Adam Smart (Director of Product at AppsFlyer) and Piyush Mishra (Lead Growth Marketing at Product Madness) take a deep dive on SKAN with David Philippson (Founder & CEO at DataSeat) and Tim Koschella (Founder & CEO at Kayzen). They look at the different approaches that people are taking to measuring user value (Conversion Values), how Apple's privacy threshold is iterating and impacting postbacks, and how to drive up ATT opt-ins.
On Android, Google allows a parameter to be pulled from the app store called RefTag/ReferID. It’s a simple way of attributing deterministically without a device ID, that doesn’t allow you to create profiles of users (or to collect long-term data on users).
What’s puzzling is that Apple has invented a completely new, extremely complex and yet still immature system to manage attribution that is off all the standards from the industry.
There are different privacy thresholds:
- The privacy thresholds that can prevent from knowing which publisher served the ad in the postback
- The privacy thresholds associated to the conversion values that the advertisers have set
For any app network, knowing which publisher drove the install is the biggest variable. But the way the privacy thresholds have been designed (probably to affect Google/Facebook), they are far too high.
DataSeat believes that the privacy thresholds are 10 installs or above, per publisher, per region (Apple’s region: Europe, North America, South America, APAC), per campaign ID, with a 24h rollout period. Above that, the installs from that publisher will start being populated.
The way the thresholds work creates an incentive to use less campaign IDs. It also skews things towards bigger publishers/networks like Facebook, Google, Snapchat and TikTok which is ironic because SKAN was probably designed to hurt these.
Ad networks were not transparent with app bundles because they were worried that advertisers could buy traffic directly from the source apps. For some, also because they balance high-quality traffic with low quality traffic.
SRNs certainly don’t want to report what is click-through and what is view-through. The biggest example is Youtube: it drives huge performance but it also skews towards views whereas the rest of the industry is judged on clicks. This is something advertisers should now be able to see if they receive the postbacks.
Product teams might need to build games/apps with ATT in mind in order to get the right signals and be able to acquire users.
There are 2 main variables to think about when you design the ML algorithm of a DSP:
- Time delay of the event, which is always a tradeoff between receiving an event early and its quality
- Frequency at which the event happens, and even without taking SKAN into account some frequencies are so low that you can’t really optimize for that
In the end it comes down to the scale a lot.
No ad network in the industry is really there yet in terms of optimization, so you need to simplify. For now, optimize for 1. Earlier events 2. More frequently occurring events.
Before SKAN, when Facebook was optimizing for AEO or VO, a lot of it is not based on data from the actual campaign but from pre-existing data sets that Facebook connects back to your campaign goals (e.g. knowing that a user has paid in another game). This is going away with SKAN so you won’t be able to feed that to the ML anymore.
The impact depends on what kind of monetization model you have:
- Ad-monetized (e.g. hypercasual game): delta between high-value and low-value users is relatively low so you “just” need to recalibrate to the fact that you’re getting lower CPMs (but you’ll probably have lower CPIs too).
- IAP-monetized (e,g, casino game at the extreme): you rely on high-paying users so the impact on both monetization and UA is much bigger.
You’ll have to rely on a mix of data to maximize the outcome of the new paradigm:
1. Maximize ATT opt-in so you can have deterministic data to work with (e.g. 20/30/40%) and extrapolate.
2. Work closely with your MMP to make sure that SKAD data is gathered for all the networks, including the SRNs (it’s worth much less if Facebook and Google are not included)
3. Have sophisticated pLTV buckets instead of just binary events
If you want to understand the actual impact of ATT on your business, you need to measure opt-in by the overall number of users you actually get opt-in for (not just who has seen the prompt) because you can’t track all the other ones (including the ones opted-out by default). If you want to understand users’ reactions to privacy measures then you should measure how many users confronted with the prompt have made the conscious choice to opt-out or opt-in.
If you look at ATT opt-in data based on ad requests, it’s yet another data set: you can’t track at the user level so you measure at the aggregate level how many users have opted-in. But maybe users that opt-in are under/over represented in terms of ad frequency.
UA managers now have much more complexity to deal with. They have to make qualitative judgement because a lot of the data they have is uncertain: what data set do you trust more, what is the underlying source, etc. It’s become more similar to what an investment analyst does.
With view through and some level of multi-touch attribution in SKAN, there isn’t much evolving SKAN can do besides allowing real-time postbacks and lowering privacy thresholds...And that’s not going to happen because Apple is not trying to compete on attribution: they’re happy with a limited attribution that helps their advertising products.
On Android, Google allows a parameter to be pulled from the app store called RefTag/ReferID. It’s a simple way of attributing deterministically without a device ID, that doesn’t allow you to create profiles of users (or to collect long-term data on users).
What’s puzzling is that Apple has invented a completely new, extremely complex and yet still immature system to manage attribution that is off all the standards from the industry.
