AI-powered advertising works well in theory. Set a goal, let the algorithm optimise toward it, watch the results come in. In practice, especially in a vertical like iGaming, it rarely goes that cleanly.
Tracking breaks mid-campaign. KPIs shift. The app you are driving users to develop store issues you only find out about after the budget is gone. And through all of this, the AI tools you are depending on keep optimising toward whatever signal they can find, which may not be the one that actually matters to your business.
Getting real value from these tools is less about which platform you use and more about the foundations you put in place before the campaign goes live.
When the setup is messy
Last year we ran a campaign for an iGaming app client with clear enough KPIs: installs at around $8, registrations at $15 to $25, first-time deposits at $45 to $55. The goal was to hold that funnel stable while scaling volume. What we were actually dealing with was tracking that dropped in and out, app store complications affecting users downstream, and KPIs that shifted as the campaign progressed.
Over three months, we delivered 97,674 installs, 21,134 registrations, and 12,701 deposits. But that result came from constant adjustment, not from setting up a campaign and letting the algorithm take over.
Optimising toward the wrong signal
CPA Goal bidding is useful as it finds traffic that converts toward a specific action quickly, which makes it well suited to testing and early data gathering. The issue is that it only optimises toward the first conversion event. If that event is an install, the system will find installs and it has no way of knowing whether those users will register, deposit, or ever return.
In iGaming, where the value of a user depends almost entirely on what happens after the install, that creates a real gap between what the algorithm is rewarding and what the business actually needs. This is a limitation that needs to be accounted for.
Data quality comes first
AI optimisation is only as good as the data going into it, and that is easy to say and easy to skip over when there is pressure to get a campaign live. In this campaign, the tracking inconsistencies meant we were working with incomplete signals for stretches of time. Rather than try to compensate with more automation, we slowed down, identified what data we could actually trust and built decision rules that could hold up even when the picture was partial.
We gave campaigns a minimum of two days to stabilise before making any changes and monitored performance across operating systems, devices and browsers to separate real trends from noise. When the data eventually improved, the AI performed better too. The tools themselves did not change, it was the quality of what we were feeding them that did.
Where human judgment fits in
The campaign worked because automation and manual optimisation were doing different jobs, not because one replaced the other. We used CPA Goal during the testing phase to identify which traffic zones were generating real conversions, then moved those zones to SmartCPM for more precise control over how we weighted and scaled the traffic. Automation was better at discovery. Manual bidding was better at refinement.
We also built decision rules for managing traffic quality. A zone spending the equivalent of two deposits without generating a single downstream conversion was cut immediately. A zone producing high registrations but no deposits was blacklisted. Zones that consistently delivered full-funnel results at good cost were whitelisted and protected.
Building that whitelist took around two months, because users who eventually deposit take time to get there. Any framework that does not account for the full conversion window will undervalue exactly the zones that produce your most valuable users.
What actually makes AI work
AI tools perform well when the conversion events they are tracking reflect what the business needs, when the tracking is reliable and when there is enough clean data for the system to work with before decisions get made. They also perform better when human judgment is being applied to the questions an algorithm cannot answer: which KPIs to prioritise, how to weight different stages of the funnel and when to trust the data versus when to push back on it.
The campaigns that get the most from AI are not always the ones with the most automation. They tend to be the ones where human and automated decision-making are clearly separated, each doing what it does best, with solid data connecting the two.
