UMT is one of the many abbreviations floating around adjoe. When I first started writing here, I tried to Google the term along with ROAS, CPI, and the rest of the kind. But for UMT, there was nothing. That was the first hint that UMT was something more special; something custom. Now I know it is, and it’s something worth sharing, especially for developers looking for users who will truly appreciate their app and pay back into it like no one else. 

So I sat down with Berkay, Director of Growth, to talk about it.

a: Hi Berkay! If I stop you in a hallway: what’s UMT?

Berkay: User Model Targeting. It’s our ability to target only the best-fit users for your game, making ad campaign performance better.

a: How much better? 

Berkay: You’ll see changes in almost all KPIs: ROAS (return on ad spend), ROAS D7, and other longer-term ROAS KPIs. UMT campaigns perform two to four times better than regular targeting.

a: So what’s the idea behind it? Why did targeting need to change at all?

Berkay: We thought: why do we treat all targeted users the same, when not all of them perform in the same way? 

Don’t get me wrong, in regular targeting, performance still meets partner expectations, but it’s just one good strategy for everyone. Let’s say you have user classes from A to F: to target all of them with the same reward curve or CPI isn’t the most effective way to distribute your resources.

For some users, advertisers just need to be more competitive, and it’s absolutely worth it. These users perform exceptionally well in the short term, and even better in the longer term, so advertisers can specifically claim that portion of users for increased profitability.

a: If UMT performs so much better, why doesn’t everyone just run User Model Targeting?

Berkay: Running on UMT alone wouldn’t be possible because of lower volume, since the model only targets cherry-on-top users. And regular campaigns still meet partner KPIs. 

UMT is not a replacement for regular targeting, it’s an add-on that lets you run a tailored acquisition strategy for your ‘VIP’ users, so you target and convert them as effectively as possible.

And since the value of these users is higher, the competition for them is as high. It’s more difficult to acquire them. 

So it’s a combination of higher competition and lower volume that makes it impossible to scale on UMT alone. But that’s also what makes it so effective in terms of performance and profitability.

a: You’ve called them ‘VIPs’ and ‘cherry-on-top’ users. Could you describe their characteristics?

Berkay: I wish there was a definition. The user model is tailored for every single campaign based on the game + geo combination, because users from different geos have different playing and consuming habits.

For a puzzle game in Japan, it can be someone spending a lot of time in the app, progressing faster and staying in the app longer than a regular user. For the same game in the US, a UMT user may have shorter session time but start spending earlier. 

So if you already run a UMT campaign, that model won’t be eligible for another country. In order for it to be as efficient as possible, we need to train it for each specific game + geo combination.

That’s the best part of UMT: it finds the perfect match for each of these combinations.

a: I see! So it depends on how the game makes money; and on the users who behave in the way that supports that. If the game’s mostly ad-monetized, it’s users who spend a lot of time and watch a lot of ads. And if it’s in-app-purchase-based, it’s users who spend a lot on purchases.

Berkay: Yes. There are many data points we collect to train the model. Even within the same genre and geo, the “good user” definition will be different for each case.

a: Since every game would have this top-performance cohort of users, what’s there to do to start with the UMT targeting?

Berkay: The initial requirement for running a UMT campaign is to run a regular targeting campaign in that geo for some time until it acquires a certain threshold of users to train the model on. Once we have enough data, we can start additionally targeting the highest-quality users at the same time. 

My strong suggestion is to keep both types of targeting running so our machine learning model keeps improving based on user behavior coming from the regular campaign as well.

a: You mentioned passing a threshold before being able to start running UMT campaigns, what is it?

Berkay: There isn’t a certain number, but the model usually has enough data for learning after acquiring 1–2K users from the app + geo combination.

a: How fast can someone get to that number and start a UMT campaign?

Berkay: It depends on the game, the time frame, and the geo, because the number of users they can get daily differs between, say, the US and the UK due to demographic reasons. 

A month is a realistic timeline. If they start more aggressively, it’s possible to meet the threshold faster. But there’s also tech work behind it; it’s not just a toggle. The model needs to be deployed, and so on.

a: Are there scenarios when UMT is the wrong move?

Berkay: Leveraging UMT is always a good thing, but if your one and only priority is scale at that time, UMT isn’t the best fit for that expectation. Its main focus is profitability.

a: What’s the most common question you get from advertisers when UMT comes up?

Berkay: The most common question must be around the suggested CPIs for UMT campaigns. Due to the high competition, they’re high as well.

It’s understandable to have doubts when someone asks you to pay double of what you usually pay for a specific type of product when you’re not 100% sure it will work. But based on our experience, UMT delivers substantial results on every geo we run. 

Then, when we share the numbers with advertisers—ROAS D7, retention D7, even ROAS D30—those concerns are quickly relieved.

a: Could you give an example?

Berkay: Yes. For games monetized with in-app purchases, UMT ROAS D7 is double of regular targeting, retention is 40% better on day seven, and on day 30-50% better. So higher CPI is not a problem when ROAS doubles.

a: Do the results ever get stale or go down?

Berkay: No, the model changes with the game for consistently good performance.

Games aren’t constant. Almost all of them have LiveOps and events, and that affects user behavior. The UMT algorithms keep learning and improving over time as partners make updates in their games. If the model doesn’t keep learning, it stays optimized to the old behavior, which we don’t allow. 

a: OK, i’m sold! One more takeaway or tip before we finish?

Berkay: See UMT as diversifying your sources and using them more efficiently. Game studios already build complex products, so a single marketing campaign or just one targeting type is never enough to match that complexity: of the product and the users you’re trying to reach. UMT is one of the best tools to optimize your user funnel as well as your marketing campaigns.

a: Thank you, Berkay!

Berkay: Any time!

Note: UMT is currently available only for Android campaigns, with iOS support launching in the coming months.