direct) channels for your whole revenue – they ignore factors like pricing and discounts, changes in stock availability , your distribution channels or your competitors activities – but in fact these influence your revenue often more than advertising.
Typical use cases for MMM are:
Typical use cases for attribution modelling are:
In today's world where privacy of users is crucial and marketers have to deal with realities of cookie restrictions, iOS changes, GDPR, CCPA, ePrivacy etc. on one hand and an increasing pressure to get the most buck for their marketing budget on the other hand, it is important to recognize that measurement methods of yesterday may no longer work or need to be supplemented by new ones that are more suitable for this environment. Relying just on the good old attribution models can and usually will lead marketers to the wrong conclusions regarding effectiveness and optimal channel spend. In this context, MMM can be the right choice to stay ahead – the solution is no longer reserved for Fortune 500 companies, today’s automated MMM solutions are suitable for mid-sized and even smaller advertisers and can be set up fairly quickly.
Whether you work with attribution or MMM, it is important to use the outputs similarly:
© 2023 by Marketingintelligence.io
Do you want to get more information about MMM or how/whether it could help your ROI & measurement issues? Lets us know – we can either answer your additional questions, share more materials with you or have a quick call if you prefer.
One of the new and exciting features of GA4 is called “predictive metrics” – with these you can learn about your customers and their shopping behavior. These are added automatically to your data and are based on Google’s machine learning expertise.
GA4 currently includes the following predictive metrics:
Google BigQuery is a cost-effective and highly-scalable cloud data warehouse optimized for high performance on very large data sets. With GA4 you can export all your event-level data to BigQuery for additional analytics or data science initiatives.
Example use cases might be
Example use cases might be
You will need to have a Google Cloud account set up and maintained for this purpose. That is why we are here and can solve all the infrastructure for you and with you!
Since 1/2022 GA4 has made a data-driven attribution model available to all users – unlike in Universal Analytics where the DDA model was only available to GA360 customers.
DDA is an algorithmic attribution model that quantifies the value of each touchpoint in the user journey (such as campaign click) and it does so by smart modelling behind the scenes without human bias that is always present in rule-based models.
The DDA model in GA4 is also better than the one on GA360 as it takes into account up to 50 touchpoints in the user journey vs. the GA360 only took 4 touchpoints.
It is also possible to set the default attribution model to DDA (or another model) from the old Last Non-Direct Click. And then all your reports and data exported to BigQuery will use the newly selected model as default.
DDA in Google Analytics 4 excludes almost all direct visits from receiving credit – you may or may not like that, but as it’s been always the case with attribution we recommend validating any model using marketing experiments.
GA4 is built for a world where more and more users opt out of cookie consent and other methods for data collection.
Google uses machine learning modeling to fill in the data gaps – some of these are already in the current GA4 others will be deployed in the future.
For example, modelling conversions allows GA4 to properly attribute conversions without user identification – this is crucial for optimized advertising campaigns and automated bidding. This covers situations such as some browsers limiting the time window for first-party cookies, conversions for unconsented users, Apple’s App Tracking Transparency (ATT) impacts, cross-device user behaviour, and others.