UA Coffee: All About LTV (Life Time Value) and pROAS with Algolift

everybody welcome to the next edition of you a coffee talk as in the last few episodes we have our typical guest Nebo and john welcome back guys if we have a new guest with us Paul Bowen from album maybe just really tell us a little bit about yourself yeah so I signed in mobile in advertising 20 years ago ten years ago I started the European office of tap joy over in London late 2015 and moved to San Francisco were tapped where we did six months there and then moved to unity where I oversaw the account management team on the ads in the division that so overseeing the MV at that time who said three years and then nice and so the topic of our discussion today is gonna

be all about LTV I thought we could first start by talking maybe starting with you Paul so you know Alcala provides LTV PL TV sorts of analysis but could we talk about what are the primary ways and game companies use LTV today and just be clear do they use LTV or do they use PL TV yes I think that the PNP on TV means predicted so you're essentially trying to understand what the future looks like based on past past behavior I think you guys did a really good job of sort of breaking down how most most current game camisa understandings of massing metrics right now so I'd say almost all of that sort of clients and people that we talk to looking at deep 7ro us as the core metric the games which they optimize their martin campaigns we have some clients who use the one row as as a metric to understand sort of the future

of item value of their users but really what we're trying to do and I think where the market is going to go over the next couple years is looking at further window than seven days which is really a problem to sort of solve and yeah I mean you guys talked about so various ways simple ways to sort of predict future lifetime value but really understanding sort of what the lifetime value of the user will be in a year's time or two-year cycle three years have you got an actual game is pretty in technology or thinking about maximizing your Martin's been and as competition on the sort of marketing platforms gets more it gets tougher yeah you want to be able to basically spend as much as you can it's for you your clients do people want P LTV the predictive LTV because you might know up to how much they can spend to acquire users that's right yeah I mean you know most clients that we would sort of have a pretty strict payback we know that

they want to be hitting so although they look at 70 robots as a short-term metric to basically move money around yeah so they'll do a review of their marketing spend over say six months and you know do you want to you and me both John you guys in your experience what you guys do or what you're you know what you hear your colleagues are doing what what are the primary applications for you guys when it comes to LTV or PLC V primitives the same thing basically just understanding what the money's gonna pay back and how well with it with our previous campaign network we should spend the money in the future and it doesn't really quickly I think that's one more point it's really important just you cannot wait for at six months to see what is something payback or not ideally you know understand what's going on after like a week or maybe a few days

and be able to like make decisions based on what's interesting that if you if he go back to like the the beginnings of like mobile smartphones right a lot of times that you had the first developers on scene that had the luxury to wait I didn't have all the competitors they didn't have to worry about the cause of acquisition as much nowadays it's very different when you have a lot of competitors casa acquisition is always rising so waiting is much more cost relevant in terms of that decision-making so moving towards KLTV is definitely something that other companies you mean include convinced I've worked before I've moved towards right but that's a great point gun because to your point what you're saying is because there's so much competition if you have a good idea of how much you can spend and be confident that you can make it back then you know because the part used is an auction-based market you

want to be able to match that up as much as possible in order to get market share right that's right just to give you one quick example they're like I remember when I joined play Studios our VP of UA at the time options said that if I can go back in time three years ago I would spend this much more are you waiting right I wish I could throw a million dollars instead of anything 100,000 and that's always that kind of hindsight that marketers will always see just more risky decisions now like pretty much every developer change as well so it's just more risky do not make so for some of the developers out there so that you

a folks out there who are thinking about how to use PL TV so like this when we talk about Facebook or Google PLT be to some degree doesn't matter so much is that right or do you factor that in and then you just use that for more like the how're you guys you know once you get that peel TV number how are you guys using that yeah I mean LTV matters on every platform so what you have so Facebook and Google but it uses short term sort of feedback loops know I could value optimization optimization so they may be a quick feedback to the algorithms to find the users for mass in campaigns however what the Apple or the Kevlar will begin but yeah that is really the long-term value of these users so whereas Facebook and Google are concerned about short term short term metrics because that's what they're taught to optimize Wars really what what game coming care about is so LTV is is