There are different privacy thresholds:
- The privacy thresholds that can prevent from knowing which publisher served the ad in the postback
- The privacy thresholds associated to the conversion values that the advertisers have set
For any app network, knowing which publisher drove the install is the biggest variable. But the way the privacy thresholds have been designed (probably to affect Google/Facebook), they are far too high.
DataSeat believes that the privacy thresholds are 10 installs or above, per publisher, per region (Apple’s region: Europe, North America, South America, APAC), per campaign ID, with a 24h rollout period. Above that, the installs from that publisher will start being populated.
The way the thresholds work creates an incentive to use less campaign IDs. It also skews things towards bigger publishers/networks like Facebook, Google, Snapchat and TikTok which is ironic because SKAN was probably designed to hurt these.
Ad networks were not transparent with app bundles because they were worried that advertisers could buy traffic directly from the source apps. For some, also because they balance high-quality traffic with low quality traffic.
SRNs certainly don’t want to report what is click-through and what is view-through. The biggest example is Youtube: it drives huge performance but it also skews towards views whereas the rest of the industry is judged on clicks. This is something advertisers should now be able to see if they receive the postbacks.
Product teams might need to build games/apps with ATT in mind in order to get the right signals and be able to acquire users.
There are 2 main variables to think about when you design the ML algorithm of a DSP:
- Time delay of the event, which is always a tradeoff between receiving an event early and its quality
- Frequency at which the event happens, and even without taking SKAN into account some frequencies are so low that you can’t really optimize for that
In the end it comes down to the scale a lot.
No ad network in the industry is really there yet in terms of optimization, so you need to simplify. For now, optimize for 1. Earlier events 2. More frequently occurring events.
Before SKAN, when Facebook was optimizing for AEO or VO, a lot of it is not based on data from the actual campaign but from pre-existing data sets that Facebook connects back to your campaign goals (e.g. knowing that a user has paid in another game). This is going away with SKAN so you won’t be able to feed that to the ML anymore.
The impact depends on what kind of monetization model you have:
- Ad-monetized (e.g. hypercasual game): delta between high-value and low-value users is relatively low so you “just” need to recalibrate to the fact that you’re getting lower CPMs (but you’ll probably have lower CPIs too).
- IAP-monetized (e,g, casino game at the extreme): you rely on high-paying users so the impact on both monetization and UA is much bigger.
You’ll have to rely on a mix of data to maximize the outcome of the new paradigm:
1. Maximize ATT opt-in so you can have deterministic data to work with (e.g. 20/30/40%) and extrapolate.
2. Work closely with your MMP to make sure that SKAD data is gathered for all the networks, including the SRNs (it’s worth much less if Facebook and Google are not included)
3. Have sophisticated pLTV buckets instead of just binary events
If you want to understand the actual impact of ATT on your business, you need to measure opt-in by the overall number of users you actually get opt-in for (not just who has seen the prompt) because you can’t track all the other ones (including the ones opted-out by default). If you want to understand users’ reactions to privacy measures then you should measure how many users confronted with the prompt have made the conscious choice to opt-out or opt-in.
If you look at ATT opt-in data based on ad requests, it’s yet another data set: you can’t track at the user level so you measure at the aggregate level how many users have opted-in. But maybe users that opt-in are under/over represented in terms of ad frequency.
UA managers now have much more complexity to deal with. They have to make qualitative judgement because a lot of the data they have is uncertain: what data set do you trust more, what is the underlying source, etc. It’s become more similar to what an investment analyst does.
With view through and some level of multi-touch attribution in SKAN, there isn’t much evolving SKAN can do besides allowing real-time postbacks and lowering privacy thresholds...And that’s not going to happen because Apple is not trying to compete on attribution: they’re happy with a limited attribution that helps their advertising products.
On Android, Google allows a parameter to be pulled from the app store called RefTag/ReferID. It’s a simple way of attributing deterministically without a device ID, that doesn’t allow you to create profiles of users (or to collect long-term data on users).
What’s puzzling is that Apple has invented a completely new, extremely complex and yet still immature system to manage attribution that is off all the standards from the industry.
There are different privacy thresholds:
- The privacy thresholds that can prevent from knowing which publisher served the ad in the postback
- The privacy thresholds associated to the conversion values that the advertisers have set
For any app network, knowing which publisher drove the install is the biggest variable. But the way the privacy thresholds have been designed (probably to affect Google/Facebook), they are far too high.
DataSeat believes that the privacy thresholds are 10 installs or above, per publisher, per region (Apple’s region: Europe, North America, South America, APAC), per campaign ID, with a 24h rollout period. Above that, the installs from that publisher will start being populated.
The way the thresholds work creates an incentive to use less campaign IDs. It also skews things towards bigger publishers/networks like Facebook, Google, Snapchat and TikTok which is ironic because SKAN was probably designed to hurt these.
Ad networks were not transparent with app bundles because they were worried that advertisers could buy traffic directly from the source apps. For some, also because they balance high-quality traffic with low quality traffic.
