The Complete Guide to Marketing Attribution

A MARKETER’S GUIDE TO REBUILDING MARKETING ATTRIBUTION

www.queryclick.com Contents 04 Marketing attribution: the basics

09 Marketing attribution models explained

The problem with current attribution 15 solutions

Why your data is the foundation for your 18 attribution success

Marketing attribution: a guide to your 26 choices

A new approach to attribution: visit-level 33 attribution

The powerful data views you need to 38 accurately drive marketing ROI

40 Closing thoughts ATTRIBUTION PLAYBOOK 3

The purpose of attribution is deceptively In this guide, we are going to take a close Overview simple: to most fairly share the value of a goal look at some of the key aspects of attribution conversion across all touchpoints that may including: have influenced that conversion. However, in the real world the complexity of touchpoints • What marketing attribution actually is and media opportunities in marketing create • Why it matters more now than ever an attribution challenge, even in a perfect • An introduction to the main attribution world of data availability. models including some of their limitations

The customer journey is now very often • Why data is key to all of this – and some of a highly complex one. And being able to the challenges and opportunities around attribute the impact of specific marketing collecting data across websites, offline touchpoints is crucial, as pressure from media, social platforms and CRM/CDP/ERP internal stakeholders - including finance and • How techniques like Machine Learning and the boardroom - to link marketing to revenue, Deterministic and Probabilistic matching and prove ROI intensifies. Unravelling the can help overcome limitations in current impact of specific touchpoints on conversion attribution approaches is priority number 1 for marketers. • How all of this provides effective attribution In fact, its relative importance was summed at a Channel, Campaign and Impression up when Google’s marketing evangelist, level to better focus and optimise your Avinash Kaushik, described solving attribution marketing efforts so that they are more to identify incremental value across all media closely aligned to the customer journey channels as the biggest problem facing itself analytics in 2020. So, attribution matters. Marketing attribution: the basics ATTRIBUTION PLAYBOOK 5

Marketing So, what is marketing attribution: the attribution? Display Put in its simplest context, marketing Paid Search basics attribution is the process of determining Email which marketing touchpoints and activities Social are contributing to conversions – which could come in a variety of shapes and sizes ??% ??% ??% ??% from sign-ups, to downloads, to purchases of a product or service – or some other An attribution model, or combination of meaningful conversion types. models, is a defined set of rules that helps you determine how conversion credit is given In a complex marketing environment where to these different marketing touchpoints marketers are using an ever-increasing mix of across the customer journey. They come in all channels simultaneously it is about unpicking shapes and sizes – but are split broadly into the impact that channels, campaigns and two types: single-click and multi-click. But even individual creative executions are having more detail on those later. across media like Display, Paid Social, Search (Organic & Paid), email and offline media. What’s important to state here is that, if you don’t have effective marketing attribution in place, then it’s a fairly safe bet you don’t know what is – and just as importantly, what isn’t – working across your marketing mix. So accurately assessing and improving “true” marketing performance isn’t really possible for you. ATTRIBUTION PLAYBOOK 6

Why marketing attribution matters. More than ever before.

So, attribution matters - and it always has. and digital is a growing part of the mix. By 2021 the volume of data in silos involved in At the most fundamental level the sheer delivering just digital marketing, will have scale of current investment in marketing increased in complexity 32-fold compared to demands effective attribution. In the UK alone a decade previously. marketing spend will top £21.3bn in 2020 The figures are staggering and, as a result, many businesses are struggling to sustain Marketing spend Vs Silo complexity (Global, US$ BN.) growth while maintaining or reducing media $900 80% spend budgets. Which – and to echo Avinash $800 70% Kaushick - makes solving attribution the $700 60% largest opportunity any marketer can engage $600 50% with today. $500 40% $400 But there are also a number of other key 30% $300 reasons why having an effective marketing 20% $200 32x increase in silo complexity by volume attribution solution in place is more important 10% $100 than it has ever been, as it enables you to: $0 0% 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023

Traditional Ad Spend Programmatic Ad Spend Revenue Amazon Revenue Other Digital Ad Spend Traditional %age of Total Digital %age of Total

Source: eMarketer Global Digital Marketing Report & Statistics. ATTRIBUTION PLAYBOOK 7

Link marketing spend directly to growth and Prove your value to the business and build revenue stakeholder trust

Which is non-negotiable now. The pressure Effective attribution builds credibility and for marketers to spend accurately, efficiently trust with the key stakeholders closest to the and to prove marketing effectiveness in marketing function including the boardroom, driving growth is greater than ever. And finance – and, perhaps most importantly, is only going to intensify, so directly tying sales. Securing the buy-in needed to support activity to revenue is no longer a “nice-to- your marketing vision and fast-track future have”. It’s essential. plans.

