by Alex Ho
(Image source: StorySet )
Table of Contents
- Advertising Attribution
- Common rules-based attribution methods
- The primary practice: Last-click Attribution
- Emerging Difficulties from the market trends
- Privacy policies formulated by tech firms
- Attribution is getting harder and harder
- Data-driven attributions methods: MTA & MMM
- MTA (Multi-Touch Attribution)
- MMM (Media Mix Modeling)
- Comparison between MTA and MMM
- Approaches related to ads attribution from tech Firms
Advertising Attribution aims to determine the contribution of each individual ad campaign touch point for a certain marketing goal (e.g. make a purchase, click on a banner, install mobile App).
There 3 reasons that it is difficult to correctly measure the marketing performances across multiple touch points
<1> Each Ad platform tends to credit same individual conversions to themselves
- Advertisers tend to simultaneously place multiple Ads in multiple channels.
- Each Ad platform (e.g. Facebook, Google) provides different attribution methods for the advertisers.
- (Take e-commerce scenario as example) If we add the converted GMV from various ads in each channels, the subtotal number could be higher than 100% of the real total GMV.
<2> Much traffic could not be correctly attributed because of the measurement limitation across devices or operating systems
- It makes the attribution model falsely credit a chunk of conversions to “direct traffic”
- Then it could make advertiser underestimate the performance from certain marketing channels.
<3> Data signal loss due to emerging privacy policies
- More and more privacy policies formulated by tech firms has made ads platforms more difficult to collect 3rd-party user data (compared with what they were able to before).
Common rules-based attribution methods
- Linear (equally credits every touch point)
- Time-decay (credits touch points by the duration between an impression and conversion)
- Position-based (40% credit each to the first and last impressions, and 20% spread over the rest)
The primary practice: Last-click Attribution
Last-click Attribution credits the conversion to only the last individual touch point that a user clicked (or saw) before taking the desired action (a.k.a driving the conversion). It is by far the most common and dominant attribution method.
There is one major drawback of Last-click Attribution: It could make advertisers (i.e. media buyers) ignore the value and incremental contribution of the touch points in the higher funnel (i.e. the marketing channels before the last touch point.)
Emerging Difficulties from the market trends
Privacy policies formulated by tech firms
Examples raised by the tech firms
- Apple ATT (App Tracking Transparency) policy
- Google will abandon 3rd-party cookie in Chrome browser soon (even if they keep delaying it)
- Google Tag Manager stops tracking tags from being fired for users who opted out the feature
- GA4 does not collect IP address information any more.
- Generally, less than 30% of iOS users are willing to let App developers track and use their device data.
- CPM cost in Meta Ads increased while impressions dropped.
- As the advertising performance drops, Advertiser (a.k.a Media Buyer) cut their overall online ads spending, and focus more on the Performance Marketing instead of Brand Marketing.
- Investors’ lack of confidence in platforms such as Meta and Snap were reflected in the stock prices.
Attribution is getting harder and harder
- Because of the privacy policies applied from the tech firms, there are more users (or visitors) are not able to be tracked by the advertisers and ad platforms.
- And as the majority of the advertisers are still relying on the
Last-click Attribution, it attributes even more conversions to the direct traffic (because we are not able to attribute it to the real campaign which should be credited).
- That makes the marketing performance of those major ad platform looks worse than before.
Data-driven attributions methods: MTA & MMM
Before we dive deeper, keep these in mind:
- No attribution model can perfectly credit each conversion to the correct touch point
- Either MTA or MMM is designed to complement the primary attribution model - the Last-click Attribution.
- Any projection based on historical data could be unhelpful when the external world changes dramatically (e.g. COVID-19 pandemic in 2020, Apple announced ATT policy in 2021, new players entering advertising market).
Some known facts from the markets
- As of July 2022, 53% of large advertisers are using MTA to track and optimize their marketing spending across channels (source: MMA Global)
(The following focus on the general concepts of MTA and MMM and will not identify any specific products from existing tech firms)
MTA (Multi-Touch Attribution)
- MTA assigns fractional credit to each touch point in the overall customer journey.
- Input - It tracks all of the individual-level data that happens in advertising campaigns.
- Output - So that advertisers can measure the contribution of each touch factors (e.g. channels, creatives, audiences).
- Drawbacks of MTA
- MTA requires access to individual-level user data tied to trackable conversion events (this could be helped by using MMM. Will be explained soon)
- Hard to handle offline campaign measurement
MMM (Media Mix Modeling)
- MMM uses supervised machine learning model to attribute the marketing performance results to multiple factors (for easier understanding, I personally separate them into marketing factors and external factors)
- Marketing factors examples: spending in different marketing channels, big promotions, special marketing-related events in the local market
- External factors examples: seasonality, market trends, and etc.
- Machine learning input and output (let’s take e-commerce advertiser’s scenario for example)
- Data input: spending in each single channel
- Data output: converted GMV from each single channel
- MMM does not need to know “what each factor actually means in the real world” in order to process the analysis and projection
- This means - if you randomly rename the “factors” that you input into the MMM model, it does not affect the output results.
- What’s more, after digesting the historical data, MMM model can project the future marketing performance results based on optimized re-allocated marketing spending suggestions
- This can help advertisers to review the contribution from each single marketing channel regardless of being online or offline
- Accordingly, advertisers can revise their spending allocation across various channels in order to maximize the marketing performance.
- If you are already using MTA, then MMM can be seen as a complement.
