Session: A Session With Lakshya Ankit, UndrAds - Understanding Ad Data Complexity.

[music] fall and fall and fall and fall and Up next, I would like to give a warm welcome to Lakshshire from Under Ads, who's going to be running a session now. So, please welcome Lra to the stage.
Hello, Helsinki. I guess it's pretty warm today. Well, I myself Lux and I'm just privileged to be here in front of you guys and presenting under ads and today I will be covering the topic understanding app data complexity and how AI and AI agents what is the role of agentic AI for speed and scale.

So perfect. So coming to Helsinki I was reading about a brief history about so so it has been a harbor for trading ground like almost like more than 200 years. So and of course as we know like we have seen some great games coming out from here like coming from like the fish boats to the clash of clans like this city definitely knows what the scale is all about. Today the gaming market has per my research what I've seen is like it's 2.85 billion worth of industry revenue which has been contributed from this particular country and the active studios well my numbers could be not absolutely correct but what I researched about was like almost 270 active studios who are building great games out here. [clears throat and cough] So what we

covered today, so why I'm here for. So I'm a fan of data and so you guys are. So when we talk about data, there are four layers uh when it comes to any kind of game. So the first one would be the user layer where you kind of check in like who is the player who is going and coming to your apps coming to your games and like what kind of device or OS what kind of country what kind of language we are talking about what is the installation installation source what is the platform uh then again we kind of break it down during the retention period like what is D1 what is D3 what is D7 and so forth and there are multiple benchmarks mult multiple metrics out there. The second layer how I break it down is like the app economy like the app events data where as in you understand like okay if a user is coming what is the number of sessions like what is the timing of the session whether it is morning whether it is [snorts] afternoon evening or at night

particularly like okay when is the session starting when is the session ending if a user completes a level if a user doesn't complete a level all kind of events what gets fired when a user is coming into your games, coming into your apps. That is something which I classify as the app economy data. The second thing is like after you kind of combine the user data, the app data of course the major chunk where we build all these games like the inapp purchases. So again like there is a lot of data which again suffices a lot of decisions like the price points the offers or the particular triggers when you kind of make those offers to the user. So again taking those decisions based on the data. So that again in itself becomes the AP economy and then coming to the last part which again like AP and IAA both of them contributes to your ARB DAO which is the holy metric for all of us.

So again in IAA it's like understanding from the advertiser viewpoint like okay what is the bid rate win rate show rate match rate or particularly like okay how much ad request CPM you are getting so all of these kind of data and combining all of this uh like the magnific like the entire magic sits on this data pool when you kind of look at this data holistically combining all of it. So it is really really important to understand like okay what's the role of data and before that to understand like okay what kind of data are we talking about. So how I would love to break it down is like there is one player doing one session and kind of completing seven kind of levels in your particular game. So how does it go? It's like like the data kind of starts scaling expanding. It would be just one user one session. But again if you'll take a look

at like okay now you understand okay there is one player there is from the app there is like one session but like after progressing from level one to level seven there would be multiple amounts of events which would be fired in a particular app in a particular game. So that particular data needs to be recorded and the amount and the scale of data you cannot imagine like it kind of it gets multiffold at every particular section when we talk about any kind of data like any level let's suppose if there was a level fail if you want to kind of trigger a particular offer like a rewarded maybe the AP over there or maybe you want the user to see certain kind of ads over there. So again combining all of this data it becomes a huge pool of data. So this is how I kind of see it. All right. So it's like in a particular one session like there are 65 events. All right. If you kind of just talk about like okay 1 million players

like daily active users on your game doing three sessions. So it easily goes to like close to 200 million. And like if you talk about the events in a month from one game, it reaches like close to 6 billion. Well, this is some data point which is really really hard to kind of manually ingest. Uh no matter like what kind of GPUs you are running, believe me like all the all of the workbooks definitely going to take a hang, right? And you would definitely earn a coffee break for yourself. Again like I would specifically come to the advertiser data on this particular bit like when you are talking about any kind of like 200 million kind of uh sessions in a day roughly even if you're doing like one placement it would be like 200 ad request 200 million ad request in a particular day for a particular ad request just to let you guys know there are 13 parameters base parameters which decides okay what is your revenue going to be. So we are talking about numerous

and a very large scale of data. Again the problem is not the scale of data. Well we all understand like okay if a game is there if a user is there if a session is being created if the levels are being kind of progressed there would be huge amount of data. But when we talk about this the main part is like the combination like when we make the combination if we are talking about 100 countries where your game is live we're talking about two platforms well this is platform is the only thing where you are limited whether iOS or Android well now we have got bunch of other options as well but majorly it's iOS and Android but if you're talking about like six placements or maybe we are talking about eight networks live on those placements be it from the US side be from the monetization side if we are talking about five device tiers maybe like iPhone 18 which got launched right now or maybe iPhone duo or maybe a Samsung Galaxy or maybe like a OnePlus. So we are talking about different device tiers. We are talking about different

