Vijay Kumar, Founder & CEO at Mile tells us what last mile optimisation actually requires...

Machine learning and real-time optimisation have been standard practice on the buy-side of programmatic advertising for nearly a decade. Demand-side platforms make millions of impression-level decisions per second - adjusting bids, selecting inventory, and optimising creatives in real time based on predictive models trained on historical auction data. Yet on the sell-side, most publishers are still managing yield with static rules updated weekly by hand.

This isn't because publishers don't understand the value of optimisation. It's because the infrastructure required to do ML-driven yield optimisation at scale is vastly more complex than most realise - and the entities that have built it (exchanges and SSPs) aren't neutral parties aligned with publisher interests.

The Neutrality Problem

Exchanges have sophisticated machine learning systems. They optimise auction dynamics, bid shading, and transaction efficiency. But they're optimising for transaction volume and marketplace health, not publisher revenue. An exchange maximises GMV; a publisher maximises RPM. These aren't always the same objective.

The result is a "last mile" optimisation gap. Publishers control the final set of decisions that determine auction outcomes - price floors, bidder participation, timeout settings, refresh logic - but lack the infrastructure to optimise them systematically. They're left relying on manual segmentation rules that can't keep pace with the impression-level variance in their inventory. This is where independent technology providers come in, offering ML infrastructure that sits on the publisher's side of the transaction.

What Last-Mile Optimisation Actually Requires

Building real-time ML yield optimisation requires a complete data and decisioning infrastructure that most publishers vastly underestimate. It starts with impression-level data capture across eight critical dimensions: auction characteristics and outcomes, bidder performance metrics, contextual signals like geo and device, user engagement patterns, revenue outcomes, demand configuration, real-time auction landscape, and external data from exchanges and third-party sources. Every impression generates a structured data record - a mid-sized publisher serving 2 billion impressions per month generates approximately 60 million feature vectors daily, each containing 40-80 individual signals depending on the decision type.

This data feeds into a complete ML systems pipeline. Feature engineering transforms raw auction data into predictive signals - a single impression might generate 50+ features combining historical CPM patterns, session-level engagement trajectories, and bidder performance indicators. Models are trained on historical outcomes using the previous 30-90 days of data, with custom models per site cluster rather than one-size-fits-all approaches. In production, models are retrained daily and deployed within hours if performance improvements are confirmed, ensuring adaptation to evolving traffic patterns and seasonal shifts.

The hardest technical requirement is the inference engine, which must return decisions in under 50ms because it sits in the critical path of the ad request. This prediction then integrates with Prebid's native module system to execute before the auction runs—setting impression-specific price floors, providing bidder-level signals through RTD modules, dynamically adjusting bidder timeouts, and selectively enabling or disabling bidders based on predicted value. All of this happens in parallel, in under 100ms, before the live auction begins.

The only way to prove ML impact is through always-on A/B testing. For price floors, 80% of ad requests get dynamic floors while 20% run without floors as a control group, using Prebid's native skipRate feature. For RTD and timeout optimisation, 80% of user sessions get ML-optimised parameters while 20% run baseline configuration, with cookie-based assignment ensuring consistent treatment. The control group isolates seasonality - market shifts, demand surges, budget cycles—so measured lift reflects true causal impact rather than marketplace movements.

Beyond Dynamic Floors

Most publishers think of ML optimisation as "dynamic price floors," but that's just the entry point. The same data infrastructure enables dynamic bidder filtering based on predicted value, timeout optimisation to balance latency versus revenue, refresh optimisation for viewability-based ad refreshes, real-time pricing rules for Google Ads and DV360, and session-level yield maximisation across multiple page views. Publishers running only static floors are leaving the majority of optimisation opportunity untapped.

Real-World Validation and Honest Tradeoffs

We've validated ML-driven yield optimisation across diverse publisher portfolios. Gaming sites with significant session times see particularly strong results - auction aware pricing on long-session inventory shows sustained mid-to-high single-digit RPM lifts at the network level, with individual high-engagement properties reaching double-digit improvements. The impact concentrates where variance is highest: sites with diverse geographic traffic, behavioural variance between casual browsers and deep-engagement users, and traffic volatility like live score sites with drastic surges. These results sustain over 6+ month validation periods through seasonal fluctuations and competitive dynamics.

However, ML-driven yield optimisation isn't magic, and it's not right for every publisher. Dynamic price floors reduce fill rates by design - typically 3-6 percentage points - though net RPM still increases because you're capturing significantly higher CPMs on impressions that do fill. Publishers serving fewer than 1 million impressions per month per site typically don't have enough volume for models to outperform well-tuned manual rules; the sweet spot starts around 5-10 million impressions per month per property.

There's also a 2-4 week cold-start period where models collect baseline data and calibrate through A/B testing. During this ramp, performance may be flat or even slightly negative. When significant market shifts occur - like the bid rate and CPM drops we observed across several publisher portfolios in January 2026 - models need recalibration, and there's a 1-2 week lag between market movement and full model adaptation. And critically, ML optimises execution but doesn't set strategy. Publishers still make decisions about demand sources, formats, and user experience tradeoffs.

Why Independence Matters

Publishers have three options for ML-driven optimisation: build it themselves, rely on exchanges, or partner with an independent provider. Most publishers can't justify building their own infrastructure unless they're operating at massive scale (5B+ impressions/month minimum). Exchanges offer optimisation, but it's optimised for their objectives, not yours, with no transparency into how decisions are made and no control over your data.

Independent providers offer the infrastructure without the build cost while ensuring publishers maintain strategic control and data ownership. At Mile, we handle the infrastructure complexity—data logging, feature engineering, model training, real-time inference, continuous validation, monitoring—while publishers set goals, monitor performance, and own their data.

What Publishers Should Evaluate

When assessing ML optimisation solutions, focus on infrastructure depth (do they have impression-level data collection or work from aggregated reports?), validation rigour (always-on A/B tests with proper control groups versus before/after comparisons), capability breadth (just floors or the full pre-auction decision surface including RTD, timeout, and shaping), independence (aligned with your revenue goals or someone else's objectives), and transparency (can you see exactly how decisions are made, or is it a black box?).

The buy-side figured out impression-level ML optimisation years ago. It's time the sell-side caught up—with infrastructure that actually works, and partners who are actually neutral.

About the Author

Vijay Kumar is Founder & CEO of Mile, an independent AI-powered yield optimisation platform serving enterprise publishers. Mile is a Google Certified Publishing Partner and a member of Prebid.org at the Technology Tier and IAB EU.