Blog: Higher Ad Costs, Slower Growth, Wrong Leads: What Happens When You Let Paid Media AI Run Without Guardrails
Key Takeaways
- The job of paid media has shifted from controlling levers to setting guardrails:
Platforms now make most of the execution decisions. What a human decides — signals, exclusions, conversion targets, campaign structure — determines what the machine is optimizing toward. Skip that setup, and the platform doesn't pause. It just runs. - AI systems optimize precisely toward whatever signal they're given — right or wrong:
Bot traffic, spam leads, and unqualified clicks all look like success if that's what the campaign is configured to reward. The platform can't tell the difference. That distinction is still a human job. - Most of the damage happens at the settings level, not the strategy level:
Brand exclusions left off, conversion feedback never set up, campaign structure too fragmented to exit learning mode — none of these get flagged by the platform. They're defaults, and defaults are still decisions.
Table of Contents
For most of Jordan Crawford’s career in paid media, the job was control. Every targeting parameter, every audience segment, every bid adjustment. Value came from how precisely you could operate all of it.
That’s changed.
“The job of paid media managers really changed from trying to control all of the levers to more focusing on what are the signals I’m sending the platform so it can learn what’s a profitable customer for me and then what are the guardrails that I’m setting.”
It’s tempting to hear “fewer levers” and assume the job got easier. It didn’t. “Maybe fewer levers, but those levers matter more.” The platforms make far more decisions now. What’s left for a human to decide is what to tell the machine before it starts pulling them on its own.
Skip that step, and the machine doesn’t pause to ask. It just runs.
What Happens When AI Controls Your Ad Creative Without Explicit Settings
The clearest example of what goes wrong sits in creative, an area that seems like it should be the easiest thing to hand over to AI.
“Meta has like 32 settings now that are all related to AI content,” Jordan says, “and you have to set them at the ad level.” The detail that matters most: many aren’t opt-in. They’re set automatically.
What does opting in by inaction look like? “Meta can just put a random song on your ad, and I’ve seen it happen. Rap music with a lot of profanity, for example, on an ad from a super polished brand that would never approve that.”
The platform did exactly what it was configured to do. That’s the point. The configuration is the decision, whether or not anyone consciously made it.
Why AI-Driven Campaigns Attract Bot Traffic and What It Costs You
If creative automation is the visible risk (the one that shows up as an embarrassing ad, for example), there’s a quieter risk underneath it, in what platforms are told to optimize toward.
“This bot traffic has gotten even worse with AI,” Jordan says. Conversion types like click-to-calls and lead form fills are “really susceptible to bots,” and if a campaign optimizes toward raw lead volume without distinguishing real prospects from bot noise, the platform delivers exactly what it’s told to value. More leads. More clicks. More calls. None of it is necessarily real.
This connects to where ads actually run. Jordan flags Google Display Network, Google Search Network, and Facebook Audience Network as placements that tend to be “bot-heavy,” places where monitoring and explicit settings matter more than a top-line number suggests.
The throughline: AI systems are extremely good at optimizing toward whatever signal they’re given. They can’t ask whether that signal is the right one. That’s still a human job.
How Machine Learning Actually Learns (and What It Needs From You)
There’s a second layer here, and it’s counterintuitive for media managers trained in the “more control, more segmentation” era.
Since iOS tracking changes limited how precisely platforms can target individuals, Meta leans harder on pattern recognition based on behavior and creative response.
Jordan’s example: “Jordan is active on Reels at 9pm, and spends a lot of time watching videos about dogs. Jordan behaves very similarly to these thousands of other people in the U.S.” The platform builds a behavioral fingerprint and matches ads to people with similar patterns. So, creative becomes the primary lever, since granular audience targeting is largely gone.
But this kind of pattern recognition needs volume. “For Meta to be good at that, it has to have enough data to build a good pattern.”
The thresholds: Meta needs roughly 50 conversions a week at the ad set level to exit “learning mode” and perform consistently. Google’s threshold is lower, but the same dynamic exists: roughly 30 conversions over 30 days.
This is where the old instinct toward control can become a liability. A media manager trained to segment heavily — separate campaigns for separate audiences, separate ad sets for separate offers — can end up with ad sets that never hit enough conversions to exit learning mode.
Each segment might make perfect strategic sense alone. But if no segment can hit 50 conversions a week, none of them get the chance to perform the way the algorithm is designed to perform once it’s learned.
The job now: “finding the middle ground between having enough control over your budget and where you want to spend, but also giving the platform enough data to learn from.”
Too much segmentation starves the system. Too little structure hands over decisions that should stay with a human. The guardrails sit in that middle ground.