- You can consider the former as micro and short-term analysis, while the latter as macro and long-term analysis
- Common drawbacks of MMM (more to be stated in the following table)
- MMM requires huge amount of data (maybe accumulated for at least 2 or 3 years) as input of the model to generate reliable results and projection. Such time-length is much longer than campaign time length in the real world, so it could not catch up with emerging external signals fast enough as new things happen. So it’s not realistic to use MMM to review the performance of short-term advertising campaigns.
Comparison between MTA and MMM
(And don’t forget. We can use both MTA and MMM simultaneously.)
Time-lagging measurement and projection
Suitable use cases
Measuring performances of short-term campaigns
Long-term and macro marketing spending optimization
Level of attribution granularity
Able to measure the impact of a specific single touch point
Not able to measure the impact of a specific single touch point
Data volume required for data input
Does not need huge amount of data
Needs huge amount of data as input of the model to generate reliable results and projection. Generally it requires data accumulated for more than 2 or 3 years.
Dependency on advertiser’s ability to track individual user identity
MTA requires access to individual-level data Therefore, MTA could be severely affected by the restrictions from tech firms (Apple and Google)
MMM can be applied even when advertisers are unable to track individual user behavior. Information of individual users are unnecessary. MMM uses aggregated date for modeling
Hard to measure offline channels Ads performances
Able to handle both online and offline channels
Approaches related to ads attribution from tech Firms
Meta provides MMM Robyn, a free and open-sourced data model, for advertisers or software companies to implement and include it into their own analysis tools.
Robyn is an automated Marketing Mix Modeling (MMM) code. It aims to reduce human bias by means of ridge regression and evolutionary algorithms, enables actionable decision making providing a budget allocator and diminishing returns curves and allows ground-truth calibration to account for causation.
In September 2021, Google replaced the Last-click attribution by Data-driven Attribution as the default conversion model in Google Ads (buy-side ad network).
- It uses in-house algorithm to assign fractional credit to touch points
- Last-click attribution is still available, not deprecated.
Google Data-driven Attribution Methodology
More Marketers Are Adopting MTA, But There's Still Some Frustration
Multi-touch attribution is reaching an inflection point (finally). The majority of large advertisers (53%) now say they use MTA to track and optimize their spending across channels, according to new research from marketing trade org MMA Global, released on Monday.
Goodbye, Last-Click Attribution: Google Ads Changes Default To Data Modeling | AdExchanger
Is this truly the end of last-click attribution? Google will no longer use last-click attribution as the default conversion model in Google Ads, its buy-side ad network, the company announced in a blog post on Monday. The change will mean that, going forward, the default attribution method for any conversion touchpoint - a new product...
Google Is Creating A Panel To Feed Its Conversion Models | AdExchanger
Everything old is new again. Surveys are on the rise, marketers are turning their attention to media mix modeling (MMM) - and Google is recruiting participants for an online panel through a partnership with Gallup. The panel will serve as a truth set to validate the models Google uses to estimate online conversions across its measurement products.
TPG 週刊 Issue 23 - 歸因模型百百種
TPG 週報會在台灣時間每週一早上 10:00 出刊，每一期將由 TPG 成員分享當週所閱讀的大小新聞與短評，還有不分新舊的優質閱讀文章分享。 上期提到的 Unity 二選一難題，馬上就迎來了結果。Unity 重申對於 ironSource 的青睞，並拒絕了 AppLovin 的提親。根據 Unity 官方說法，主要拒絕的理由是在他們分析完 AppLovin 的提案後，認為在財務面上及戰略面上都沒有更佳的優勢，因此拒絕此項提案。（ Richard ） 華爾街日報近日刊出一篇重磅報導，揭露了三年前 Apple 與 Facebook 曾經試圖達成若干協議，讓蘋果可以從廣告營收為主的 App 當中獲取利益。蘋果一開始提案由 Facebook 推出一個免費廣告版的訂閱功能，讓蘋果透過 App Store 取得 30% 的分潤，後續也曾提案是否 Facebook 能夠將 30% 的 iOS 廣告費分潤給蘋果。然而此二提案並沒有被 Facebook 接受，因此也被認為是後續蘋果推出 ATT 及 IDFA 緊縮的相關政策的原因。（ Richard ） 世人皆知串流影音服務終將超越有線電視收視，而這天已到來。根據尼爾森的數據顯示，今年七月美國串流影音收視 首次超過有線電視，34.8%
Study: Effectiveness of Apple's App Tracking Transparency
Does it stop third-party tracking? Or is it just an illusion of privacy? In April 2021, Apple released the App Tracking Transparency ("ATT") feature with iOS 14.5. ATT claims to give users choice and transparency for third-party tracking in their apps, and it was lauded by many as a step forward in protecting user privacy.
In 2019, Amex started a project to replace MTA with MMM
Amex Is Planning For The Cookieless Future With An Eye On MMM
American Express has been investing in measurement technology to try and give credit where credit is due. The increase in regulatory scrutiny, the end of third-party cookies and the paucity of other identifiers means that "what used to work doesn't work anymore," said Abhi Juneja, VP of performance marketing and ad tech data science at Amex.
- Software Product Manager in tech.
- 10 years of experience as software product manager in these areas: #ecommerce #SaaS #SoutheastAsia #delivery #MarTech #SuperApp
- Work experiences in Taipei, Singapore, and Shanghai.
- Currently based in Taipei City, Taiwan.
- My Linkedin profile: Alex Ho
Notes & Articles