Keep full control over the nature of marketing Secure and ring-fence budgets based on spend sound marketing analysis

Not only is there broad evidence that internal Marketing budgets are under the microscope

“Internal stakeholder pressure restricts my option to stakeholder pressure on spend - from outside more now than any time previously. By employ marketing activity that has a longer payback of marketing - is increasing but there are also accurately being able to show how marketing period than last-click measures.” worrying signs that it is directly influencing is contributing you have the capability to Strongly agree 24% the nature of spend. In fact, research carried ringfence and even grow your budget. out by QueryClick points to the fact that Somewhat agree 43.5% 67.5% of Marketing Directors report that Improve your marketing ROI internal stakeholder pressure actively restricts No opinion either way Ultimately, effective attribution helps you the option to employ marketing activity with 19% drive improved marketing ROI by finding the longer term payback. Somewhat disagree parts of your program that are working - and

11.5% those that are not. This the allows you to Strongly disagree redirect spend to where it is going to have the 2% Base: all respondents, n=200 biggest revenue impact. ATTRIBUTION PLAYBOOK 8

Reduce reliance on perceived short-term Take control of marketing complexity media ‘fixes’ Finally, better attribution enables you to For many businesses, media like Pay-Per Click take control of what is ultimately a more is the main channel. In fact, everyone loves it challenging and complex media environment. because it provides immediate results. Switch it on and revenue flows when you spend. But Where customer journeys are taken on and it is no longer the panacea it once was. High offline, and research can be done across a spending competitors are not only eroding multitude of devices including smartphones, market share but driving up CPAs. So, ROI is tablets, work and home PCs, and mobile, being driven downwards. Effective attribution effective attribution enables you to bring opens the door to wider media opportunities, together the data silos across this complex with longer payback cycles – because you environment to define a single, data-driven can pinpoint the data you need to make the view of the customer journey. business case for it more effectively. And confidently build your marketing approach around it. Marketing attribution models explained ATTRIBUTION PLAYBOOK 10

In this section we are going to take a closer Single-channel attribution Marketing look at attribution models. These models are relatively simple, and work attribution One of the key things to understand upfront on the basis of allocating 100% of the credit about attribution models is that there for a revenue conversion or other customer models isn’t a definitive, one-size-fits-all answer behaviour to a single touchpoint with the for everyone. And, in practice, identifying explained customer, either at the beginning or the end which model is right for your brand or your of the customer journey. All other touchpoints business is going to be impacted by a range in between are ignored for attribution of different factors from the stage, scale and purposes. maturity of your marketing activity to how

much time, budget and effort you are able to Works well Where you have a simple funnel apply to it. environment with very few marketing activities. And businesses who are scaling from a low base may well employ them early in their growth. Below we take a whistle-stop tour of each in

turn. Including where they might work well Possible limitations However, they oversimplify and where they also have limitations. and ignore the impact of multiple touchpoints on complex, longer customer journeys.

In broad terms they split into two broad types There are two types of single channel models – single-channel attribution and multi-touch – First and Last-touch attribution. attribution. Let’s look at these in turn.

ATTRIBUTION PLAYBOOK 11

First-touch attribution Last-touch attribution

100% 0% 0% 0% 0% 0% 0% 100%

Built around the premise that the first Similar to First-touch attribution, Last-touch touchpoint with the customer is the crucial is predicated on allocating 100% of the credit determinant in a customer’s decision to do for conversion to a single touchpoint. business with you. However, in this case it allocates credit on the So, 100% of the credit for revenue is directly basis of the last touchpoint prior to purchase attributed to the first interaction in the or revenue. And similarly, it gives zero credit customer journey. For example, if a customer to other interactions along the way. comes in on an eBook download it gets Works well Where you have a heavy focus on all the credit – regardless of all the other optimising conversion around the bottom end of the touchpoints in between that and the sale. funnel.

Works well Where you have a simple, early stage Possible limitations But again, it’s going to be marketing model that is heavily focused on growth and much less appropriate for analysing longer and more quickly scaling your business. And where there is a complex journeys accurately. strong focus on growing top of funnel. So, taken together, Single Channel attribution Possible limitations However, it is going to be models do have their benefits – and you may much less appropriate where you have a lot going on in your marketing funnel and a more complex and well start your journey using them in early extended buyer journey (of 90 days plus). growth stage scenarios in smaller businesses ATTRIBUTION PLAYBOOK 12

– but it is also highly likely that even an initial Linear spurt of growth will see you outgrowing their usefulness. As your marketing evolves and matures, your need for more sophisticated attribution is likely to follow suit.

25% 25% 25% 25%

Multi-touch attribution Very often you don’t want to give single models channels or campaigns all the credit for conversion and a Linear model moves the These models take a broader view of principle on by applying equal weighting and attribution and attempt to place a relative credit to all contributing touchpoints. value on the impact of all touchpoints across Works well Where you are looking to measure a the customer journey, although these are campaign holistically and get a feel for the customer weighted differently depending on the model journey, including seeing which channels contribute you select. across it

Possible limitations Where you already If you are operating with a more complex understand the customer journey and have a more model, longer cycles and a multitude of detailed need to dig into the data on individual channels to understand which ones are critical so you can focus channels it’s likely that this is where you find your effort on them yourself gravitating to. Let’s take a quick look at each of the commonly used multi-touch models in turn. ATTRIBUTION PLAYBOOK 13

U-shaped Time decay

40% 10% 10% 40% 10% 15% 20% 25% 30%

Takes view that some touchpoints have more Also weights each touchpoint differently and influence on the way to conversion than is driven by the assumption that touchpoints others and scores them separately. Here closer to sale should get more credit. The first touch and lead conversion matter most assumption being they have a heavier impact and are both heavily weighted with 40% of on conversion as the buyer moves down the the credit being applied to each. While the funnel. remaining touchpoints are given an equal Works well With a short sales cycle less than share of the remaining 20%. 90 days and if you are looking to identify what drives conversion in the lower funnel Works well Where you are looking to identify what drives channel acquisition or have heavy focus on Possible limitations Not so appropriate for longer nurture B2B customer journeys where it ignores the fact that customers can make important decisions earlier in the Possible limitations If you are using Google process Analytics but have a buying cycle that extends over 90 days you can potentially lose the data view you need ATTRIBUTION PLAYBOOK 14