SRNs certainly don’t want to report what is click-through and what is view-through. The biggest example is Youtube: it drives huge performance but it also skews towards views whereas the rest of the industry is judged on clicks. This is something advertisers should now be able to see if they receive the postbacks.
Product teams might need to build games/apps with ATT in mind in order to get the right signals and be able to acquire users.
There are 2 main variables to think about when you design the ML algorithm of a DSP:
- Time delay of the event, which is always a tradeoff between receiving an event early and its quality
- Frequency at which the event happens, and even without taking SKAN into account some frequencies are so low that you can’t really optimize for that
In the end it comes down to the scale a lot.
No ad network in the industry is really there yet in terms of optimization, so you need to simplify. For now, optimize for 1. Earlier events 2. More frequently occurring events.
Before SKAN, when Facebook was optimizing for AEO or VO, a lot of it is not based on data from the actual campaign but from pre-existing data sets that Facebook connects back to your campaign goals (e.g. knowing that a user has paid in another game). This is going away with SKAN so you won’t be able to feed that to the ML anymore.
The impact depends on what kind of monetization model you have:
- Ad-monetized (e.g. hypercasual game): delta between high-value and low-value users is relatively low so you “just” need to recalibrate to the fact that you’re getting lower CPMs (but you’ll probably have lower CPIs too).
- IAP-monetized (e,g, casino game at the extreme): you rely on high-paying users so the impact on both monetization and UA is much bigger.
You’ll have to rely on a mix of data to maximize the outcome of the new paradigm:
1. Maximize ATT opt-in so you can have deterministic data to work with (e.g. 20/30/40%) and extrapolate.
2. Work closely with your MMP to make sure that SKAD data is gathered for all the networks, including the SRNs (it’s worth much less if Facebook and Google are not included)
3. Have sophisticated pLTV buckets instead of just binary events
If you want to understand the actual impact of ATT on your business, you need to measure opt-in by the overall number of users you actually get opt-in for (not just who has seen the prompt) because you can’t track all the other ones (including the ones opted-out by default). If you want to understand users’ reactions to privacy measures then you should measure how many users confronted with the prompt have made the conscious choice to opt-out or opt-in.
If you look at ATT opt-in data based on ad requests, it’s yet another data set: you can’t track at the user level so you measure at the aggregate level how many users have opted-in. But maybe users that opt-in are under/over represented in terms of ad frequency.
UA managers now have much more complexity to deal with. They have to make qualitative judgement because a lot of the data they have is uncertain: what data set do you trust more, what is the underlying source, etc. It’s become more similar to what an investment analyst does.
With view through and some level of multi-touch attribution in SKAN, there isn’t much evolving SKAN can do besides allowing real-time postbacks and lowering privacy thresholds...And that’s not going to happen because Apple is not trying to compete on attribution: they’re happy with a limited attribution that helps their advertising products.
Notes for this resource are currently being transferred and will be available soon.
David
Started in 2018 or even prior, but no one paid attention. It was a placeholder by Apple.
People paid attention when Apple made the IDFA announcement.
Before, they were using UDID/IMEI then Apple introduced the IDFA which was privacy compliant because you could reset it.
First bigger move towards privacy was the introduction of the LAT toggle. It was limited to tech savvy people mostly, which actually made LAT on users more valuable in some instances.
Tim
Apple is a big company, with lots of teams and different opinions. The way things have evolved shows that there wasn’t a masterplan from the start. Only in the last couple of years there might be one, before that it was more “following the industry”. Now they’ve started to think ahead in terms of privacy.
[💎@07:30] On Android, Google allows a parameter to be pulled from the app store called RefTag/ReferID. It’s a simple way of attributing deterministically without a device ID, that doesn’t allow you to create profiles of users (or to collect long-term data on users).
[💎@08:10] What’s puzzling is that Apple has invented a completely new, extremely complex and yet still immature system to manage attribution that is off all the standards from the industry.
David
Apple initially created IDFA for attribution purposes. In itself it does not have huge privacy implications, but what does is what the SRN (Self Reported Networks) started with their own attribution: asked MMPs to send all installs before deciding if an install came from their network.
Then, other ad networks and DSPs had to compete and started gathering data through suppression lists. It served a purpose for the advertiser but it was also a disguise to have more data and compete with Google and Facebook.
That’s where Apple’s issue is with IDFA and Apple basically introduced a complex system to counteract Google and Facebook.
Piyush
Google and Facebook are still trying to differentiate themselves from DSPs and Ad Networks by not communicating all the data.
Tim
It took a lot of time to understand the full impact, due to the complexity of SKAN. Example: SKAN postback doesn’t have a 1:1 matching with an impression or a click and that’s what the industry is built on.
Adam
It’s indeed complex, but the elephant in the room is the privacy thresholds.
David
Apple had ITP for Safari on the web, so it was only a question of time before something came to the app industry.
