If you need a definition: Attribution modeling is an area of expertise which seeks to uncover the factors which influence customers prior to a sale. Generally speaking this involves the investigation and measurement of the entire customer journey: from the first time they hear the brand mentioned, through to an in-store visit or sale. If these brand contact points are accurately measured, they can then in-turn be ‘attributed’ back to sales revenue.
Growth-focused marketing and sales professionals are particularly interested in attribution to improve their budgeting, strategy and tactics.
The Story So Far
First we need to set the scene.
Prior to the digital age, much of the ‘marketing roi’ work was performed by research agencies who would poll a sample of customers and ask them questions like, ‘when did you first hear about us?’ etc. Data was also gathered from media research companies such as Nielsen or Roy Morgan who routinely sampled consumers over a period of time and provided research report data to perticipants in the industry vertical such as media agencies, TV stations, associations etc.
During this time, most brands’ marketing expenditure was heavily allocated toward the purchase of advertising media – usually via an outsourced agency arrangement. So by using both of these sources of data, brands could make reasonably informed decisions regarding the effectiveness of their marketing expenditure. Smaller businesses who could not afford to purchase these higher-priced media or hire a market research firm were left to make blind guesses as to the validity of their marketing investments.
Then there came the digital age and with it, unrivaled granularity and the ability to track users based on their digital signatures. Google Analytics measured where the traffic originated and the number of digital advertising tools exploded. Much of the early attribution tools relied heavily on browser cookies, IP addresses and tagged urls (UTMs). However, as mobile phone adoption increased, so too did the complexity which ushered in the need for ‘cross-device’ tracking systems which could detect a person’s desktop/laptop use and their mobile phone use as the same person. Device IDs, location markers and other technical data was married together to provide more certainty, but there was still no comprehensive foolproof system. Social media advertising systems matured and marketers who mastered pixel tracking, had a large advantage. Businesses of all sizes had access to free products like Google Analytics to make more accurate decisions when it came to their digital marketing expenditure. There were still challenges when it came to visibility around organic traffic from search engines and social media.
As the digital era matured, the costs of previously expensive attribution tools reduced and Google also added features to their free product. Marketers transitioned from a focus on ‘last click’ attribution models and transitioned toward weighted, multi-touch or even regression based systems instead. CRMs and email marketing tools started to include user tracking features into their products which allowed for customer journey visualizations. Marketing automation tools amalgamated multiple data sources together and took this a step further. Business Intelligence (BI) dashboards were increasingly adopted and attribution tools such as Datorama, Google Data Studio gained in popularity. Online market research tools could be used cheaply to poll users via email, lessening the demand for market research firms. An obsession with measurability also ensued during this time and some argue a short-term ethos took root at the expense of ‘brand’ investments. As traditional media faded in popularity, the focus also shifted away from offline brand interactions and the critical influence these can have on the purchase decision.
Attribution Modelling Goes By Other Names
Marketing mix modeling (MMM), digital attribution, multi-touch attribution, marketing econometrics, Marketing Performance Measurement (MPM), Return on Marketing Investment (ROMI)
How the Attribution Modeling Process Works
This of course depends on the maturity of your brand and how sophisticated your internal systems are.
1 – Implement and configure digital data collection systems
There’s a very large range of digital tools which can track users based on their cookies and other identifiable digital signatures. A lot of CRMs and marketing automation tools have more advanced systems which can track the same user over longer periods of time to prevent holes when cookies expire or when the user switches devices.
2 – Implement and configure offline data collection systems
Existing sales teams can be a treasure trove of information. Mix these efforts with professional primary market research for the best results. Secondary data can be acquired also via a multitude of providers and other internal systems.
3 – Sales Process Mapping
While it’s common for there to be differences between customers, what you’re looking for here is identification of commonalities. Perhaps there’s one exposure or trigger that disproportionally moves the prospective customer along the sales journey. The effect of the overarching brand can distort this process. You should consider then weighting each part of the journey based on the data collected above. Sales process mapping is also called customer journey mapping. A large decision occurs in terms of deciding upon an appropriate ‘time lag’ or period of time between the first interaction and the last. This can be quite a lengthy time period with enterprise B2B sales and much shorter with convenience B2C purchases.
4 – Isolate Performance Metrics
Examine the correlation between different metrics and how they contribute toward the sales process. If there are holes in your conversion tracking data you may need to reconfigure your systems. We would recommend employing a highly experienced attribution expert during this phase (and the next phase) as this is where a high degree of error can creep into the modeling creation. If in doubt, use isolation testing and other smaller experiments to prove or disprove hypotheses. In a previous role we changed the target growth metric from gross estimated GMV, to average projected GMV to GMV + 3 opt-ins.
5 – Model Creation
Create models manually using statistical data programs like SPSS or Excel. Alternatively, feed relevant data into your visual dashboarding BI tool. Even if it is a digital attribution tool, there will be an ability to import other datasets. Next you will need to make a judgement call and perhaps test different types of regressions (Shapley Values etc). What you don’t want to do is simply tally each channel’s output against a top funnel metric like ‘reach’ or ‘digital conversions’ and see which channels produced the highest result.