W-shaped Custom model

This is where you take the view that you need a very personalised view for your marketing needs. And the discussion here opens up into a wider consideration of more complex 30% 5% 30% 5% 30% attribution models that include Statistical An extension of the U-shaped model with or Data Driven Attribution, Media Mix/ the addition of weighting for the opportunity Econometrics Attribution and Probabilistic stage. Here first touch, lead conversion and modelling. Which we will return to in more opportunity creation share 30% credit each, detail later on. with 10% being shared out between the other Works well Where you have a very direct need to touchpoints. tie things back to your end goals and need the flexibility to create tailored weightings to create a highly refined Works well Where you are looking to identify and accurate view of things touchpoints at the 3 critical points that help you build audience, create leads and then drive to conversion Possible limitations Requires an adequate allocation of time, thought and resource to make it Possible limitations If you are using Google effective. So not to be undertaken quickly or lightly Analytics but have a buying cycle that extends over 90 days you can potentially lose the data view you need For now, it’s probably just important to know that, in practice, this is going to be a blend of elements from the models above. Customised in a way that works for your own marketing scenario. The problem with current attribution solutions ATTRIBUTION PLAYBOOK 16

The section above provides a good overview attribution model for Google Analytics, and The problem of marketing attribution models, each with is a complete oversimplification where 100% their benefits and limitations. However, the of the value of the conversion is given to the with current reality is that existing flawed attribution last marketing engagement. In essence, this attribution models are part of the problem marketers ignores all the previous points in the buyer face. journey: almost the same as applying no solutions attribution at all. For a couple of key reasons: Lookback windows are too short: Even Historical and oversimplified views: At tools like Google 360 have limitations due the most fundamental level almost all current to their short lookback window and only attribution models are historical in nature partial attribution in relation to what is very and all the evidence is that they are woefully often a long and complex customer journey – inadequate in terms of doing what they where the customer is influenced by a broad should. For example, Last-click is the default spectrum of media.

“Attribution is broken and it’s screwing up your entire approach to digital marketing” ATTRIBUTION PLAYBOOK 17

So, rebuilding attribution – in a way that reflects the complexity of the customer journey – has to be a key goal for all marketers. Unfortunately, “in the wild” the vast majority of marketers - in fact, according to research by QueryClick, over 90% of those responsible for improving marketing ROI in the UK - are unable to get to sophisticated attribution models such as these. Or even simple rules-based models such as First-click, Position Based or Time-Decay Weighted. In an effort to provide better outcomes than a basic Last-click view.

But flawed attribution models are only one part of the story. It’s also highly likely that 80% of your core analytics data is wrong too.

So, the quality of “data in” to your attribution model has a huge bearing on the quality of your analysis. Which is why we are going to DOWNLOAD take closer look at this in the next section. Why your data is the foundation for your attribution success ATTRIBUTION PLAYBOOK 19

Our extensive work with clients around data So, the seeds of success in effective Why your enrichment consistently tells us that poor marketing attribution are sown firmly in the models are only one part of the issue with way that you collect the data to feed into your data is the attribution. model. foundation for Just as important is the quality of the your attribution underlying data that you are feeding into your Challenges and attribution model – and in purely layman’s opportunities in data success terms if you put rubbish in, you will simply get rubbish out. Let’s take the practical example collection of Google’s approach to collecting the vast At this stage it is worth considering some of majority of digital marketing data which is the challenges and opportunities around data based mainly on a pixel plus cookie-based collection across a number of channels and approach. The problem is that measurement techniques including: in this model is rudimentary and based on tracking a device accessing a specific web • Websites property. Not around identifying an individual • Offline media on the complex, multi-device journey they are • Mobile and social on. • Fingerprinting & 1st Party Enrichment

• CRM, CDP & ERP ATTRIBUTION PLAYBOOK 20

Limitations in the cookie/pixel And this is the point at which the world’s approach most commonly used analytics packages start to break down and generate incorrect The vast majority of digital marketing data or incomplete data. The cause? What is is what is called ‘deterministic data’ and is being measured is a device accessing a web collected using a pixel working in tandem property. Not the person themselves. Who with a cookie. With the exception of log-file will use multiple devices in the course of even analytics, which uses data collected directly simple transactions. by the server hosting the web property. Cookies are supposed to enable ‘joining’ HTTP Request of multiple sessions generated by devices

Third Party Server interacting with a web property but, in reality, they do a poor job, as is shown in the

First Party Server following chart:

Assigned ID Requested Content Daily attributable data (i.e. Repaired broken Tracking Cookie sessions & enhanced cross device)

Corvidae A variety of data including:

• measurement of dwell time

• previous site visited Google 360 • any ad campaign that may have been encountered by the visitor

• any repeat visit made by that same person K £2m £4m £6m £8m £10m £12m £14m is all captured in this way. Non-Attributable Data Attributable Data ATTRIBUTION PLAYBOOK 21

This chart shows that using deterministic journeys involve a plethora of media and analytics data only (i.e. pixel and cookie it’s worth now considering some of the data) to create a picture of the individual’s challenges – and opportunities – in data behaviour behind their multiple devices. Even collection across mobile applications, CRM/ where advanced session joining ERP and offline, to build a clearer view of TECH FOCUS are employed - as with Google 360 - attribution across the customer journey. Random forest modelling generates data that is around 80% incorrect.

