RTB Intelligence Engine AI Plugin for Revive Adserver

The RTB Intelligence Engine brings machine-learning-driven pricing intelligence to your real-time bidding pipeline. It combines two complementary capabilities — Predictive Bid Optimization (Bid Shading) on the demand side, and Dynamic Floor Price Optimization on the supply side — to help both DSP and SSP operations price inventory more accurately, in real time, without adding latency to the auction path.

1. Predictive Bid Optimization (Bid Shading)

In a real-time ad auction, every bidder has a maximum price they're willing to pay — usually the full eCPM. If everyone simply bid their maximum every time, whoever wins would often pay far more than necessary, since most auctions only require bidding slightly more than the next-highest bidder to win.

Bid shading means bidding less than the true maximum value, but by just the right amount — low enough to save money, but still high enough to keep winning close to the same rate. A machine learning model estimates that number — the likely minimum clearing price — for each incoming bid request, and the bid is shaded down from the raw value accordingly.

What the model looks at:

• Publisher / site ID — a premium news site's inventory is generally worth more than a low-traffic blog's, even for an identical ad size
• Ad size — a large video ad and a small banner carry very different market values
• Device / geo — mobile vs. desktop vs. tablet traffic prices differently, and so does geography
• Hour-of-day — prime-time and lunchtime peaks see more advertiser competition than late-night hours
• Historic win rate — tells the model whether current bidding levels are winning too easily (room to shade more) or losing too often (shading too aggressively)
• Floor price signals — the publisher's bid floor is usually the strongest signal of where the clearing price will land

Together, these let the model shade differently for a premium metro video slot at 9pm versus a low-value tier-3 banner slot at 3am, instead of applying one fixed percentage rule everywhere.

2. RTB Dynamic Floor Price Optimization

The floor price is the minimum an SSP/publisher will accept for an impression. A static floor (e.g. “Publisher A → Floor = $1.00 CPM”) ignores the fact that inventory value changes continuously by publisher, format, geo, device, time, and demand. Set it too low, and valuable inventory sells too cheaply; set it too high, and auctions go unfilled.

Dynamic Floor Price Optimization predicts the appropriate floor for each incoming bid request instead of applying one number everywhere — for example, a premium US video placement in the evening might get a predicted floor of $4.20, while a low-value mobile banner late at night gets $0.65. The goal is the best balance between CPM/revenue and fill rate.

What the model looks at:

• Publisher / Site ID and Ad Zone / Placement ID — premium publishers and placements (e.g. homepage video) support higher floors than lower-value inventory
• Ad Format / Ad Size — video, native, banner, and CTV all carry different market values
• Device Type and Geo / Country — demand and clearing prices vary significantly across both
• Hour of Day — peak traffic supports higher floors; low-demand periods may need lower floors to maintain fill rate
• Historical Clearing Price, Win Rate / Fill Rate, and Bid Price — some of the strongest signals for where a floor should sit and whether the current one is too high or too low
• Current Floor Price — the model optimizes around the existing baseline rather than making unrestricted changes
• Publisher Revenue / eCPM and Traffic Volume — helps the model judge whether a floor change actually improves monetization, and how much confidence to place in low-volume inventory

Safety constraints are applied on top of the raw prediction — for example, if the existing floor is $2.00 and the model predicts $3.50, the system can cap the increase (e.g. to $2.50) rather than jumping straight to the raw prediction, preventing unexpected price swings from abnormal predictions.

Built for Production Safety

Both features are designed to only ever make bidding better, never introduce a new way for it to break:

• Runs as a local, low-latency inference step inside your existing Go-based DSP/SSP — no external API call needed at bid time
• Integrates with a simple on/off configuration switch, so operationally nothing changes about how the feature is enabled
• Automatic fallback — if the model fails to load, errors out, or exceeds the latency budget, the system automatically falls back to your existing bidding or floor-price logic rather than failing the auction
• Supports gradual rollout via A/B testing — expose the feature to a small percentage of traffic first, monitor fill rate, win rate, clearing price, eCPM, revenue, and latency, and scale up only once improvement is confirmed

Benefits

• Lower cost per win — bid shading reduces what you pay to win an auction without meaningfully reducing your win rate
• Better-balanced floor pricing — floors adjust to actual inventory value instead of one static number, protecting revenue on premium inventory and fill rate on lower-value inventory
• No added latency — both models run as local inference inside your existing Go codebase, not as external calls
• Safe by default — automatic fallback to existing logic means the AI layer can never be the cause of a failed auction
• Improves over time — models are designed to be retrained periodically on new auction outcomes, adapting to shifts in demand, seasonality, and market conditions
• Low-risk rollout — built-in A/B testing lets you validate real revenue and fill-rate impact on a small slice of traffic before going all-in