it's sort of foundational to I guess so one of the challenges with women's at would say Facebook Google is they're not looking across all the channels so even if they did build their own healthy meanwhile which they could do they're not telling you what the LTV you cannot really input like that specific number let's say you know your relative ease $20 M for like four hundred eighty days and you're like you take $20 for that type of traffic that you can every control the price that way for the cost that way and as a result in just so much harder to to apply learnings from from like to appeal to be you know models to also know there's like with differences regime like I networks in Google Facebook is that like it's significantly harder to predict that LTV on the lt1 Facebook where you don't have a look at

for himself like how they change their optimization model Google specifically you don't have a lot of control in terms yeah as much whereas like with reward a video or had networks ago I observe that like their algorithms are typically rule-based and they're typically much more static in terms of like how they evolve over time so you actually have a lot more time to determine why the pl TV is for reward of video person it's like a Facebook Google where things can be changing so on that side on Facebook Google side from an out of the lift do you guys do stuff like pyramids or like you have any products that can help people on yes should I tell you what we do and then some yeah so tungaluuk so we so there's a couple of different ways to sort of predict sort of LTV and I think you guys last week one of them is sort of using

this output curve slash using a multiple on g7 robots turn your head future LTV another might be using attention an opt-out to sort of model out what the TV looks like they're both good but they're both terrible models and they're limited in terms of how flexible they are living in a campus or than anomalies where you have whales inside them and you cannot be powerful to break down those models you come into sort of to really understand what's going on in the cobalt so what we do is we approach it from the user level and we make user level projections on a daily basis for every gate flavor in the game so every every single user everyday gets an updated PL TV projection at a 30 60 90 180 360 i we do project that further than that there's a lot that's inherent risk in sort of making projections and making marketing decisions when you're buying on say a two-year window so we tend to

advise against that so we on a daily basis will ingest data from our clients the most recent of attribution transaction data say whether that IQ purchases subscriptions or we run the moles we share so that's the first thing that we do which is sort of user level LTV projections the benefit of user level and and there's some some companies in the market then this is a really interesting case nothing but you can sort of rolled up your users on any dimension so any geo platform any anything appended to sort of use ready really understand what your video of each code looks like so that has real flexibilities in UAE and so yes they so then the second thing

we do is we connect to the reporting and managing so then we'll put in the spend data historical spend later and then on a daily basis so that allows us at the user level to understand the periods per user we can then grow because we have to spend you know a year two years time in order to do that you would have needed some you know training meter or you would have needed like how long would it take or how much time does it take before you get a good whether it's a yeah so I mean I think you guys sort of bit about this last week I mean the more historical dates are the better for us in terms of you need to trade later we can use priors from similar apps than new types of games so we do so if a

client is a new game it's an actually we have a bunch of matches on our platform so we only had an understanding of what as the coop matures obviously can get better that's we're predicting that we tend to ingest at least a year's worth of data when we start working with the client that means that we can see temporal changes so you know we see especially on the on the marketing side you know to send that the seventh of January prices get generally cheaper to users because so we ingests at least years worth of data but we start making projections really on the new part that's about weaves up we build another one we even use prize to do that but the month something like that has about three months all right and so you guys

did sound like you guys are using machine learning based approach and maybe John and Eva for the bulk of getting companies I don't think a lot of them are using machine learning based ways of doing pretty to melki be at all right and so you know I happy are they doing a lot of that I mean I know from my experience it seems like there's there's a few different approaches that I've seen that maybe you guys comment on what you guys are saying but one is we talked about a simple retention knock down model so you check the arktech check their attention draw the retention curve yeah that's one I think the second way we talked about is you know you try and get a are pure and try to extrapolate that out that's 2/3 methodology is if you've had an existing game before then I don't know what it's I call it a factory base model where you take a bunch of different factors from a similar game and you regress against the LTV or whatever and then you just kind

of waited that way but that's basically all I've seen and the various levels of sophistication but generally speaking I would say the majority the companies I've worked with generally you just you know markdown retention and it's very hand wavy so I don't know buddy what did you guys I think everyone started with the art that attention yeah I don't know anyone who and this is a third especially when you're like a new gaming company like you do Paul Spong if you lack a certain level of sophistication like you you have to start somewhere I think it's directionally it's good to have however the limitations to that is like how it's it's too inflexible shrinks - Richie you're also making a lot of assumptions about like the opt-out of the user a channel or are you looking at do you have to a grenade that information and that's why I think it's a lot more directional the second one you talked about which Paul would

