Static floor prices are costing publishers an average of 30 to 40 percent of potential revenue. After analysing 1.4 billion daily auctions across our publisher network, the data is unambiguous: the era of one-size-fits-all floor pricing is over, and the publishers who adapt first will capture enormous competitive advantage.
Most publishers set their CPM floors once, maybe twice a year. They pick a number based on historical averages, gut instinct, or what their ad ops team negotiated with their primary SSP. This approach made sense in 2018. In 2026, it is actively destroying revenue.
The programmatic ecosystem has changed fundamentally. Demand fluctuates by hour, by geography, by content vertical, by user segment, and by device type. A floor price set to capture average demand will consistently leave money on the table during peak demand periods and block legitimate low-CPM fill that would otherwise monetise unsold inventory.
Our analysis shows that for a typical mid-market publisher with 50 million monthly impressions, static floors result in three compounding losses:
Publishers using static floors are optimising for the average. But programmatic advertising does not reward average. It rewards precision.
A machine learning floor pricing system replaces a single static value with a continuous prediction engine. Instead of asking "what should our floor be?" the system asks "for this exact impression, at this moment, with this user context, what is the minimum price we should accept?"
The model ingests dozens of signals in real time: historical clearing prices for similar impressions, current demand pressure from active DSPs, time-of-day curves, seasonal patterns, content category premiums, geography-specific demand indexes, and advertiser category activity. It produces a per-impression floor recommendation within the bid evaluation window, typically under 8 milliseconds.
The output is not a single floor but a distribution: a conservative floor to maximise fill, a target floor to optimise eCPM, and an aggressive floor to capture peak demand. The system selects dynamically based on current inventory pressure and publisher-defined revenue objectives.
We ran a controlled 90-day trial across 42 publishers, splitting inventory 50/50 between static floors and ML dynamic floors. The results were consistent and decisive.
Publishers on dynamic floors saw eCPM increases of 28 to 41 percent, with the median at 34 percent. Fill rates remained stable, with a slight positive lift of 2 to 3 percent as the system more accurately identified impressions where accepting lower floors made sense for total revenue optimisation.
The most dramatic gains came in two areas. First, during high-demand windows (weekday mornings, Sunday evenings, major news cycles) where static floors consistently under-captured demand. The ML model recognised these periods and raised floors accordingly, capturing clearing prices 60 to 80 percent higher than the static baseline. Second, in premium content categories where brand safety signals drove advertiser premiums that static floors were not capturing.
Moving to ML-driven floor pricing does not require replacing your entire ad tech stack. The approach we recommend follows three phases:
Before training any model, you need clean auction-level data. This means logging every bid request alongside the clearing price, winning DSP, and key impression attributes. Many publishers are surprised to find they do not have this data in a usable form. Start here before anything else.
With 30 or more days of auction data, you can begin training. We recommend starting with a gradient boosting model (XGBoost performs well on tabular auction data) and validating against held-out historical auctions before going live. The goal in this phase is not a perfect model but a baseline that outperforms static floors on historical data.
Deploy to 10 percent of inventory first. Monitor eCPM, fill rate, and unfilled impressions against your static floor control group. Expand to 25 percent, then 50 percent, then full rollout as the model stabilises. The ramp period is also your retraining window: the model improves significantly with live feedback data.
The publishers who moved to dynamic ML floors in 2025 did not just earn more per impression. They built a data infrastructure that compounds in value as the model trains on more auctions.
Dynamic floors introduce new failure modes that static floors do not have. The most common issue is model drift during unusual demand events (major holidays, global news moments) where the model lacks training examples. Build in a fallback to static floors with manual override capability for these windows.
The second common failure is over-optimising for eCPM at the expense of buyer relationships. If your dynamic floors consistently block DSPs that bid below your targets, those DSPs reduce bid density on your inventory over time, which undermines the long-term demand pool. Calibrate aggressiveness carefully, especially with key demand partners.
Static floor pricing is a 2018 solution to a 2026 problem. Publishers operating at scale cannot afford to leave 30 to 40 percent of revenue on the table because their pricing infrastructure has not evolved. The technology to fix this exists today. The question is execution speed.
Publishers who adopt ML-driven floors in 2026 will have 12 to 18 months of model training advantage over those who wait. In programmatic advertising, that data compounding effect is a durable moat.
Monthly analysis on programmatic trends, publisher strategies, and advertiser playbooks. No fluff. Just data and insight from the team building at the frontier of adtech.
No spam. Unsubscribe anytime. Monthly cadence.