David
[💎@17:45] There are different privacy thresholds:
For any app network, knowing which publisher drove the install is the biggest variable. But the way the privacy thresholds have been designed (probably to affect Google/Facebook), they are far too high.
[💎@18:45] DataSeat believes that the privacy thresholds are 10 installs or above, per publisher, per region (Apple’s region: Europe, North America, South America, APAC), per campaign ID, with a 24h rollout period. Above that, the installs from that publisher will start being populated.
[💎@19:20] The way the thresholds work creates an incentive to use less campaign IDs. It also skews things towards bigger publishers/networks like Facebook, Google, Snapchat and TikTok which is ironic because SKAN was probably designed to hurt these.
There seems to have been an increase in the conversion value privacy thresholds at the beginning of June, 2021.
Piyush
On TikTok/Snapchat, where there is only one source app id Product Madness seems to be receiving conversion values for more than 90% of the installs. On the DSPs it’s closer to 30%.
In terms of receiving the source app id, from Snapchat/TikTok they’re receiving about 50-60%. For DSPs it’s close to 1-2%.
Tim
Some networks are probably not comfortable sending the source app ids to MMPs.
Many of the ad networks have not shared the app bundle before, under the reasoning that their job is to bring installs at a certain CPM buy (which doesn’t require transparency). The same way, Facebook never broke down the source apps from FAN.
Tim believes they won’t share that now, either. Why would they provide more transparency now?. Note: this podcast episode was before Apple’s announcements that advertisers could get the postbacks directly.
Many of the advertisers would pass hashed identifiers to advertisers for the source app bundle so that they could still manage their campaigns (blacklist some, etc.), but advertisers would not know the exact app.
David
[💎@25:30] Ad networks were not transparent with app bundles because they were worried that advertisers could buy traffic directly from the source apps. For some, also because they balance high-quality traffic with low quality traffic.
[💎@26:15] SRNs certainly don’t want to report what is click-through and what is view-through. The biggest example is Youtube: it drives huge performance but it also skews towards views whereas the rest of the industry is judged on clicks. This is something advertisers should now be able to see if they receive the postbacks.
95% of the industry has been building user graphs, and that’s now damaged. But the SKAN change is leveling up the playing field even though it’s negative for most.
Once iOS 14.5 rolls out to 90%+, there will be significant detrimental effects on behavioral-based bidding ad networks and SRNs. This means that there will be unspent budget, and this brings opportunities for players that design with privacy in mind.
Tim
At the time of the episode, too early to tell the impact. But the status quo has been challenged and so the interest in Kayzen has increased.
The density of competition for iOS inventory has gone down, due to uncertainty. Budgets have been reduced (or switched to Android) until they have more clarity on the data.
David
Now we’re asking to predict LTV with buyer events within the first 24 hours, which is really hard. Companies are looking at cohorts over 365 days to identify if there were common patterns in the first 24 hours.
Slots might have an advantage because you can have players deposit money in the first 24 hours. Even better for companies like Booking or Expedia.
Adam
[💎@35:17] Product teams might need to build games/apps with ATT in mind in order to get the right signals and be able to acquire users.
Piyush
You need to get the first conversion value within the first 72 hours. From a DSP perspective, should you focus on just one event (e.g. revenue) or on several events?
Tim
[💎@36:40] There are 2 main variables to think about when you design the ML algorithm of a DSP:
In the end it comes down to the scale a lot.
[💎@38:35] No ad network in the industry is really there yet in terms of optimization, so you need to simplify. For now, optimize for 1. Earlier events 2. More frequently occurring events.
[💎@38:57] Before SKAN, when Facebook was optimizing for AEO or VO, a lot of it is not based on data from the actual campaign but from pre-existing data sets that Facebook connects back to your campaign goals (e.g. knowing that a user has paid in another game). This is going away with SKAN so you won’t be able to feed that to the ML anymore.
David
If networks only rely on SKAN, then they have to optimize just based on these first 24 hours.
If there is merit to measure 64 conversion values, there is no penalty to it. But if you change your conversion values, then it creates a delay and that will affect the campaign optimization.
There will be more sophisticated solutions coming up, based on pLTV.
Machine learning can help define which conversion value pLTV “bucket” a player should be placed based on his behavior (e.g. how many spins, tutorial, deposit, etc.). This is what AppsFlyer offers with PredictSK.
Adam
Is there a whole new set of KPIs that need to be defined?
Tim
[💎@45:55] The impact depends on what kind of monetization model you have:
David
[💎@48:40] You’ll have to rely on a mix of data to maximize the outcome of the new paradigm:
Adam
Do you calculate the opt-in based on how many users have seen the prompt or how many have come in vs. how many have accepted?
Tim
[💎@50:30] If you want to understand the actual impact of ATT on your business, you need to measure opt-in by the overall number of users you actually get opt-in for (not just who has seen the prompt) because you can’t track all the other ones (including the ones opted-out by default). If you want to understand users’ reactions to privacy measures then you should measure how many users confronted with the prompt have made the conscious choice to opt-out or opt-in.