Random forest is a regression technique that uses multiple decision trees and merges their The alternative approach being used by Difficulties in integrating offline predictions together to get a more accurate and Corvidae in this example, is a Machine data stable prediction rather than relying on individual decision trees. Learning model approach, which takes the For most businesses, marketing activity core deterministic data returned into the tree tree tree 1 2 n is a blend of offline and online activity. clickstream, and rebuilds it using random Offline increasingly offers rich data as forest probabilistic modelling (see Tech focus) measurement moves beyond traditional and regression analysis. panel-based approaches to incorporate

k1 k2 kn This approach is not applied by existing Web Beacon technology, opt-in Wi-Fi Tracking Analytics suppliers such as Google, IBM and and RFID data. These technologies, deployed In a Marketing Attribution context – and when compliantly, enable good data mapping to combined with the Machine Learning capability Adobe, as compute intensity is high and initial of tools like QueryClick’s Corvidae platform model development has until very recently digital marketing data. - it offers the possibility of taking deterministic involved manual ‘tuning’. data gathered from the clickstream and applies The largest marketing reach touchpoints Machine Learning to rebuild it in conjunction with techniques like device characteristics to This analysis highlights why even core 1st outside of store are linear TV and Radio significantly improve attribution and identify party web analytics data for web properties activity. Which typically offer either panel- individual customers journeys. As well as led, econometrics data. Or increasingly providing the predictive data insight needed to fails to provide good data to base marketing target incremental growth by using propensity to decisions on. But more complex customer purchased, direct from manufacturer data - convert models to drive revenue. ATTRIBUTION PLAYBOOK 22

for example from Smart TVs - for inclusion in Challenges and opportunities in marketing mix analysis. mobile application & social data collection There are challenges with econometrics data including the inability to look beyond channel level impact as part of A/B channel or messaging strategy adjustment, a lack of immediacy in terms of delivering data outcomes and the fact that they operate in a silo to all other data collection strategies. Which creates an increased issue around cannibalisation. In addition to being expensive and less appropriate for smaller budget media spends.

However, if they are applied to enrich Data collection in-app or by a social an already unified marketing strategy, platform is more tightly associated with an econometrics can be a valuable tool to unlock individual who is required to login. Which customer behaviour, environmental influence, creates a much more robust measurement competitive intelligence and meaningful environment. signals for targeting. However, the challenge today with activating

mobile app data is building good quality data to connect it to. For example - if purely deterministic data collection is being used in the wider web, then the individual’s app ATTRIBUTION PLAYBOOK 23

based activity which it is being “joined onto” APIs - such as the Facebook one - and ingest will remain 80% incorrect for the reasons we their data into a wider analytics view. highlighted in the section above.

In addition, practical issues around how Fingerprinting & 1st Party social platforms utilise “share” functionality Enrichment of Data across web properties. Combined with the One associated data collection technique - fact that significant volumes of retargeting which is highly relevant to social profiling - is TECH FOCUS activity occurs against social platform fingerprinting. Essentially a way to combine “Walled Gardens” on Social pixels - essentially short-cutting the need to certain attributes of a device to uniquely Media platforms understand the customer journey and simply identify it. targeting them based on their demographic The broader definition of a Walled Garden is a or sociographic profile and proximity to a closed Ecosystem in which all operations are There are two common forms of controlled by the ecosystem operator. conversion event - means individuals are fingerprinting: device and canvas. already some way down the conversion path. In media terms, this is very much the approach that highly successful social platforms like Device-led fingerprinting Facebook, , YouTube etc. are taking in Social platforms also typically take a ‘walled marketing terms. By keeping their information garden’ approach to data availability for and data analytics firmly closed for privacy but - This approach uses “hit-level” data from a analysis outside of their own analytics just as importantly, commercial value reasons. In person’s device to determine its uniqueness, practice, this means reporting only aggregated (see Tech Focus). Which is geared towards and then match it to subsequent visits. The data with no identifiers to aid the process of enabling native advertising to their users. identifying individual customer journeys. level of data available is very high and this is a very accurate way of joining multiple It is only by using the power of Machine Learning However, these platforms offer a potentially sessions together when included in the type in Attribution platforms like QueryClick’s Corvidae rich data view of customers and it is now that you can begin to use sophisticated datasets of probabilistic models we outlined earlier. from social – using data on page views, location possible, using sophisticated Machine details, device details, on-site behaviour etc. - to Learning techniques to ‘unbundle’ data from derive specific user journeys. ATTRIBUTION PLAYBOOK 24

Canvas fingerprinting All of this means that 3rd party data collection - essentially data collection from domains A single device offers a wide range of data that are not the web property domain - will when preparing to render a web property and become ineffective, unless they switch to 1st when combined with canvas fingerprinting party data collection processes. it can become entirely unique. Although it does have potential data compliance issues 1st Party Data 2nd Party Data 3rd Party Data to address.