app versions like maybe some will have like uh two ad placements, some will have no ad placement, some will have only no AP just the ads. So we are talking about like different versions of app and again like coming back to it like the user cohorts like okay maybe uh someone where like uh it is a demography coming from US where the user is like 13 to 21 and then we are talking about some other user demography possibly like uh women between the age of 25 to 35. So these kind of user coercing all right who who's going to pay who is not going to pay who is watching the ads completely who is not watching the ads completely. So the problem becomes when you combine all of this data and you try to kind of sync in some kind of inference from the data. So in a particular let's suppose you can just check the chart like if you kind of just keep on multiplying everything like if you just keep on combining all the all

the columns all the rows in your workbook. So each segment needs its particular pricing. All right? Be it IEP, be it IIA. So you cannot just simply go ahead and you cannot say like okay everyone in US is going to buy at $5 or 50. You cannot go and say like okay everyone is in US is going to see my 90cond rewarded ad. So you need to kind of make sure like everything is break down and all the decisions all the triggers what you're making is based on the combinations. But how do you kind of take those decisions? So it's like you understand the data. You kind of uh make sure like everything what sits with you is kind of getting churned in a way where you are getting more and more decisions out of it. Not just looking blatantly looking and doing some convenience-based or emotions or feeling based kind of decisions like hey it worked in the past maybe like it will work today. No that's not how we go. Like a lot of things which were working in the past today we don't use it. All

right. So the very simple thing is like okay how do we kind of get there? So like I said like the abundance of data you have versus the data you actually put to use. All right it's like a reference uh like between the sun and the earth. So of course if you're tapping all the data a lot of guys whom I meet it's like yeah we record all the data we have it on game analytics we have on bunch of different sources. We have it on bigquery. But my question, next question to them is like okay how are you looking at this data? How are you churning this data? Is it like any kind of manual algorithms you're running or is it like uh you're you're kind of uh just selecting some segments of data or maybe like because the data is huge you just go by like okay I'll just go and check my top apps maybe like I have 20 apps but I'll just go and check like top five. So that is how like currently I have seen majorly what people optimize and that's absolutely true as well right

you cannot put yourself and especially when you are the person who is kind of digging into data and then drawing out decisions you cannot just simply go ahead and you cannot do everything all at once in a particular single day or a single hour. So now there is a need like okay how do we put things together? So we at under ads particularly we are a fan of uh the great machine learning the causal learnings uh we put data to a great use. We do not let anything pass away by just like okay uh some event happened something triggered uh maybe a ad was shown maybe advertiser did a bid but we didn't tap it all right but the user came and user left. We don't do that. So whenever you are doing whenever you're sitting I'll leave you with some kind of uh algorithms over here like you can use a very simple predict and produce kind of a thing whereas in like anything any event you can put it to use

use like machine learning over there you can use causial learnings right now it is a new thing uh possibly all of you guys would be already aware about it so it's not just looking at the data patterns but it's also looking at the patterns of like okay what will change and how it will forecast before doing actually any kind of change. So you can use this at your end. uh possibly it's uh like there are particularly good books which I read particularly which can help you analyze the data in a way where honestly like you would have not imagined like the systems telling you like okay maybe we need to do this change and even before you do the change it will predict like okay this is a particular thing will like this is how the data pattern would change if you are doing any kind of change like maybe like the eCPM forecasts you kind of just bring in all the data all the session request according to the CPMs and then possibly if you want to know like okay by country what is the particular CPM and maybe like just combining the hour

of the day like maybe like in the 9:00 a.m. morning the CPMs would be different for your users in a particular let's suppose you're talking about 3 p.m. then there would be a different kind of use case. And if you're talking about like maybe like the first 3 days of the month, maybe your AP revenue could be higher because your users are getting paid like they are getting their salaries. All right, but how do you know this? You just need to make sure like you're looking at all the data points and you need to create systems which can analyze the data all by itself and just throw out results and inferences to you so that you can create you can validate your hypothesis. you can create certain kind of AB tests and again like keep a loop running over there according to data again I'm a big fan of Beijian theorems uh like uh it always helps you whenever like you feel like okay you are getting lost in data go back to the combinatorics and the go back to the probability systems which has been running like since a very long time I'm

a big fan of a algorithm which is multi-arm bandit algorithm it helps you explore it helps you exploit like whatever the next thing you're exploring. So possibly this is something which I find as a very great use. Let's suppose you're running like four price points for a particular geography for a particular day for a particular hour on a particular placement. Possibly you can check in like okay maybe a $3, maybe a $5, maybe a $7 or maybe a $10 floor and you can allocate very simple traffic to them. The arms of this algorithm would be checking in like okay what is the revenue which you're driving from these kind of CPMs and the one let's suppose like the $5 one is giving you the high payout then particularly it would be it will be pushing all the traffic to that particular $5 floor instead of putting it to three 7 or 10 and but still somehow like it will keep on like a very minuscule of traffic running on those particular arms to check like in real time like okay if the user is paying you

