Modeled Conversions on Google Marketing Platform
September 28th 2020 marks a pivotal day for advertisers/brands using the Google Marketing Platform either through its adserver Campaign Manager (CM) or its DSP Display & Video 360 (DV360). Slightly under the radar of all the data legislation & privacy changes in the ecosystem, Google will be moving towards Modeled Conversions for attribution. Considering both CM & DV360 have the largest market share in both the adserver & DSP markets, this is a fundamental change. So let’s break it down & talk about why this is only the start of how attribution is evolving in the face of change within adtech.
What are Modeled Conversions on CM/DV360?
Modeled Conversions are as they sound, a predictive based approach to probabilistically make assumptions around attributed success. Google will utilise machine learning to focus on all conversions that are unattributed by nature and attempt to attribute them towards specific campaigns where they feel they deserve credit. There is not so much detail shared publicly on how it works within Google but the announcement for DV360 was in August release notes found here.
What is the expected impact of this change?
Google have communicated that they expect “single digit percentage increase” in conversions to be seen from those that include modeled conversions in their attribution models. They will also be rolling it into any existing conversion reporting so don’t be on the lookout for any new conversion metrics that specifically report on modeled conversions.
Can I opt out of this change?
Yes you can but you have until September 28th 2020 to do so. You are opted in by default so anyone that does wish to make a change will have to configure this. This is done dependent on the platform:
For Campaign Manager, navigate to the Floodlight Configuration of the Advertiser you wish to change and scroll down to the Attribution section. Within here are the Attribution Models, including the default Floodlight Standard Model. There is a new setting as shown in the image above, which can be edited to opt out.
For DV360, navigate to the Floodlight Group tab under Resources in a DV360 Advertiser. Under Basic Details, you can edit the Attribution Model to remove modeled conversions. Note if you are using CM as the adserver, you don’t need to do this as DV360 inherits from CM so this is only for DV360 advertisers not using CM.
For advertisers/brands that may be sceptical around these changes, there is one recommendation that could be interesting to look at, which is to create 2 separate attribution models within the platforms, where one is opted in to modeled conversions & the other is not. Then it may be possible to see the actual impact to this change. Note at the time of writing this, this is unproven but something that I will be looking at come the first week of October.
Why are Google doing this?
Essentially this change in methodology comes at a time when the measurement / attribution of digital advertising especially on web, is becoming much more challenging due to the combination of data legislation such as GDPR & data privacy changes from the different browsers. The latter is very much more aligned to attribution, with concepts such as Apple’s Intelligence Tracking Prevention (ITP) on Safari having severe implication on all forms of attribution. And with the latest announcement from Apple in terms of iOS14 meaning ITP is now on by default across all browsers (not just Safari), this will even more restrict what is possible to measure / attribute on web. For more information on that, check out Simo Ahava’s blog post on it here.
This should be very much seen as a change for web based conversions on both Desktop & Mobile. Mobile App is not the focus here & with iOS14 opt in delayed until 2021, expect more shifts in the app world to be seen at this point.
Is this the first time Google have done this?
In short no. In fact a lot of Google’s methodology over time has slowly moved away from a deterministic approach revolving around cookies towards a probabilistic modeled approach as seen now. To name a few:
Back in 2017 when Google first started to make major changes to their pixel on all of their solutions (Global Site Tag), Google Ads noted that it would start to use Google Analytics data to model conversions. Here there was also use of anonymised Google sign in data. This is called out in this Google Ads support link here.
In both mid 2019 & early 2020, focusing on Mobile App campaigns on iOS, Google announced some dramatic changes to how attribution would work on Google Ads App campaigns (formerly UAC), which would become more probabilistic. A good Adweek article talking about it is here.
So what type of conversions exist on CM / DV360?
There are actually quite a few types of conversions within CM & then extended if you use DV360 as your DSP also. Modeled conversions will fit into this but not as a brand new field:
Standard Conversions are based on the legacy approach of tying a cookie on an ad impression or click to a Floodlight event e.g. transaction. You could say this is deterministic by nature but with all the cookie limitations, this is getting harder to do. Modeled conversions will enter this specific metric in reporting. This is available on both CM & DV360.
Cross Environment Conversions are Google’s main report for looking at potential overlap between both devices & browsers. This is a separate report that can be pulled but on a more aggregated level & often reports conversions back in decimals. Modeled conversions will not be part of this at launch but expected to over time. This is available on both CM & DV360.
Same Device Conversions are a DV360 specific methodology for attribution, where Google will attempt to validate attribution if an user was to use the same device but be exposed / convert within 2 different environments i.e. web & app. This has been the default setting in DV360 for a while and cannot be split out in reporting. There has been no indication on how modeled conversions will fit in with this as of yet.
What are other platforms / vendors doing?
The idea of modeled / probabilistic measurement is not new & many vendors have specialised in it for a number of years. However what is potentially quite scary about Google’s shift to do this, is that they hold more data than anyone else, which should theoretically empower their machine learning to be more accurate vs a smaller vendor who is trying to do the same thing with less data or stick to a deterministic approach
Facebook are very much 2nd in terms of data that they have on individuals vs Google but their methodology for conversion attribution at least on its main product Ads Manager remains deterministic by nature. This is generally helped by the fact that the large majority of ads are served in an environment that Facebook knows everything about however they still have the same challenges when cookie based pixels are involved to do attribution. However expect this to change over time, with Facebook already announcing earlier this year that other products/solutions such as lift studies & Facebook Attribution have already taken a probabilistic approach as seen here
In the adserver market, Campaign Manager’s main competitor has very much become Flashtalking over the past 18 months whilst Amazon work on rebuilding Sizmek back up. And Flashtalking has its core attribution product fTrack with a probabilistic approach at the heart of it. This is an answer to solving attribution without reliance on cookies which makes use of other signals instead. But considerations to have on this include the fact that it is not free & could be argued that there is an element of fingerprinting going on, which is being cracked down further by the browsers.
Closing Thoughts
Web based attribution & conversion measurement is approaching a new normal with the decline of cookies accelerating based on browser changes. Whilst 3rd party cookies are the main focus of these changes, which has much more significant implications on impression based attribution, do not underestimate the fact that 1st party cookies are being targeted/limited also. As digital marketing evolves away from cookies but also away from channels that there are no clicks to fallback on e.g. audio & out of home, a probabilistic approach is the only logical approach without the common currency of an universal ID, which looks ever-increasingly unlikely to occur despite some of the work being done in the ecosystem due to privacy concerns.
What is more of a talking point can be seen in 2 areas. The fact that Google and soon to be Facebook are doing this, as the two largest vendors with the most amount of data are having to turn to probabilistic methodology. So how can anyone else try to do it in a more deterministic manner if those 2 can’t? The other area is very much the theme that is prevalent in the ecosystem, which is that you will be ultimately forced to take the numbers you see and live with it, with no real way to validate or pick them apart. Thus giving more power to the Walled Gardens.