Most publishers and game studios already have the demand they need. Several networks, a mediation layer, a Google Ad Manager account, sometimes a direct AdX seat. The revenue they lose rarely comes from a missing partner. It comes from the hours between a performance change and someone noticing it.
Take a studio earning $200 an hour from ads. A tag starts decaying at 6pm. The ad ops team reviews performance the next morning. Six hours of degraded fill and lower CPMs go by with no correction, and roughly $720 disappears with no way to recover it. That cycle repeats a few times a week at most studios. Nobody on the team is doing anything wrong. A human reviewing dashboards every four to six hours simply cannot react faster than that.
This is the specific gap AI closes. Everything below is about where the technology actually moves the number, in what order to adopt it, and what it will not do for you.
Why the timing matters right now
Two things changed recently that make automated pricing worth more than it was a year ago.
First, Google removed Unified Pricing Rules from Ad Manager in December 2025 after antitrust rulings in the US and EU. Publishers can now set bidder-specific floors again, so one buyer can be held to a $5 minimum while another competes at $2. That flexibility had been gone since 2019. Managing hundreds of bidder-specific rules by hand is not realistic, which makes this a change that only pays off with automation behind it.
Second, the money at stake keeps growing. Mobile game ad revenue reached $62.1 billion globally in 2025, and programmatic networks now run in 73% of the top 200 games. Small percentage gains on a base that size are worth real money.
Where AI actually makes an impact
1. Floor prices that move with demand
A static floor is a guess about what your inventory is worth, and it goes stale within days. Set it too high and fill collapses on inventory that would have sold at a lower but still profitable CPM. Set it too low and you hand away impressions that buyers would have paid more for.
A machine learning system reads live bid data by geo, device, ad unit, and hour, then adjusts the floor per query instead of per week. The mechanism is simple: every auction outcome feeds the model, so the next floor is calibrated against what buyers actually paid rather than what someone estimated last month. Our breakdown of how AI floor price optimization works covers the math in more detail, including why the failure mode of a too-high floor looks like high CPM paired with low fill in the same demand tier.
For web inventory where refresh is in play, the floor has to be recalculated on every reload as well, since a second impression on the same page has a different value than the first. That combination of viewability-gated refresh and per-reload pricing is what UndrAds runs on the ad refresh layer.
2. Catching revenue drops before they compound
This is the highest-value lever for anyone running a waterfall. Demand partners go cold. Tags decay. eCPM falls over a weekend. The system knows within seconds; the team finds out at the next review.
An always-on model watches your key metrics continuously, flags the anomaly the moment it starts, and switches tags or reorders the waterfall inside the hour. More advanced setups go further and learn the decay curve of specific tags across geos and time windows, then switch preemptively before the decline shows up in reporting. What autonomous ad operations means in practice walks through the difference between reacting to a drop and preventing one.
The gain here is not a higher peak CPM. It is a shorter loss window on every incident, several times a week, compounding across the year.
3. Format and timing decisions per user
Ad format drives eCPM more than almost anything else. Business of Apps put 2024 in-app eCPMs at roughly $2.80 for banners, $4.80 for interstitials, and $10.50 for rewarded video. In US markets the spread is wider still, with rewarded video averaging $16.49 on Android and $19.63 on iOS.
Knowing rewarded pays more is the easy part. The harder question is when to show which format to which player, and AI handles that better than a fixed rule. Showing an interstitial to a player who just failed a level produces a different completion rate than showing it after a win. Models that segment on session depth, spend history, and past ad engagement pick the moment, and AI-based ad personalization has been measured lifting eCPM by 12% across Android.
Retention is the constraint to respect here. Games that keep interstitials under three per session retain 27% more users, so any system optimizing purely for impression count will cost you more in churn than it earns in fill. If rewarded placements are underused in your app, the role of rewarded ads in app monetization is a good place to start, and our eCPM benchmarks by format and region give you numbers to compare against.
4. Waterfall management and the move to bidding
In-app bidding removes most of the manual tier maintenance that eats an ad ops week. That does not eliminate the problem, since almost every studio runs a hybrid setup with legacy waterfall lines still carrying meaningful volume, and those lines still need reordering as partners shift.
AI handles the reordering continuously instead of on a weekly cadence. It also balances Google bidding against external partners in real time and applies timeout protection so slow bidders stop adding latency. If you are weighing a mediation change as a way to fix revenue, read our take on AppLovin MAX alternatives first. Switching mediation platforms usually costs 60 to 90 days of algorithm relearning, and it does not address reaction speed at all.
5. Fill rate and regional pricing
Tier 2 and Tier 3 traffic has thinner demand and lower CPMs, which means the optimization goal changes. Chasing peak eCPM in those geos leaves inventory unsold. Automated systems set separate floors per region and prioritize fill, and offerwall placements in particular have lifted ARPU by 17%, concentrated in Tier 2 and Tier 3 markets.
Publishers who cannot get a direct AdX seat are usually leaving the deepest demand pool untouched. UndrAds is an authorized Google partner for that access, with the same agent setting floors and demand priority per query once the seat is live.