describe it as like a extrapolating DX robots and then traveling and LTP out of it that's actually where I've seen most companies that have been doing they would look at their if he's doing nomads and think of multiple ahead and say okay that is that would be being very okay also limitations to that and very directional but how do you determine the flexibility of the multiples how do you differentiate them between Geo's channels platforms or optimization levels and how often should you update them so these are big challenges when you're looking at a very rigid way of looking at LTV or determination of recoup again directional but can you make the day-to-day decision on it I wouldn't think so the last one you mentioned is actually what I saw my last six months at place to do is is really team was really moving towards a machine learning taking

ingesting user level behavior right and determining the LTV of those particular users or some examples of things that we considered is like after session time and session that number of spins frequency of payment amount of payment and I don't mean remember I think there were likes it was kind of say you know like 80 different things that they take into account and they trained it over the course of the year and they didn't look back to look at the accuracy of it and at the time I believe they were like 90% less in terms of the accuracy of what they predicted after years old and no a lot of company is trying to move in that direction but do they have the resource to build something that and even if you have the resource to build it shoot you and those are all fundamental questions I've seen a lot companies asking themselves you more yes it's like both

of these love completely nor the UA side which is what the problem is with names like one-size-fits-all solutions are just a simple you know simple you know using the function to figure out what you know what the other thing that's important to mention is that a lot of companies don't really have enough data to actually build machine learning models so that's probably a good question for Paul as well it's like you know which what install volumes you need and like what your indium needs to be in order for these models to work well with Network we have a we're in SDK play as the case solution so we have our own solution so we track you know user journey and we understand their behavior and then make predictions based based on that which means it a little easier and that's why we can do a model but the thing is that for a lot of smaller companies that's not a not really an option which is why they have to use models so in terms of sort of the cohort size so I mean really depends on the

product conversion rate but generally sort of the five minutes install they were two week cohort is sort of what we need to make projections the way that we use to we leverage the LTV projection to then inform user acquisition campaigns so I mentioned so we connect algorithm actually make changes to those to those campaigns so we're also sort of dark directionally is more important is way more important accuracy so accuracy is very key but also understanding directionally the PLT via cross so we are in new poll has a very good

point about a lot of LT beam models not truly taking into account the newest I yeah yeah Paul pretty much mention when he said like these things are directional it's like there is a completely another side of like you way which is basically changing when its own

and your job when you're using these models is to like optimize things in a way that that yields the highest possible return to the company and not like not it does not return a specific number because it's impossible to guarantee it simply because you know in December you a more expensive in January it's a cheaper I feel a lot of there's and other home escapes it spends like I know 20 million dollars in a month and just completely you know changes the market dynamic so you cannot really take these factors into account when building the model which is why your job when building model is to figure out you know where it's been money in the most efficient way given the current market circumstances and I think that's why these models have to be directional and not exact well directional first directional first yeah and also like that my question for you is what is what does the output for like Facebook alone especially for

Facebook do you you know send more like spend more money in VO campaigns or there's like you have a specific output in terms of what that you mean to rise should be or like a should be yeah so - I mention we do so Delta you mean modeling the second piece we do is is optimization of campaign so we work we have algorithms that look at last - C over c pn CBR and CPI and so we're determining basically if we increase the budget on this yo campaign what does appear oh I'd look like as we continue to increase the budget and then taking into account that asset as part of the labor across those three platforms what what do we do in terms of increasing the budget on that campaign - we work with clients as they'll say that I want to hit like say you want to pay

back max by 365 Pro s so we we didn't we relied on the Facebook and Google algorithms to find those to find excuses through there and then we layer on top of that to determine which asset type with the exits being Adsense campaign Thank You woods which asset type should we buy more or less up there's a bet based on the so yeah we don't we we tend to ask our clients or have notes of mineralize on their campaigns and just basically create a campaign and then this is sort of trying to change the app and you know change the runners so we basically say we want more or less of this asset based on its

projected rise so that's how we approach provide customer perspective are they asking more for this service relative to the PL DEP realize that for sake is it's the PM TV is that you know the P Ltd thing is challenging because people have a solution that that works until there's a cost benefit of how much do I really want to invest in this yeah so the application of activity we need to it's much more interesting to clients because automating user acquisition is it's something that exact some people do you ain't no you just want to save time yeah most people haven't built you a conservation so yeah they may have some level of automation but most people don't have a programmatic solution to it so most clients on the other side because they've managed to build a detention model in it you feel like what works enough other things that you see a