[💎@52:05] If you look at ATT opt-in data based on ad requests, it’s yet another data set: you can’t track at the user level so you measure at the aggregate level how many users have opted-in. But maybe users that opt-in are under/over represented in terms of ad frequency.
Piyush
Since you receive IDFA data for some apps and not for others (users might make a different choice based on the app), is there any potential to extrapolate data and use it for retargeting?
David
It is technically possible to match users (IDFA and IDFV from different apps you own, or mobile browser and mobile device, etc.) through probabilistic matching but you shouldn’t try that.
David
The UA manager role is going to change significantly and become far more of a data role: you’ll still have to look at D7 ROAS and optimize based on it, but you’ll also have to understand pLTV, incrementality (uplift over baseline), etc.
It’s hard for in-house UA to compete against ad networks. But when you can then it means you’re bringing more control in-house.
Tim
The amount of skills you need to be a good UA manager now really matters again. Before there wasn’t some huge value in having a very skilled and experienced UA manager.
[💎@1:03:04] UA managers now have much more complexity to deal with. They have to make qualitative judgement because a lot of the data they have is uncertain: what data set do you trust more, what is the underlying source, etc. It’s become more similar to what an investment analyst does.
Piyush
At a campaign level, you have much less data: only media, source, level. A limited number of campaigns. So there will be a movement towards creatives: new formats, etc.
David
[💎@1:06:00] With view through and some level of multi-touch attribution in SKAN, there isn’t much evolving SKAN can do besides allowing real-time postbacks and lowering privacy thresholds...And that’s not going to happen because Apple is not trying to compete on attribution: they’re happy with a limited attribution that helps their advertising products.
Tim
Apple is not looking to create an attribution tool they own and can sell. It’s going to evolve because there are a lot of challenges, including fraud. Apple might be addressing this but it will require more data and time.
David
Started in 2018 or even prior, but no one paid attention. It was a placeholder by Apple.
People paid attention when Apple made the IDFA announcement.
Before, they were using UDID/IMEI then Apple introduced the IDFA which was privacy compliant because you could reset it.
First bigger move towards privacy was the introduction of the LAT toggle. It was limited to tech savvy people mostly, which actually made LAT on users more valuable in some instances.
Tim
Apple is a big company, with lots of teams and different opinions. The way things have evolved shows that there wasn’t a masterplan from the start. Only in the last couple of years there might be one, before that it was more “following the industry”. Now they’ve started to think ahead in terms of privacy.
[💎@07:30] On Android, Google allows a parameter to be pulled from the app store called RefTag/ReferID. It’s a simple way of attributing deterministically without a device ID, that doesn’t allow you to create profiles of users (or to collect long-term data on users).
[💎@08:10] What’s puzzling is that Apple has invented a completely new, extremely complex and yet still immature system to manage attribution that is off all the standards from the industry.
David
Apple initially created IDFA for attribution purposes. In itself it does not have huge privacy implications, but what does is what the SRN (Self Reported Networks) started with their own attribution: asked MMPs to send all installs before deciding if an install came from their network.
Then, other ad networks and DSPs had to compete and started gathering data through suppression lists. It served a purpose for the advertiser but it was also a disguise to have more data and compete with Google and Facebook.
That’s where Apple’s issue is with IDFA and Apple basically introduced a complex system to counteract Google and Facebook.
Piyush
Google and Facebook are still trying to differentiate themselves from DSPs and Ad Networks by not communicating all the data.
Tim
It took a lot of time to understand the full impact, due to the complexity of SKAN. Example: SKAN postback doesn’t have a 1:1 matching with an impression or a click and that’s what the industry is built on.
Adam
It’s indeed complex, but the elephant in the room is the privacy thresholds.
David
Apple had ITP for Safari on the web, so it was only a question of time before something came to the app industry.
David
[💎@17:45] There are different privacy thresholds:
For any app network, knowing which publisher drove the install is the biggest variable. But the way the privacy thresholds have been designed (probably to affect Google/Facebook), they are far too high.
[💎@18:45] DataSeat believes that the privacy thresholds are 10 installs or above, per publisher, per region (Apple’s region: Europe, North America, South America, APAC), per campaign ID, with a 24h rollout period. Above that, the installs from that publisher will start being populated.
[💎@19:20] The way the thresholds work creates an incentive to use less campaign IDs. It also skews things towards bigger publishers/networks like Facebook, Google, Snapchat and TikTok which is ironic because SKAN was probably designed to hurt these.
There seems to have been an increase in the conversion value privacy thresholds at the beginning of June, 2021.
Piyush
On TikTok/Snapchat, where there is only one source app id Product Madness seems to be receiving conversion values for more than 90% of the installs. On the DSPs it’s closer to 30%.
