Streaming killed the thing that made TV easy to buy. For about fifty years, you could point at a show, a slot, and a price, and everyone in the room understood the deal. That world is gone. Your audience now watches a prestige drama on a free ad-supported app at 6:40 a.m. on a phone propped against a coffee mug, and the person next to them on the couch is scrolling a cooking video on a different device entirely. If your media plan still assumes one screen and one sitting, you’re paying for impressions nobody is actually present for.
Here’s the good news buried under all that fragmentation. The same mess that broke TV buying also produced a flood of behavioral signals, and machine learning is the only realistic way to read it at scale. Advertisers who treat connected TV as a data problem rather than a slot-buying problem are pulling better results from smaller budgets. That’s the whole shift, and it’s worth understanding before your next planning cycle.
Why the Old TV Buying Playbook Stopped Working
Traditional TV worked because scarcity did the targeting for you. Three networks, a limited number of prime-time slots, and a broad audience meant a decent creative could carry a mediocre plan. Reach was the metric, and reach was easy to buy.
Streaming inverted all of it. Viewership spread across subscription video on demand, free ad-supported TV channels, virtual multichannel video programming distributors, and network apps, each with its own inventory, its own measurement quirks, and its own idea of what a completed view means. You can still buy reach. You just can’t assume reach equals attention.
The Federal Trade Commission has spent real energy on how ad tech describes its own targeting and data practices, which tells you something useful: the machinery behind modern advertising is complicated enough that regulators feel the need to keep an eye on the promises vendors make. If you can’t explain how a platform decides who sees your ad, you’re trusting a black box with your budget.
What Deep Learning Actually Does With Your Data
Strip away the vendor vocabulary and most AI advertising platforms do one of two things. They sort you into a prebuilt audience segment, or they build a model around your specific outcomes. The second approach is where the interesting work happens.
A model trained on your own conversion data learns the difference between a viewer who watched thirty seconds and left and a viewer who watched, wandered into the kitchen, came back, and then visited your site the next morning. On its own, each signal looks like noise. Stacked across thousands of households and dozens of touchpoints, patterns show up. Households that saw a particular creative sequence during a specific daypart tend to convert at a higher clip. Households exposed to the same ad five times in two days tend to go quiet.
That’s not clever marketing language. That’s just what happens when you let a system chew through first-party data long enough to find the structure inside it. The National Institute of Standards and Technology publishes foundational work on how AI systems behave, including where they fail, and it’s worth a read if you’re evaluating any vendor claiming machine learning is doing the heavy lifting on your behalf.
My honest read after watching these platforms up close: the ones that win are not the ones with the fanciest model. They’re the ones that can explain what their model learned. If a sales rep can’t tell you which signals matter most for your account, walk.
The Three-Bucket CTV Budget Split I’d Use
Most marketers either over-fund CTV or treat it as a rounding error. Neither works. Here’s a split that holds up across account sizes, and it’s the one I’d defend in a planning meeting.
- Prospecting bucket (roughly half your budget): Broad reach across premium streaming inventory, optimized by the platform’s model rather than by your assumptions about which networks matter.
- Signal collection bucket (about a third): Deliberately spread across dayparts, device types, and creative variations so the model has something to learn from. This bucket funds knowledge, not impressions.
- Retargeting and sequencing bucket (the rest): Tight windows, specific creative sequences, aimed at households that already showed interest.
Review the split monthly, not quarterly. Streaming audiences move faster than quarterly planning cycles, like to admit, and the models get sharper as data accumulates. If you’re already working with a performance CTV ads agency powered by AI algorithms, ask them how they’d weight these three buckets for your account. The answer tells you a lot about whether they’re actually modeling your business or just renting you inventory.
A Tuesday Morning Flight Change, and What It Taught Me
Earlier this year I watched a mid-size advertiser move roughly a fifth of its streaming budget from evening to early morning slots. Everyone involved hated the idea. Early morning streaming inventory is cheap for a reason, or so the assumption went.
The model had flagged something odd.
