Why Building Your Own AI Ad System Is a Worse Idea Than You Think

The question comes up in every founder conversation about AI monetization: why not just build my own ad system? Take the AdMob data, make it look native, problem solved. It sounds reasonable until you understand what “ad system” actually means at the infrastructure level.

Nic Baird, CEO of Koah Labs — the company building a full-service ad network for AI-native applications — has a concrete answer to why DIY isn’t the shortcut it appears to be. His team of 12 engineers spent over a year building their ad infrastructure, and that’s their sole focus.

The technical stack is deeper than it looks

The surface-level problem — making an ad look native — is maybe 10% of the actual work. The real complexity sits in the targeting pipeline, brand safety, copy generation, and real-time relevance matching.

“You could probably do it in a couple of months. Great developers could do it in a couple of weeks,” Nic says about standing up a basic programmatic connection. “The problem is that while you might be able to stand up something that sort of works, there’ll be a lot of issues around brand safety, a lot of issues around verification.”

Koah’s pipeline runs multiple layers of analysis on every query before deciding whether to show an ad at all. First, it checks commercial intent — is the user actually in a buying mindset, or just learning? Then it runs brand safety filters for profanity and inappropriate content. Then it matches against its advertiser database using a combination of semantic search, keyword matching, and demographic signals. Then it generates custom copy. Then it runs one final check to make sure the generated copy makes sense in context.

They only show an ad about 30-35% of the time. The rest of the queries either don’t have commercial intent or don’t have a strong enough match to justify showing anything.

The hallucination problem nobody talks about

When you use LLMs to generate ad copy — which you need to do for native AI ads — hallucination becomes an advertising problem, not just an accuracy problem.

Nic shares a specific example: they had a supplements advertiser, and a user asked about the best national parks. The system correctly identified an opportunity to recommend hiking provisions, but the generated copy started by giving an opinion on which national park in Brazil was best — before getting to the actual product recommendation.

“You wouldn’t expect the advertisement to come in and say, ‘Here’s the opinion that I have on what the best national park is,’” Nic says. These edge cases are what separates a prototype from a production system. Finding and fixing them requires monitoring infrastructure, feedback loops, and constant iteration across thousands of advertiser-query combinations.

The go-to-market wall

Even if you solve every technical problem, you still need advertisers. And getting advertisers means either building programmatic connections (a dozen separate integrations with demand-side platforms) or doing direct sales to brands — which requires a dedicated sales team.

“Going out and actually making the deals with all these demand side players, whether it’s programmatic players — you’d have to make deals with a dozen programmatic players — or if you wanted to go direct, which is how you get the best CPMs, that means having an entire advertising sales team,” Nic explains.

Direct deals are where the real money is. Koah works with brands like Progressive, Choice Hotels, and Intuit. An indie developer building their own ad system isn’t calling Progressive’s media buyer. They’re stuck with whatever programmatic demand they can plug in, at commodity CPMs.

The counterintuitive economics

Here’s the part most founders miss: even if you save the platform fee by building in-house, you’ll likely make less money. Koah’s network-level data — seeing patterns across dozens of publishers — enables targeting precision that a single app can’t replicate. Their 2% CTR across the network (versus 0.4-0.5% industry average for display) is a function of aggregate data, not just better formatting.

“Even if you can do it all, you might still make less money by yourself,” Nic says. “Hopefully we’re driving enough efficiency that even with our platform fee, you’re making more money with us than you would by yourself.”

The build-versus-buy decision for AI advertising isn’t really about capability. It’s about whether your engineering hours are better spent making your product better or reinventing ad infrastructure that already exists.

FAQ

How long does it take to build an AI-native ad system from scratch?

Koah Labs’ team of 12 engineers spent over a year building their ad infrastructure, and advertising is their sole focus. A basic programmatic connection can be stood up in weeks, but production-grade targeting, brand safety, copy generation, and monitoring requires sustained engineering investment well beyond initial setup.

What technical components does an AI ad system need?

A functional AI ad system requires: commercial intent detection, brand safety filtering, demographic and location targeting, semantic ad matching, LLM-based copy generation, a verification layer to catch hallucinated copy, real-time monitoring for quality issues, and connections to demand-side platforms or direct advertiser relationships. Each component adds complexity.

Can you use AdMob data to build native AI ads?

No. AdMob and similar mobile ad networks don’t provide raw advertisement data for custom rendering. They control the ad display format. Building native AI ads requires connecting directly to programmatic demand-side platforms via specs like OpenRTB, or establishing direct relationships with advertisers — both require significant infrastructure.

What is the difference between programmatic and direct ad sales for AI apps?

Programmatic connects your inventory to automated ad exchanges at commodity CPMs. Direct sales means negotiating individually with brands like Progressive or Choice Hotels for higher CPMs and more relevant placements. Direct deals generate significantly more revenue but require a dedicated sales team and established credibility.

How does Koah Labs achieve 2% click-through rate vs industry average 0.4-0.5%?

Koah’s CTR is roughly 4x the display ad industry average because they combine first-party query data with semantic matching to show only highly relevant sponsored content. They also suppress ads on approximately 65-70% of queries where relevance isn’t strong enough, meaning every shown ad has high intent match.

What are the brand safety risks of AI-generated ad copy?

LLM-generated ad copy can hallucinate claims about products, express opinions the advertiser didn’t authorize, generate copy in the wrong language, or produce contextually inappropriate recommendations. Koah encountered cases where ad copy offered opinions on travel destinations before promoting a relevant product — harmless but unprofessional.

Is it cheaper to build your own ad system or use a platform like Koah?

Building in-house saves the platform fee but typically generates less total revenue. Network-level data across dozens of publishers enables targeting precision a single app can’t match. The engineering cost of maintaining brand safety, monitoring, and advertiser relationships also exceeds the platform fee for most teams.

What does $5 CPM mean for an AI app developer?

At $5 CPM (cost per thousand impressions) with a 30% fill rate, a developer earns approximately $1.50 per thousand queries. For an app processing 100,000 queries per day, that’s roughly $150 daily or $4,500 monthly in ad revenue — often enough to cover inference costs and reach sustainable unit economics.

Why do AI ad startups need 12+ engineers?

The challenge spans targeting algorithms, real-time bidding infrastructure, LLM copy generation, brand safety systems, SDK development for multiple platforms, monitoring dashboards, and advertiser-side tooling. Each component requires specialized engineering, and they must work together in real-time at sub-second latency to avoid degrading the user experience.