Canvas fingerprinting data can also be collected via the delivery of ad inventory in the wider digital ecosystem. However the bulk of canvas render requests today are for ad visibility verification purposes. Facebook also has tracker pixels performing the same fingerprinting across 30% of the top 1,000 Leveraging the wealth of data in websites by visits globally. Giving their ad your CRM, CDP & ERP targeting data deep richness. Similar to mobile application data collection, However, Pixels that are set from outside a Customer Relationship Managers (CRM), web property’s domain are already blocked Customer Data Platforms (CDP), and by default for Safari users. And so, by default Enterprise Resource Planning (ERP) in some also on the vast majority of iOS devices. In senses make data collection and achieving addition, the world’s most popular browser, data accuracy simpler. As customer Chrome, which is owned by Google, will be identification or sign in is usually a pre- moving to the same default block position requisite. through the course of 2021. ATTRIBUTION PLAYBOOK 25

However, these systems’ attempts at ‘joining’ increase the volume or frequency of or crossing over customer IDs into wider purchases web analytics, is also affected by the 80% As well as enabling a deeper understanding inaccuracy issue highlighted in an earlier of marketing ROI. So, exploring what is section. possible here is well worthwhile.

However, there is a wealth of potential data available to build into your analysis including: Finally, be careful with data sampling and vendor selection CRM & CDP data Across email, SMS, targeted ads, phone calls Before signing off on this section on the or direct mail challenges and opportunities of data collection, it is worth making a small note

ERP data on the impact of sampling and vendor Across clubcard schemes, Point of Service selection. There can be issues with “sampling (POS) transaction data, inventory and stock Inventory Service in” analytics data which differ from vendor availability data, returns and fulfilment cost to vendor. So, vendor selection when data Product considering all of your data collection needs Sales is a key part of getting attribution working for That can be complemented by 3rd party data you. Financials ERP that can be associated with customer profile

HR data to help drive:

• high quality customer segmentation and

MRP cohort analysis Purchasing

CRM • marketing automation to improve conversion and engagement and to Marketing attribution: a guide to your choices ATTRIBUTION PLAYBOOK 27

Once you have identified and gathered your Without rebuilding the clickstream, Marketing data, how you apply attribution to it is equally however, they suffer from increasing levels important. Marketers today have 4 options of inaccuracy as we can see in this side by attribution: a to get attributed data on their marketing side comparison using the same dataset as guide to your activities. in the chart below – where attributable data identified by Google Analytics 360 was a • Single/multi-touch or rules-based fraction of that uncovered by QueryClick’s choices attribution Corvidae platform. • Statistical or data-driven attribution

• Media mix/econometrics attribution Daily attributable data (i.e. Repaired broken sessions & enhanced cross device)

• Probabilistic modelling GA360: Time Decay Weighted

Let’s take a quick look at each in turn.

GA360: Last Click Non Direct

Single/multi-touch or rules-based attribution Corvidae Attribution

We looked at these models earlier on and K £5M £10M £15M they include a range of options including Attributable Unattributable First-touch, Last-touch, Linear, Time Decay When used in combination with and W/U shaped models. segmentation, channel attribution can still These approaches are segmenting data from be powerful and enables budget reallocation core clickstream data into scoring for each across channels for paid marketing activity channel touchpoint and therefore typically that will affect overall marketing impact. offer only channel attribution outcomes. However, it offers no insight for individuals, ATTRIBUTION PLAYBOOK 28

preventing the use of this attribution type for available in digital ad tech like Google Ads automated customer acquisition or insight and is based on using vendor specific tags on individual impressions. Or the impact of – which have the very real limitation of not specific pieces of marketing content on that being aware of data collected elsewhere of individual and their subsequent behaviour. course.

Conversion credit model = Data-driven Path length > 2 touchpoints

Statistical or data-driven Channel

attribution 1 Organic search 100% Direct x2 0%

2 Organic search 50% Email 50%

Increasingly marketers have access to ‘data 3 Direct x2 100%

driven’ models which essentially apply 4 Organic search 100% Direct x2 0%

channel attribution rules that are determined 5 Referral 100% Direct x4 0%

by the data available to them. Instead of 6 Referral 100% Direct x5 0%

rules-based assumptions about channel mix, 7 Paid social 100% Direct 0%

data-driven models determine the model’s 8 Organic social 100% Direct 0%

weighting based on existing data. 9 Paid Social x2 100%

10 Email 100% Direct 0%

Conversion Beginning of path to conversion

Paid Search 31% 12% 8% — 171,650.00 Channel attribution using data-driven approaches allow for Display 32% 16% 5% 6% 55,505.25 Referral — — 24% 47% 13,824.46 path grouping like this in reports. Organic Search 24% 24% 8% 29% 2,647.27 Direct — — — 48% 23.02 — 9% — — 9.50 Data Driven channel attribution in GA and

N/A 0% 10% 20% 30% 40% 50%+ Adobe suffers fromthe same data quality challenge as rules-based options described Even the free version of Google Analytics now previously – the modelling it relies on can offers this functionality, and it is increasingly only work with the supplied data. Which as ATTRIBUTION PLAYBOOK 29

we have seen is inaccurately representing the But, again, should not be used to assess individuals behind the clickstream. And so, revenue outcomes. the data-driven models are showing only part Orders Assisted Orders of the real picture. Last Touch | Visitor Marketing Channel 2,009 2,009 Feb 1 Feb 28 Feb 1 Feb 28