In terms of receiving the source app id, from Snapchat/TikTok they’re receiving about 50-60%. For DSPs it’s close to 1-2%.
Tim
Some networks are probably not comfortable sending the source app ids to MMPs.
Many of the ad networks have not shared the app bundle before, under the reasoning that their job is to bring installs at a certain CPM buy (which doesn’t require transparency). The same way, Facebook never broke down the source apps from FAN.
Tim believes they won’t share that now, either. Why would they provide more transparency now?. Note: this podcast episode was before Apple’s announcements that advertisers could get the postbacks directly.
Many of the advertisers would pass hashed identifiers to advertisers for the source app bundle so that they could still manage their campaigns (blacklist some, etc.), but advertisers would not know the exact app.
David
[💎@25:30] Ad networks were not transparent with app bundles because they were worried that advertisers could buy traffic directly from the source apps. For some, also because they balance high-quality traffic with low quality traffic.
[💎@26:15] SRNs certainly don’t want to report what is click-through and what is view-through. The biggest example is Youtube: it drives huge performance but it also skews towards views whereas the rest of the industry is judged on clicks. This is something advertisers should now be able to see if they receive the postbacks.
95% of the industry has been building user graphs, and that’s now damaged. But the SKAN change is leveling up the playing field even though it’s negative for most.
Once iOS 14.5 rolls out to 90%+, there will be significant detrimental effects on behavioral-based bidding ad networks and SRNs. This means that there will be unspent budget, and this brings opportunities for players that design with privacy in mind.
Tim
At the time of the episode, too early to tell the impact. But the status quo has been challenged and so the interest in Kayzen has increased.
The density of competition for iOS inventory has gone down, due to uncertainty. Budgets have been reduced (or switched to Android) until they have more clarity on the data.
David
Now we’re asking to predict LTV with buyer events within the first 24 hours, which is really hard. Companies are looking at cohorts over 365 days to identify if there were common patterns in the first 24 hours.
Slots might have an advantage because you can have players deposit money in the first 24 hours. Even better for companies like Booking or Expedia.
Adam
[💎@35:17] Product teams might need to build games/apps with ATT in mind in order to get the right signals and be able to acquire users.
Piyush
You need to get the first conversion value within the first 72 hours. From a DSP perspective, should you focus on just one event (e.g. revenue) or on several events?
Tim
[💎@36:40] There are 2 main variables to think about when you design the ML algorithm of a DSP:
In the end it comes down to the scale a lot.
[💎@38:35] No ad network in the industry is really there yet in terms of optimization, so you need to simplify. For now, optimize for 1. Earlier events 2. More frequently occurring events.
[💎@38:57] Before SKAN, when Facebook was optimizing for AEO or VO, a lot of it is not based on data from the actual campaign but from pre-existing data sets that Facebook connects back to your campaign goals (e.g. knowing that a user has paid in another game). This is going away with SKAN so you won’t be able to feed that to the ML anymore.
David
If networks only rely on SKAN, then they have to optimize just based on these first 24 hours.
If there is merit to measure 64 conversion values, there is no penalty to it. But if you change your conversion values, then it creates a delay and that will affect the campaign optimization.
There will be more sophisticated solutions coming up, based on pLTV.
Machine learning can help define which conversion value pLTV “bucket” a player should be placed based on his behavior (e.g. how many spins, tutorial, deposit, etc.). This is what AppsFlyer offers with PredictSK.
Adam
Is there a whole new set of KPIs that need to be defined?
Tim
[💎@45:55] The impact depends on what kind of monetization model you have:
David
[💎@48:40] You’ll have to rely on a mix of data to maximize the outcome of the new paradigm:
Adam
Do you calculate the opt-in based on how many users have seen the prompt or how many have come in vs. how many have accepted?
Tim
[💎@50:30] If you want to understand the actual impact of ATT on your business, you need to measure opt-in by the overall number of users you actually get opt-in for (not just who has seen the prompt) because you can’t track all the other ones (including the ones opted-out by default). If you want to understand users’ reactions to privacy measures then you should measure how many users confronted with the prompt have made the conscious choice to opt-out or opt-in.
[💎@52:05] If you look at ATT opt-in data based on ad requests, it’s yet another data set: you can’t track at the user level so you measure at the aggregate level how many users have opted-in. But maybe users that opt-in are under/over represented in terms of ad frequency.
Piyush
Since you receive IDFA data for some apps and not for others (users might make a different choice based on the app), is there any potential to extrapolate data and use it for retargeting?
David
It is technically possible to match users (IDFA and IDFV from different apps you own, or mobile browser and mobile device, etc.) through probabilistic matching but you shouldn’t try that.
David
The UA manager role is going to change significantly and become far more of a data role: you’ll still have to look at D7 ROAS and optimize based on it, but you’ll also have to understand pLTV, incrementality (uplift over baseline), etc.
It’s hard for in-house UA to compete against ad networks. But when you can then it means you’re bringing more control in-house.