lot of different I mean that sort of well buzzword in the industries is so creative that's so any insights so I'd like to ask mr. insights I'm creative that is a real that's a challenge because it's very hard to sort of you we we make you level predictions and so if you have creative reporting at the user we there is the possibilities of what the creative there's really hard to sort of make the determination that you can be said in that creative across multiple different access or campaigns so just rolling that that creative up on the user level I'm saying this PLT but this creative like you can't really you can't really do that so we were working on insights to clients in terms of the

creative level that's of the number one is a we've sort of Apple the other asset we have is yeah platform so we're currently integrating with fungal right now and we can start working with dinosaurs soon so basically I think I think that that's the panacea user acquisition is the automation of all platforms you know I think you oughta Mason tends to it reduces the amount of people that need need to do the work it will never come you still mean so the way that we currently work with clients right now and we'd like to automate some items of this but we still need campaigns creative we don't have a

automated way to create campaigns that's something we would like to do that's quite hard problems so because there's a lot of unknowns around creating campaign and there's the strategy for the strategic element to that that's off line and we don't have to historical things on that and then we need creative refresh which you're getting humans are better eridium descending it's also one of the things that they're one of the problems so we've gotta have the same approach we're just like it's not worth it you have power everything changes yeah what what what might be interesting for us to build is giving insights on what they've mined to create this is actually a creating them so you know looking at everything in up with it on that platform I'm basically making a recommendation of weeping this if you create this sort of campaign it will start gaining traction this is this greater this nation a lot of agents

bought time saving right so when you're thinking about the savior for campaigns with like 30 line each telling something yes psi you either have a person who will go in and make a creative change one with every single line item by a campaign or you find a way to do it via an API building that out is a incredible time-saver and and what the you a person will essentially do is instead of doing this execution ad off-site work will shift towards more analysis right understanding like okay how can I spend my dollar mortgage I think you guys just become more strategic a little bit so it changes the role of what you do so I would say like with what five years ago or even three years ago you have these massive massive QA teams because you needed that amount of people and hands and to actually pour levers what we're trying to do is to eliminate the number of like manual lever pulling and

automate that so that you can have less teams less people in the team but much more efficient so last question what's the future in terms of whether it's how we think about Piro as RPO TV how do you weigh with these types of backups the next service and you guys are gonna offer yeah well I do I think start with sort of market terms of cost of user acquisition rising over time means that the companies are going to need to get better at predicting when they let's marketing spend they're gonna get it get it back so I think I think there's gonna be a lot more investment you know really that I think inside companies as well so communicative scientists to be able to make these predictions and make them more accurate and directly correct so that would be the first thing on the LTV so I think there's gonna be investment

on that so driven driven by and then I think overall automation is going to be there's a lot that's around automation so how do we sort of cross-platform to automate user acquisition what can we do and creative what can we do one sort of creating campaigns I think we just got just a general I mean you're the industry continues to consolidate here the fewer players they're much bigger I think as you mentioned automation is gonna be really important you're probably going to be some monetization of field duty like more people will be able to use luchon of all solutions such as on the length arts you know any other solution you know and and it's gonna be really important to understand where can this money across the portfolio than just games and even that's where a makes you plays a really important role and understanding where and how to spend money yeah it's pretty that's pretty

much it you know just like there can be fewer channels that you will need to understand better how stand money in scale and agree with you both I think most of the competition in ramallah combination is really about increasing sophistication or what it boils down to is efficiency right time management how can you do the most work with at least amount of time and that's the hockey of the amalgamation and it's funny it's like the word automation been around for three years and we have truly I guess stopped until now seen any companies make big strides in that front I'm sure Network and a few other companies have done quite a bit but most companies have yet to really fill that in so I think like a solution is like our live would be very very helpful for life as a midsize companies to really get dip their toes into having a solution that

could allow them to make much more efficient decisions right actually one thing that you mentioned really struck me okay so my personal opinion would be bad the one bit of analysis that I haven't seen enough oh and you know I started doing some of this work before which was about what you were talking about elasticity I think so you know yeah yeah yeah

[Laughter]

[Music] [Music]