1. Email 624 31.1% 480 23.9% There is still value and insight to be gained 2. Referring Domains 542 27.0% 355 17.7% 3. Social Networks 217 10.8% 141 7.0%

from these models, however. As they can 4. Paid Search 190 9.5% 136 6.8% give an understanding of the contribution 5. SEO 179 8.9% 103 5.1% 6. Display 174 8.7% 134 6.7% of channels that are typically poorly served 7. Direct 42 2.1% 42 2.1% 8. None 41 2.0% 41 2.0% by Last-click - Paid Social in particular - in your marketing mix. As an overall measure of volume of interactions. Media Mix Modelling (MMM) / Econometrics This means that relying on data-driven models in web analytics today is not advised This approach applies statistical techniques for assessing revenue outcomes. Without to attempt to find incrementality in large data additional processing from the likes of sets using mathematical techniques. Often QueryClick’s Corvidae platform. relying on integrating panel data for offline activity. Other types of statistical attribution are available in digital marketing, for example In all cases, the object of MMM or assisted conversions reports like the one Econometrics is to isolate a test case that is shown here from Adobe. That can be usefully to be measured when it is applied in two or applied to get a better understanding of the more variations. And the outcome is often volume impact of impressions on conversion expressed as a measure of ‘incrementality’ to paths. aspects of the media mix: ATTRIBUTION PLAYBOOK 30

For example, a business may ask what the Profit Revenue Profit Expense Net cost per acquisition is for an offline catalogue. Factor In creating a test to answer that question, Catalog + Email $12.00 30% $3.60 $1.01 $2.59 Catalogs Only $10.00 30% $3.00 $1.00 $2.00 different mixes of media channels will be Email Only $8.50 30% $2.55 $0.01 $2.54 used for cohorts of customers and the overall No Catalogs $5.00 30% $1.50 $0.00 $1.50 acquisition outcomes measured. This will No Email involve mix variations including: A simple experiment example of test variations and measured • the use of catalogue only metrics from an A/B or Media Mix test. This example shows a 2% incremental benefit from running offline catalogues with an • catalogue plus some marketing channel email campaign. Source: kaushik.net • and the same channels with no catalogue In this way we can understand that media mix modelling is not just a case of running A/B By comparing the results of these variations, or multivariate testing - it is a science with a an overall value can be determined for the great volume of subtleties that can be used to incremental acquisitions made thanks to answer a particular question for a particular the catalogue itself. And that result will be period of time. enriched with data about:

• the cost of delivery Due to the complexity and volume of data required to create balanced tests and models • operational cost that are statistically sound, media mix and • marketing costs and so on. econometrics studies done correctly are an This gives a good and rich level of detail expensive, ongoing proposition. That should to the insight gained from the media mix only ever be run in parallel to attribution modelling. Ultimately offering an actionable activity that is intended to allow tactical outcome that is statistically true for those optimisation day to day. tested cohorts for the period of time the test was run. ATTRIBUTION PLAYBOOK 31

TECH FOCUS

Markov Chain modelling In short: A/B Testing and Econometrics A very broad brush explanation of a typical

Markov Chain is an example of a sophisticated, studies should be used infrequently to inform Shapley approach is to find an average probabilistic modelling approach that looks at strategy. While attribution should be used to incremental value for a particular channel individual events in the customer journey and inform day to day decision-making in service touchpoint by looking at all conversion identify patterns of user behaviour to build journey profiles. of optimisation. Which leads us to our last paths that exist, their conversion path rates option: probabilistic modelling. and their mix of channels. Comparing the By evaluating the probability to move from one different mixes and conversion values brings state to another in the customer journey. you an average incremental value for each Probabilistic modelling 0.2 channel for each conversion path. Which is then aggregated into the reporting you see in SEO Recent advances in Machine Learning and the analytics interface. cheap deployment in SAAS architectures 0.2 0.6 have allowed probabilistic data processing to 0.2 0.1 Another common model is a Markov Chain be applied to data challenges in marketing 0.7 model. Markov Chains are named after day-to-day. 0.1 PPC Social 0.6 the mathematician Andrey Markov, and in 0.3 essence, enable modelling a sequence of v(A) = ∑ C(S), C(S) ≥ 0 events in which the probability of each event For example, if a website visitor is identified as depends only on the state attained in the coming in on Organic Search, the Markov Chain S○A . would then understand and look at the other previous event. channels the visitor could have come in on and the likelihood they would come back from For example, Google Analytics uses Shapley And practical application of the Markov Model Organic. It would also look to see if they would Modelling for their data-driven channel come from Paid or if they would never come back models. This is based on Game Theory, and can be powerful. Particularly when you are or if they would transact. approaches marketing as a game, where interested in channel level attribution. each channel is a player in the game, and This type of technique can be used to predict the set of all players/channels can be seen as behaviours that lead to conversion, and the data Let’s look a practical example. In the case though working together in order to drive the can be fed back into the marketing model to below, a Markov Chain based approach optimise marketing performance. conversions. ATTRIBUTION PLAYBOOK 32

was used in conjunction with the Corvidae For example, using Markov chains for platform to assess and more accurately marketing conversion paths will find lots of 100% probability value chains due to the very attribute revenue contribution by channel for long tail of customer conversions that exist for a well-known Online Fitness retailer. all businesses - and are limited in their ability to look at more granular value answers than As you can see below, compared to Last simply channel value. Click Non Direct (the rules based model used by GA 360 and Adobe by default), channel Most importantly however, both are limited attribution has changed by £39.9m. in that they must be re-modelled to provide

Top 6 Channel attribution comparison by channel more granular outcomes than just channel

£50m attribution. Which is computationally very £40m expensive - and which also causes these £30m