6 – Model validation testing and improvements
Once the base model has been created, it should be tested regularly for accuracy and validity. Any initial assumptions should be questioned until proven. Over time, it is common for the effectiveness of certain tactical campaign activities to vary due to the associated fluctuating channel yields. As the customer base grows and changes, it’s prudent to create a separate attribution model for each segment. As the strength of the brand grows, keep in mind it will have an overarching effect on each channel’s effectiveness. Brand equity should not be excluded from any modeling. For the most accurate model, the manager should use scientific methods to examine correlation versus causation, with the latter being more pivotal in the sales process.
Shortcomings and Risk
Digital marketers can display an inherent bias toward digital media as will traditional media stalwarts. They will also tend to have a preference toward some channels over others. It’s important that channels are not only viewed independently, but also their contribution toward other channels. Google Analytics has an ‘Assisted Conversion’ section which can give more clarity on cross-channel contributions, but this will still not always exclude holes in visibility such as branded search traffic vs unbranded search traffic nor will it show by default the uplift a TV campaign might have on website traffic.
Last click vs multi-touch
There is a legacy tendency for digital marketers and inexperienced data analytics professionals to only measure the click traffic that was the last action before a conversion event i.e. ‘Last Click’. In reality, all proceeding digital interactions from that user will contribute to a conversion event include offline interactions too. It’s important to not ignore the effect of ‘view-through’ conversions also which will contribute to click traffic that may appear to originate from other channels.
A common mistake we encounter is erroneous or even duplicate data being fed into attribution models, especially on the digital side. This is commonly the result of misconfiguration of pixels, tags, goals and events. Bot traffic needs to be also excluded. Tag and event management is a highly technical area which can be overlooked in a preference for campaign execution.
You will be swamped with many different types of metrics for each channel. Choosing the right performance metrics is important here. Sometimes the top-of-funnel or top-line metrics are referred to as ‘Vanity Metrics’ and ignored in favour for lower funnel metrics. A highly experienced channel specialist should be used to critique the assumptions you have made before they are fed into your attribution model. Also choosing which metric at which point in your sales funnel will completely change up-stream decisions. For example, there is a big difference attributing a marketing qualified lead (MQL) instead of a sales qualified lead (SQL), or an opportunity versus a closed opportunity, or a new customer versus a high CLV customer. You need to be clear which metric this is and be prepared to change this over time. In the growth movement, this metric is called a ‘Northstar’.
Correlation vs Causation
It’s common for people to confuse the correlation of two variables which may appear to have a direct relationship vs one variable directly causing the movement in another. Baseline measurements, exclusion testing and longitudinal data can be very useful when trying to calculate a distinction between both.
We’ve noticed inexperienced marketers use DIY digital research tools instead of hiring market research professionals. This can work, however, internal employees should be aware of concepts such as ‘response bias’ and other market research pitfalls which can collect data which is not valid. If this data is fed into your attribution model, it’s accuracy will be compromised.
Unless the brand is large, it’s common for a small number of channels to contribute disproportionately to the sales process (Pareto principle often holds true here). Through experimentation and refinements to the marketing process, this will become more apparent over time. Exclusion testing may be required to prove these hypotheses. Search marketing may be far more relevant to a service you would need once in your lifetime and are not familiar with any suppliers, verses a convenience purchase of bottled water where long-term advertising and distribution will sway your purchase decision.
In a similar vein to the above, simply comparing certain channels against one another without considering the way they are being used is short-sighted. Often, each channel will require a specialized approach and optiimization for it to perform in the most effective way. For example, taking a 15-30 second TV commercial and using the same cut for a Youtube ad would be ineffective in a lot of circumstances unless a quick 3 second hook is used prior to the appearance of the ‘skip ad’ button. Channels are most effective when a specialist is involved, who is adept at extracting the very most from that medium.
Measurability Obsession Trap
It’s important to realize that it will be extremely difficult to attribute all marketing efforts to sales or brand building efforts. A distinction must be made between contributions toward short-term sales and longer term brand building which will pay back at a later date. Brand building efforts are often used a convenient excuse for poor campaign performance. While they may be more difficult to measure, at least some effort should go into measuring them. Even at an event, it doesn’t take much to poll a sample of attendees after the event via email or face-to-face interview and measure things like ‘brand preference’ or ‘brand likeability’ or future referral tendencies. While it may not result in instant sales, the value of these longer term measures should not be discounted. The same goes for analysis paralysis where attribution will only consider descriptive data without considering primary research or customer anecdotes. Even Jeff Bezos, who sits on a goldmine of data, claims to personally read some of the messages who customers send to his email address.
Relevance and Future
While there is currently a gap in the available measurement granularity of offline marketing efforts vs online mediums, we have already seen technological developments creeping into this space. Many TV stations and outdoor media can be bought online and their reporting has improved. As the world becomes more digital, you can expect these attribution black-spots to fade.
Attribution will always be highly relevant to any business and as the availability of performance data improves, so will the requirement to report on it.
The political costs of improved attribution should not be ignored. After an accurate attribution model reveals the extent of previous wastage, buyers remorse often surfaces within the managerial ranks. Previous decisions may be scrutinized and some will want to distance themselves from these decisions or question the accuracy of the new model.
As transparency and the respect for a modern marketing function improves within organizations, attribution will continue to be viewed as the holy grail must-have for high performing marketing departments.