The Paid Media AI Guardrail Checklist: Five Settings Most Accounts Get Wrong
A consistent set of things to check because each is so often left on a default nobody chose deliberately.
Brand Exclusions
Performance Max, left at default, “will automatically serve both branded and non-branded terms,” and because branded terms tend to be the cheapest to win, PMax gravitates toward them fast.
Roughly 90% of accounts audited don’t have this addressed, meaning a campaign meant to reach new prospects often functions as a branded search campaign in disguise.
Platform Optimization Targets
An estimated 80-90% of accounts optimize toward leads and phone calls without feeding back which converted further down the funnel, or which were spam.
Without that feedback, the platform can’t tell a real prospect from a bot and has no reason to try.
Where Ads Run
Display Network, Search Network, and Audience Network placements carry higher bot exposure. These placements need active monitoring and explicit settings.
Left on default, the platform decides where to spend.
Tracking Infrastructure
With up to 50% of Americans using ad blockers that block pixel-based tracking, server-side tracking — a direct connection between the ad platform and a business’s server, bypassing the browser — recovers data that would otherwise be lost.
Without it, the platform decides based on an increasingly incomplete picture.
Campaign Structure & Learning Thresholds
Instead of maximizing segmentation, build structure around whether each ad set can realistically hit the conversion volume (50/week for Meta, 30/30 days for Google) needed to exit learning mode and perform as designed.
None of these are complicated on their own. What makes them matter: almost none of them are decisions a platform will surface or flag.
They’re settings, and settings, left untouched, are still choices. Nobody made them on purpose.
What Good Paid Media Management Looks Like in an AI-First World
If the work shifted from pulling levers to setting guardrails and reading signals, what makes someone good at the job shifted, too.
“Previously a successful media manager used to over-index on detail-orientation, because a lot of the job was about control,” Jordan says. “Now we’re really screening for big-picture thinking, connecting what’s happening in a campaign to what it means for the business.”
Client communication matters more, not less, as automation increases. “At the end of the day, businesses are going to need someone to hold accountable, someone to make decisions, and communicate those decisions well.” As more execution becomes invisible, someone still has to explain what’s happening and why.
Deductive reasoning and structured diagnosis matter too: “These AI tools are probably going to spit out thousands of suggestions, but knowing what to do with that is going to be a big one.” The volume of output isn’t the bottleneck anymore. Judgment about what matters is.
And creative strategy and customer research (understanding who the audience is and how to reach them) is “more and more an important job of the media manager,” precisely because creative is now the primary lever left for influencing how these systems perform. The data backs this up — see what $1.3 billion in Meta ad spend reveals about creative performance.
The Bottom Line on AI-Driven Paid Media
This isn’t an argument against AI in paid media. These systems perform well when given what they need. The argument is narrower, and more urgent: these systems need to be given something. Signals. Exclusions. Volume thresholds. Guardrails.
None of that happens automatically, no matter how automated the rest of the campaign looks.
“Automated” describes what happens once the inputs are set. It says nothing about whether those inputs were ever chosen on purpose. The businesses getting hurt by AI-driven paid media right now usually aren’t the ones using AI badly.
They’re the ones who never realized there was a decision to make in the first place.
FAQs
What paid media settings should I audit first if I'm running Performance Max or Meta campaigns?
Start with brand exclusions on Performance Max — roughly 90% of accounts audited don't have them set, which means a prospecting campaign is often quietly spending on branded terms. On Meta, review your AI creative settings at the ad level; many are opted in by default, including music overlays and creative enhancements that may not match your brand. Then check what conversion event your campaigns are optimizing toward, and whether you're feeding back any downstream signal (appointments, qualified leads, purchases) beyond the raw form fill or call.
How do I know if my campaign is stuck in learning mode?
Both Meta and Google surface learning mode status in their interfaces, but the more practical signal is performance instability — results that swing significantly week to week without a clear external cause. Meta requires roughly 50 conversions per week at the ad set level to exit learning mode and perform consistently; Google's threshold is approximately 30 conversions over 30 days. If your campaign structure is too fragmented to hit those thresholds in any individual ad set, no amount of creative testing or bid adjustment will unlock stable performance.
If AI is handling more of the execution, what should a paid media team actually be spending time on?
Three things matter most now.
- 1. Signal quality: making sure the platform is receiving accurate, downstream conversion data that reflects real business outcomes, not just top-of-funnel events susceptible to bot traffic.
- 2. Creative strategy: because creative is now the primary targeting lever, the work of understanding what message resonates with which audience is more important than ever.
- 3. Structured diagnosis: as platforms generate more automated recommendations, the skill is knowing which suggestions reflect a real opportunity versus the platform optimizing toward its own metrics.