or particularly the advertiser is paying you. you you can check like okay at a particular given point of the day what exactly is paying you highest and the algorithm in itself which will make a switch that's what we do actually at our system again like uh there was a question somebody asked me yesterday over here it was a quick fire round it was uh someone else so they asked me like okay dashboards or decisions so how I would say now is like you can create any number of dashboards what you want to because of all the chat layers what we have got like the codecs the cloud code like you get multiple options like where you can create multiple dashboard but instead of creating dashboard what I would say is like just go and make a system for yourself where decision- making becomes easy for you not just looking at the data like dashboard is nothing just a display of data in some fancy manner all right it's better that you kind of use machine learning do

predictions Use Beijian theorems. If you are anywhere in doubt, particularly use causal learnings to prove your hypothesis. Even before you kind of do any kind of AB testing, you can prove your hypothesis before that. And the multi-arm band always helps you to kind of explore and exploit. But move out of dashboard, get into data and get these systems in place where they can help you take more decisions day by day. All right? Just don't look sitting at the data on a fancy dashboard. Maybe it could be a blue and white if it's under ads and it could be like anything else like coming in from any other company. How I would say is like the IP view the cohorts and the pricing points. So the very basic question which I get over here is like okay what is the first offer we make to our user like maybe like on D1 maybe on D3 what is what is that particular price point what we offer. So it's like you can get it

basically from a lot of data point maybe like if the user is working how many times like the user skip your AP button maybe that kind of data point if you're recording it maybe it can give you like a aspect like okay what could be the price point maybe if you know like uh if you're segmenting your user in on a basis like okay the demography the you like uh the hour of the day particularly or the uh day of the month whether it is like let's suppose the end week of the month maybe like you might not push much of the AP but if it's the starting of the month maybe you can because your users might be getting paid so these kind of inferences like the cohortings and the price points particularly like what what is a user valued at so this everything comes from the data like I keep on repeating this multiple times but it's never like do not take any hunch base or any past experience based or any convenience-based decisions data

will tell you if you're if you'll dig deep if you'll take a look holistically it will tell you what has to be done what needs to be done when we this is the very basic thing which which I was explaining like okay what is the path to a first purchase like install sessions levels first purchase repeat right like how do you decide like whether it is going to be a 50 cent pricing how do you decide whether is going to be a $1 pricing again go back to your data Like take a look over there, you'll find your answers. This is my favorite part which is like the ad auction. Well, under ads is a ad monetization company. We really understand the psyche of the advertisers coming from Google, Meta or Amazon or Pubmatic, Rubicon, anyone you name it like we work with them and we work with the direct buyers a lot of OEM companies. So one thing which we understand like okay in one ad auction what happens? So we get one request we amplify that request to every particular network and every particular network comes back with a lot of data like in a

particular bid request it's like not just the ad request it's like what's the bid rate we are getting from a particular buyer what is the win rate of that particular buyer what is the show rate of that the match rate the show rate and then eventually combining it to the fill rate what is the eCPM we are getting from that particular buyer so combining all of these data we kind of get and understand the viewability also gets gets counted in. So every one ad request generated creates like 13 base metrics related to it. So if we are talking about 1 million ad request a day, it automatically becomes like 14 million data points in a particular day. So we need to take decisions over there like low bid rate buyers don't want this inventory. Low win rate bids possibly are below the floor. low show rates. It's like you're showing your user a 90-cond ad where the user is not engaging and it will just kill the app and like reopen it. So these are the kind of inferences very short inferences which you can get. So buyers price

players and you your agents if you configure it properly they can score your buyers. Well I'm short on time so I'll just cover my this particular case study. So there was this Apple campaign, Apple Duo, which ran from like September 18 to September 27. And what we saw was like there was a certain user segment where we are we were getting like on a $20 floor, we were getting a $119 CPM coming from Apple. So what we did was like what our agents did, we just simply took everything from there. Uh we took like okay, what is the exact metric? What is the exact cohorting? We passed it to the studio and the studio did a bit minor change in the UA and the revenue jumped from a $35,000 to $11,000 in a particular day. This is a very real case which happened like last month and this is what the power of data is. And again like at under ads it's like we do

not just kind of see the pieces of the puzzles we put them together and how I would like to say is like uh we never miss any particular data point. We configure, we deliver, we measure, we do it hourly. So we are not looking at the data for previous day and we are not optimizing for the next day. We are optimizing for the next hour and looking at the data which is coming from the next 4 hours. So that is what we do and currently like we are at a scale where we kind of process more than 29 billion data points in a day. So yes, it's going good and particularly I would be open to any kind of questions any kind of thing like uh if you have. Thank you so much for the time.
Well, I would love to say like kito that's one thing which I learned.
Um yeah, we probably have time for one quick question if anyone has one.
I think you covered everything. Um but uh if you do have any questions for Laksha um please do connect with him

throughout the rest of the event. Uh I'm I'm you I'm sure you're going to be here all day, aren't you? So
yeah. Yeah, I'm just here around. All right. And possibly like I'm easy to find out given my height is like significant.
Yeah. Yeah, it's definitely taller than me. So, please give a round of applause.
All right. Thank you guys. and [music] I'm holding down and