YieldLab

Programmatic ads in a world fueled by AI

A hands-on tour of Connected TV ad tech: first the basics โ€” DSPs, SSPs, exchanges, and how an impression is bought. Then the machine learning powering those decisions today. Finally, where agentic AI is taking programmatic buying next.

Built end-to-end (FastAPI + scikit-learn + Next.js) on fully synthetic data โ€” models, simulation, a live auction engine, and an autonomous buying agent.

Act I ยท Ecosystem 101Act II ยท The ML pipelineAct III ยท Agentic AI
Act I ยท The basics

Who's who in a programmatic transaction

You can't reason about AI in ad buying without the plumbing. Start here.

Before the ML: a 60-second map of who's who in a programmatic ad transaction. Every impression is a tiny auction that happens in the time it takes a stream to load.

โ†’
โ†’
โ†’
โ†’

The buy-side software an advertiser (or agency) uses to purchase impressions across many publishers โ€” setting bids, targeting, budgets, and pacing. This is where ML decides how much to bid and how to buy.

Money flows left โ†’ right; the bid request flows right โ†’ left. The DSP and SSPare the two "brains" where AI/ML makes the fast decisions.

Programmatic

Buying and selling ads via software and auctions instead of manual insertion orders โ€” the default for digital and CTV today.

RTB (open auction)

Real-time bidding: DSPs bid per impression in a live auction above a publisher floor. Efficient and flexible, but delivery is variable.

PMP

Private marketplace: an invite-only auction where select buyers bid on premium inventory, often at negotiated floors.

Programmatic Guaranteed

A locked deal โ€” fixed price and volume, no auction. Guaranteed delivery, predictable cost, less flexibility.

CPM / eCPM

Cost per thousand impressions is how inventory is priced. eCPM (effective CPM) blends price with the odds an impression actually performs.

CTV

Connected TV โ€” streaming on a TV screen. High completion rates and premium CPMs, increasingly bought programmatically.

Where AI/ML plugs in

Publishers use ML to forecast inventory and optimize yield; buyers use it to predict outcomes and set bids โ€” all in milliseconds, at billions of auctions a day. The pipeline below shows those decisions; the last act shows where agentic AI takes it next.

Next up: see how machine learning actually runs the buy.