£20m models to run into data volume limits which £10m prevent their practical functioning. Small K Affiliate Search PLA Search Text Direct Email Organic Paid campaigns would not model with either First Touch Last Touch Last Touch Non Direct Corvidae (GA360/Adobe) (Markov App) approach for example. And understanding Given total turnover is only £136.7m across the engagement an individual visit has - and the measured channels, this radically alters that individual’s probability to convert at the ROI implications of media spend, each individual stage of their own, unique particularly in Paid Search and PLA (Paid conversion path - is not being modelled. Search Shopping Ads). But what if it could be? Fundamentally both approaches, Shapley and Markov, are predictive models and have much in common, however they both have drawbacks. A new approach to attribution: visit-level attribution ATTRIBUTION PLAYBOOK 34

It is now possible to take an approach Stage 1 – Rebuild A new approach that learns from econometric modelling techniques but is focused on providing near- Using previously developed session stitching to attribution: live data for tactical marketing insight using Machine Learning processes, data from visit-level bleeding edge data science. And which almost any marketing channel may be added leverages Random Forest Machine learning to the Corvidae system for inclusion. Because attribution techniques and session-stitching capability the system first creates a customer journey plus the capacity to join any offered data from the ground up using raw clickstream Advances in technology have source to fix the incoming clickstream data data, it is then possible to further ‘stitch’ in marketing content that overlaps with made this a reality. being generated by analytics platforms. And join data to any marketing channel online or an individual’s journey. Where supplied off. data overlaps in time, geolocation or other fingerprinted patterns match. This - visit-level attribution which is used by Corvidae – is the creation of a fully attributed This scores each engagement a visit has - map of a customer journey. With a scored and the probability of its next engagement value against every piece of marketing - based on existing data. This completely collateral that individual has engaged with. rebuilds core marketing data from the ground up around a true picture of customer And has 3 clear stages as follows: behaviour.

• Stage 1 – Rebuild Providing the basis for building a single view • Stage 2 – Unify of each individual’s journey to conversion.

• Stage 3 – Attribute

Let’s take a closer look at each in turn. ATTRIBUTION PLAYBOOK 35

Stage 2 – Unify where geo and time data – from a range of offline media sources - overlaps with the This probabilistic approach then enables 500 measured markers held against each joining of data from offline activity, store customer across a two year lookback window. activity and even enables the joining of third party ‘walled garden’ digital data such as data Where predicted customer conversion from Facebook, Google Ads or YouTube. behaviour is measured to have changed for a customer, and they have likely been exposed So, how does this work in practice? to an offline marketing event - a TV ad, or instore activity - then that activity is ‘stitched’ Let’s take an example – where rebuilt data in into their conversion journey. a platform like Corvidae can be accurately mapped very precisely to offline events When attribution modelling is subsequently applied that offline event is allocated value Uplift by Station (Most impact at least 50 slots) just as a digital channel would. Which opens

8 Station film4 up the possibility to quantify offline activity dave 7 channel 5 discovery science effectiveness, and allows digital and offline itv hd 6 itv sky 1 to be held to the same, unified, account for more4 5 ch4 ch4+1 performance optimisation. 4

Mean uplift (absolute) 3 And the impact is significant in terms of the 2 insight it can provide. 1

0 1 2 3 4 5 6 7 Minutes past ad display As a result of joining offline data to online, QueryClick were able to identify significant If supplied offline activity data is sufficiently granular, platforms like Corvidae can be used to apply further performance opportunities for spend reallocation in future segmentation. ATTRIBUTION PLAYBOOK 36

TV and Radio media buying campaigns for Stage 3 - Attribute a major life insurance customer. In fact, as the table below shows, 39% of TV spend was While the customer engagement paths identified as ineffective and available for generated by the Rebuild and Unify allocation to other channels for no negative phases are greatly improved data, they still revenue impact. represent a vast and complex relationship which must ultimately be scored to give TV Spend Effectiveness: Reallocation Opportunities a single attributed revenue outcome for Ineffective Spend

Total Yearly TV Spend £2,264,000 each engagement point. Considering the Savings from Time adjustment £430,160 incrementality or otherwise of a piece Savings from Station adjustment £311,753

Savings from Day adjustment £152,209 of content on one particular prospective Total £3,158,122 customer journey.

More starkly, no correlation between radio And that is a complex task. activity and conversion impact enabled £2.42m of radio media spend to be For example. Consider the challenge reallocated with no negative revenue impact. of valuing the incrementality of a single impression for a single piece of content - The rebuilding & joining phases above create which may have many millions of impressions the most complete possible journey for every - on a single prospective customer. Who may well behave differently both before, and after, identified individual, including an accurate that impression than every other prospective attribution score for each and every content customer. And will almost certainly be in a interaction for each and every individual. different conversion phase of their conversion path - or non-conversion path - from every other prospective customer again!