What AI will not fix
Being clear about the limits saves you a wasted quarter.
It will not create demand that does not exist. If your geo mix is thin and no buyer wants that inventory at any price, better pricing produces a smaller loss rather than a gain.
It will not rescue bad placements. An ad that covers a gameplay button or fires mid-level will lose you users regardless of how well it is priced.
It will not repair a broken app experience. Retention drives lifetime ad revenue more than eCPM does, and no monetization layer compensates for a game people stop opening.
It will not replace your ad ops team. What it removes is the review-and-react loop, which frees that team for partner negotiation, placement testing, and direct deals. How AI agents are changing ad operations covers where the human work moves.
A rollout order that works
- Pick one app or one site section. Testing across your whole portfolio makes attribution impossible and raises the stakes for no reason.
- Record a clean 14-day baseline. eCPM, fill rate, ARPDAU, impressions per session, and revenue by geo. Without this you cannot prove anything later.
- Start with floor pricing. It is the fastest lever to see results on and needs no app changes. Look for movement inside two weeks.
- Add drop detection and automated tag switching. This is where the largest recovery sits, and it only becomes measurable once floors are stable.
- Layer in format and frequency optimization last. It touches user experience, so it needs a longer read window and retention monitoring alongside revenue.
- Compare at 30 days, decide at 60. Anything shorter is noise. Anything longer is a delay costing you money.
Keep the comparison on total revenue per daily active user rather than eCPM alone. eCPM can rise while revenue falls if fill drops at the same time.
Manual ad ops versus an AI layer
| Manual ad ops | AI-assisted | |
|---|---|---|
| Performance review | Every 4 to 6 hours | Continuous |
| Floor price updates | Weekly or monthly | Per query |
| Reaction to a drop | After a person spots it | Within the hour |
| Weekend and overnight coverage | Limited or none | Unchanged |
| Bidder-specific floors | Impractical at scale | Standard |
| Loss per incident at $200/hr | $720 and up | Materially reduced |
Industry estimates for automated optimization workflows sit in the range of 20% to 30% above conventional rule-based systems, though that figure varies widely by starting setup. A studio already running tight bidding will see less than one running a stale waterfall.
How to evaluate an AI monetization vendor
Ask these five questions before signing anything.
What does integration actually require? API access to your existing GAM or mediation account should be enough. If a vendor asks for an SDK and an app release, your timeline just grew by weeks and your dev team is now involved. Our primer on what Google Ad Manager is and how it works explains where that API layer sits.
Can you see every action it takes? Tag switches, floor changes, and waterfall reorders should be logged and visible to your team in real time. Vendors who will not show you the actions are asking you to trust an unaudited system with your revenue.
Who owns the data? You should be able to verify the lift in your own reporting rather than in a vendor dashboard.
Can you turn parts of it off? Slot-level and placement-level control matters. A system with no override is a risk on day one.
Is there a trial with no contract? A 10-day or 30-day test on one property tells you more than any case study.
FAQ
How quickly does AI optimization show measurable results?
Floor pricing changes usually show inside 7 to 14 days because the feedback loop is short. Format and timing models need 30 to 60 days to gather enough behavioral data. If a vendor promises meaningful lift in 48 hours, ask what they are measuring.
Does automated tag switching risk policy violations?
Not if the system is built to operate inside platform rules. Floor adjustments, pricing tiers, and mediation ordering are all standard publisher controls. The risk sits in ad refresh and frequency, where firing too fast or reloading out of view can trigger policy problems, so viewability gating on refresh is the thing to verify.
Do we need to leave our current mediation platform?
No. An optimization layer sits above mediation and works with MAX, LevelPlay, AdMob, or a GAM waterfall as they are. Switching mediation platforms costs 60 to 90 days of algorithm relearning and does not address reaction speed, which is usually the actual problem.
What happens to our ad ops team?
The monitoring and manual switching work goes away. Partner relationships, direct deals, placement strategy, and reviewing what the system decided all stay. Most teams end up doing higher-value work with the same headcount.
Is this worth it for smaller publishers?
Below roughly $5,000 a day in ad revenue, the recovered amount per incident is often too small to justify a vendor relationship. The floor pricing and format principles still apply, and you can implement a simplified version manually. Larger operations are where the delay cost compounds fast enough to matter.
Does AI help with web inventory as well as in-app?
Yes, and the levers overlap. Floor pricing, demand allocation, and drop detection work the same way on web. What is specific to web is refresh logic, where floors need recalculating on every reload and viewability has to gate the refresh itself.
Find out what the delay is costing you
The number worth knowing is how much revenue leaks between a performance drop and your team’s reaction. It is calculable from your existing reporting: average hourly ad revenue, typical drop depth, and hours to correction, multiplied by incident frequency.
Request a revenue leak audit and we will run that calculation against your GAM or mediation data, then show you which of the levers above applies to your setup. No SDK, no app changes, and no commitment to see the numbers.