Tim
The amount of skills you need to be a good UA manager now really matters again. Before there wasn’t some huge value in having a very skilled and experienced UA manager.
[💎@1:03:04] UA managers now have much more complexity to deal with. They have to make qualitative judgement because a lot of the data they have is uncertain: what data set do you trust more, what is the underlying source, etc. It’s become more similar to what an investment analyst does.
Piyush
At a campaign level, you have much less data: only media, source, level. A limited number of campaigns. So there will be a movement towards creatives: new formats, etc.
David
[💎@1:06:00] With view through and some level of multi-touch attribution in SKAN, there isn’t much evolving SKAN can do besides allowing real-time postbacks and lowering privacy thresholds...And that’s not going to happen because Apple is not trying to compete on attribution: they’re happy with a limited attribution that helps their advertising products.
Tim
Apple is not looking to create an attribution tool they own and can sell. It’s going to evolve because there are a lot of challenges, including fraud. Apple might be addressing this but it will require more data and time.
David
Started in 2018 or even prior, but no one paid attention. It was a placeholder by Apple.
People paid attention when Apple made the IDFA announcement.
Before, they were using UDID/IMEI then Apple introduced the IDFA which was privacy compliant because you could reset it.
First bigger move towards privacy was the introduction of the LAT toggle. It was limited to tech savvy people mostly, which actually made LAT on users more valuable in some instances.
Tim
Apple is a big company, with lots of teams and different opinions. The way things have evolved shows that there wasn’t a masterplan from the start. Only in the last couple of years there might be one, before that it was more “following the industry”. Now they’ve started to think ahead in terms of privacy.
[💎@07:30] On Android, Google allows a parameter to be pulled from the app store called RefTag/ReferID. It’s a simple way of attributing deterministically without a device ID, that doesn’t allow you to create profiles of users (or to collect long-term data on users).
[💎@08:10] What’s puzzling is that Apple has invented a completely new, extremely complex and yet still immature system to manage attribution that is off all the standards from the industry.
David
Apple initially created IDFA for attribution purposes. In itself it does not have huge privacy implications, but what does is what the SRN (Self Reported Networks) started with their own attribution: asked MMPs to send all installs before deciding if an install came from their network.
Then, other ad networks and DSPs had to compete and started gathering data through suppression lists. It served a purpose for the advertiser but it was also a disguise to have more data and compete with Google and Facebook.
That’s where Apple’s issue is with IDFA and Apple basically introduced a complex system to counteract Google and Facebook.
Piyush
Google and Facebook are still trying to differentiate themselves from DSPs and Ad Networks by not communicating all the data.
Tim
It took a lot of time to understand the full impact, due to the complexity of SKAN. Example: SKAN postback doesn’t have a 1:1 matching with an impression or a click and that’s what the industry is built on.
Adam
It’s indeed complex, but the elephant in the room is the privacy thresholds.
David
Apple had ITP for Safari on the web, so it was only a question of time before something came to the app industry.
David
[💎@17:45] There are different privacy thresholds:
For any app network, knowing which publisher drove the install is the biggest variable. But the way the privacy thresholds have been designed (probably to affect Google/Facebook), they are far too high.
[💎@18:45] DataSeat believes that the privacy thresholds are 10 installs or above, per publisher, per region (Apple’s region: Europe, North America, South America, APAC), per campaign ID, with a 24h rollout period. Above that, the installs from that publisher will start being populated.
[💎@19:20] The way the thresholds work creates an incentive to use less campaign IDs. It also skews things towards bigger publishers/networks like Facebook, Google, Snapchat and TikTok which is ironic because SKAN was probably designed to hurt these.
There seems to have been an increase in the conversion value privacy thresholds at the beginning of June, 2021.
Piyush
On TikTok/Snapchat, where there is only one source app id Product Madness seems to be receiving conversion values for more than 90% of the installs. On the DSPs it’s closer to 30%.
In terms of receiving the source app id, from Snapchat/TikTok they’re receiving about 50-60%. For DSPs it’s close to 1-2%.
Tim
Some networks are probably not comfortable sending the source app ids to MMPs.
Many of the ad networks have not shared the app bundle before, under the reasoning that their job is to bring installs at a certain CPM buy (which doesn’t require transparency). The same way, Facebook never broke down the source apps from FAN.
Tim believes they won’t share that now, either. Why would they provide more transparency now?. Note: this podcast episode was before Apple’s announcements that advertisers could get the postbacks directly.
Many of the advertisers would pass hashed identifiers to advertisers for the source app bundle so that they could still manage their campaigns (blacklist some, etc.), but advertisers would not know the exact app.
David
[💎@25:30] Ad networks were not transparent with app bundles because they were worried that advertisers could buy traffic directly from the source apps. For some, also because they balance high-quality traffic with low quality traffic.