ATTRIBUTION PLAYBOOK 37

But it is possible, by applying advanced AUROC Graph - Retargeting Model

Machine Learning – and neural network based 100 – techniques which are predictive in nature. 9 80 Like the ones we have outlined to this point, 8 7

and that are used in sophisticated attribution 60 6 software like QueryClick’s Corvidae platform. 5 4 40 Conversions (%) Conversions

Engagement Level 3 So, what does that look like in practice? 2 20 1

0

Let’s take a closer look at the example 0 0 20 40 60 80 100 of accurately assessing conversion in % of Visitors - Ordered by Predicted Conversion Probability

retargeting activity for a major life assurer. In Corvidae’s ability to see the whole conversion picture of each this case, Corvidae made available a Machine individual allows very high performance retargeting. Learning based, predictive view of customer Prospecting for incremental revenue via prospect conversion for ingestion into the programmatic display was also highly client’s Google’s DV360 platform. In order to successful, delivering a 3.8:1 ROI, and overall run an A/B comparison against the in-built Corvidae’s automated customer acquisition ‘data-driven’ ML targeting offered by the feeds returned 5.8:1 ROI against media spend client’s existing platform. in the highly competitive life insurance sector in the UK. And the results were outstanding. Corvidae

was able to deliver 4x more conversions at Programmatic Display Corvidae API Performance ½ the CPA of Google’s model. Delivering a Spend Revenue ROI 50.5:1 ROI and clearly demonstrating the Retargeting £4,951 £254,770 50.5 : 1 Prospecting £110,000 £529,900 3.8 : 1

value in taking a predictive, neural network Total £114,951 £784,670 5.8 : 1 based approach to attribution. The powerful data views you need to accurately drive marketing ROI ATTRIBUTION PLAYBOOK 39

The powerful By building powerful visit-level attribution Channel Level Attribution data you are able to unravel the complexity of Removes 3rd Party Cookie dependency the wide variety of touchpoints on individual Optimisation of media spend data views customer journeys and accurately attribute by channel. ONGOING ROI the correct level of impact to each. • Spend reallocation (manual) you need to • Predictive Paid customer acquisition (automated) accurately drive This enables you to make the direct link between marketing spend and revenue, providing the capability to significantly marketing ROI Campaign Level Attribution improve your marketing ROI by making Optimisation to individual informed judgements based on sound data campaign level. analytics. • Granular campaign performance data (manual) ONGOING ROI • Bid adjustment (automated) It also opens up access to a broader, tiered level of attribution at a Channel, Campaign

Store and Impression level. As shown to the right. Impression Level Attribution OOH Organic Lookalike (Voice) (RTB/Prog) TV/R TV/R Paid Shopping • Optimisation of social OOH Store TV/R (Digital) display and video Paid Social Organic Display (Brand) • Optimisation of individual Paid Social Organic creatives Paid Shopping (Generic) ONGOING ROI

Awareness Consideration Purchase Service Loyalty (See) (Think) (Do) (Care) Expansion

Paid Social Organic Paid Shopping Remarketing RTB Programmatic (Generic) (Paid Search) (Brand) Organic Corvidae provides the ability to work at both Store Direct mail Earned Paid Search (Programmatic) (Content) (Brand) strategic and tactical levels to improve your Email (Referral Offer) Remarketing Email marketing programmes from the top down (RTB/Prog) (Rate & Review) Work Mobile 3rd Device In Person and bottom up. Closing thoughts ATTRIBUTION PLAYBOOK 41

A range of factors from increasing media on solving the attribution issue for leading Some closing complexity to the need to directly link brands including Tesco, Wiggle and Vitality marketing to revenue mean marketing Life. thoughts attribution is the key issue for marketers to solve right now. But our data shows marketers Reaching beyond Google 360 and Adobe lack access to effective attribution and are Analytics, Corvidae is the only attribution facing a myriad of challenges including: solution which uses sophisticated Machine Learning techniques to completely rebuilds • seriously flawed existing attribution models your marketing data to eliminate inaccuracy that are based on oversimplified and that prevents effective attribution and enables historical approaches that simply aren’t you to: suited to the length and complexity of the customer journey • view the complete customer journey across on and offline media • an increasingly long and complex customer journey that is undertaken across multiple • maximize marketing ROI by eliminating channels and devices both on and offline wasted spend and optimising budget allocation • poor underlying data that is largely predicated on a Cookie/Pixel based • correctly value content and build top of approach that generates data which is funnel activity around 80% incorrect • automate customer acquisition and convert Founded in 2008, QueryClick is an customers based on probability to convert independent digital marketing agency trusted • enable dynamic data science analysis and by brands over 32 markets and focused Machine Learning in action ATTRIBUTION PLAYBOOK 42

Goal: Conversion to Sale 01 Jan 2020 - 28 Mar 2020 Attributed Performance Vs Last Click Compare to: 01 Jan 2020 - 28 Mar 2020

#All +Segment Playbook Suggestions -> Select Region +

Total Change Direct Cannibalisation Talk to us about $4,901,417 £23,835,328 61% 13%

Corvidae is QueryClick’s SaaS marketing ROI Revenue Spend

Corvidae $4,000,000 attribution platform and the only attribution $3,900,000

$3,800,000 solution which completely rebuilds your $3,700,000 $3,600,000

$3,500,000

marketing data, reaching beyond Google 360 Revenue Total $3,400,000

$3,300,000

and Adobe Analytics. $3,200,000

$3,100,000

$3,000,000 Feb ‘20 Jan ‘20 Dec ‘19 Nov ‘19 Oct ‘19 Sep ‘19 Aug ‘19 Jul ‘19 Jun ‘19 May ‘19 Apr ‘19

Channel Revenue Cannibalised Revenue $12,356,228 / Last Click $8,356,228 GET STARTED £22,356,228 +32% 38%

“Corvidae reveals up to 334% more data for attribution than market-leading competitors. Providing game changing accuracy and predictive performance.”