[💎@26:15] SRNs certainly don’t want to report what is click-through and what is view-through. The biggest example is Youtube: it drives huge performance but it also skews towards views whereas the rest of the industry is judged on clicks. This is something advertisers should now be able to see if they receive the postbacks.
95% of the industry has been building user graphs, and that’s now damaged. But the SKAN change is leveling up the playing field even though it’s negative for most.
Once iOS 14.5 rolls out to 90%+, there will be significant detrimental effects on behavioral-based bidding ad networks and SRNs. This means that there will be unspent budget, and this brings opportunities for players that design with privacy in mind.
Tim
At the time of the episode, too early to tell the impact. But the status quo has been challenged and so the interest in Kayzen has increased.
The density of competition for iOS inventory has gone down, due to uncertainty. Budgets have been reduced (or switched to Android) until they have more clarity on the data.
David
Now we’re asking to predict LTV with buyer events within the first 24 hours, which is really hard. Companies are looking at cohorts over 365 days to identify if there were common patterns in the first 24 hours.
Slots might have an advantage because you can have players deposit money in the first 24 hours. Even better for companies like Booking or Expedia.
Adam
[💎@35:17] Product teams might need to build games/apps with ATT in mind in order to get the right signals and be able to acquire users.
Piyush
You need to get the first conversion value within the first 72 hours. From a DSP perspective, should you focus on just one event (e.g. revenue) or on several events?
Tim
[💎@36:40] There are 2 main variables to think about when you design the ML algorithm of a DSP:
In the end it comes down to the scale a lot.
[💎@38:35] No ad network in the industry is really there yet in terms of optimization, so you need to simplify. For now, optimize for 1. Earlier events 2. More frequently occurring events.
[💎@38:57] Before SKAN, when Facebook was optimizing for AEO or VO, a lot of it is not based on data from the actual campaign but from pre-existing data sets that Facebook connects back to your campaign goals (e.g. knowing that a user has paid in another game). This is going away with SKAN so you won’t be able to feed that to the ML anymore.
David
If networks only rely on SKAN, then they have to optimize just based on these first 24 hours.
If there is merit to measure 64 conversion values, there is no penalty to it. But if you change your conversion values, then it creates a delay and that will affect the campaign optimization.
There will be more sophisticated solutions coming up, based on pLTV.
Machine learning can help define which conversion value pLTV “bucket” a player should be placed based on his behavior (e.g. how many spins, tutorial, deposit, etc.). This is what AppsFlyer offers with PredictSK.
Adam
Is there a whole new set of KPIs that need to be defined?
Tim
[💎@45:55] The impact depends on what kind of monetization model you have:
David
[💎@48:40] You’ll have to rely on a mix of data to maximize the outcome of the new paradigm:
Adam
Do you calculate the opt-in based on how many users have seen the prompt or how many have come in vs. how many have accepted?
Tim
[💎@50:30] If you want to understand the actual impact of ATT on your business, you need to measure opt-in by the overall number of users you actually get opt-in for (not just who has seen the prompt) because you can’t track all the other ones (including the ones opted-out by default). If you want to understand users’ reactions to privacy measures then you should measure how many users confronted with the prompt have made the conscious choice to opt-out or opt-in.
[💎@52:05] If you look at ATT opt-in data based on ad requests, it’s yet another data set: you can’t track at the user level so you measure at the aggregate level how many users have opted-in. But maybe users that opt-in are under/over represented in terms of ad frequency.
Piyush
Since you receive IDFA data for some apps and not for others (users might make a different choice based on the app), is there any potential to extrapolate data and use it for retargeting?
David
It is technically possible to match users (IDFA and IDFV from different apps you own, or mobile browser and mobile device, etc.) through probabilistic matching but you shouldn’t try that.
David
The UA manager role is going to change significantly and become far more of a data role: you’ll still have to look at D7 ROAS and optimize based on it, but you’ll also have to understand pLTV, incrementality (uplift over baseline), etc.
It’s hard for in-house UA to compete against ad networks. But when you can then it means you’re bringing more control in-house.
Tim
The amount of skills you need to be a good UA manager now really matters again. Before there wasn’t some huge value in having a very skilled and experienced UA manager.
[💎@1:03:04] UA managers now have much more complexity to deal with. They have to make qualitative judgement because a lot of the data they have is uncertain: what data set do you trust more, what is the underlying source, etc. It’s become more similar to what an investment analyst does.
Piyush
At a campaign level, you have much less data: only media, source, level. A limited number of campaigns. So there will be a movement towards creatives: new formats, etc.
David
[💎@1:06:00] With view through and some level of multi-touch attribution in SKAN, there isn’t much evolving SKAN can do besides allowing real-time postbacks and lowering privacy thresholds...And that’s not going to happen because Apple is not trying to compete on attribution: they’re happy with a limited attribution that helps their advertising products.
Tim
Apple is not looking to create an attribution tool they own and can sell. It’s going to evolve because there are a lot of challenges, including fraud. Apple might be addressing this but it will require more data and time.