Ask an ad ops team how their floors got set and you usually get a history lesson rather than an answer. Someone raised the floor on a mobile app segment during a strong quarter. A partner complained about a country, so that country got blocked. A spreadsheet went around, and three of its rows became rules. All of it is still running, and nobody has measured any of it since the day it was written.
That is the normal condition of yield management at an exchange, and it is not a competence problem. A hand-written floor or block is cheap to create and expensive to judge. Judging it honestly means knowing what would have happened if you had never written it, and the traffic that would have told you was suppressed by the rule itself. So rules accumulate, the market underneath them moves, and the exchange keeps clearing at prices set for conditions that expired months ago.
Set the goal, not the rulebook
Floxis inverts that arrangement. For each endpoint you pick one objective: net revenue, fill, or win volume. That is the whole instruction. From there the engine blocks, scores and floors every auction toward the goal you chose, and it keeps doing it as demand moves, without anyone reopening the spreadsheet.
The objective matters because those three goals pull against each other. Maximizing net revenue raises floors toward observed clearing prices and accepts the fill it costs. Maximizing fill eases floors where demand is thin. Maximizing win volume balances floors and margin toward the highest cleared transaction count. An optimization product that never asks which of the three you want is not optimizing anything in particular.
Three levers, and what each one actually touches
Segment blocking drops the segments whose demand economics do not justify sending the request. This is the same discipline as traffic shaping, applied at the segment level rather than the partner level.
Dynamic per-segment floors set a reserve for each segment from that segment’s own observed clearing prices, instead of one global number applied to inventory that has nothing in common. Your manual floors always win. The engine only ever raises above the minimum you set, so a commercial commitment you made to a publisher or a partner is never quietly undercut by an algorithm.
Adaptive margin self-tunes to rolling trade patterns within a minimum and maximum that you define. You own the bounds; the engine finds the best point between them.
Off, Shadow, Enforce
Every lever moves through three states, and it earns each one. Off does nothing. Shadow scores and records the decision the lever would have made but still sends everything, so you get the evidence without paying for the mistake. Enforce actually acts.
Underneath all three sits a live holdout. Requests hash deterministically into arms, and arm 0 is the holdout: the exact status quo, no engine blocking, no engine floor, no adaptive margin. Every other arm is read against arm 0, on the same traffic, in the same hours, under the same demand conditions. That is what makes the number trustworthy. Comparing this week to last week is not a measurement, it is a coincidence with a chart attached.
This is exactly where floor changes normally go wrong. Raising a floor reliably lifts your average CPM whether or not it made you money, because it removes the cheap clears from the average. We wrote about that trap at length in our guide to bid floors. A holdout is the only thing that separates a floor that earned revenue from a floor that only flattered the report.
The breaker, and the ledger
If a test arm’s bid rate falls below 80% of the holdout’s, Enforce reverts itself to Shadow automatically. The lever keeps measuring and stops acting. The worst case of an experiment that goes wrong is therefore a bounded period of reduced demand that ends by itself, rather than a slow bleed that somebody notices at the end of the month.
And every arm is visible in the log-level bid, win and drop ledger. You can see which arm a request landed in, what floor was applied, whether a bid came back, and why anything was dropped. Optimization you cannot audit is a black box you are being asked to trust with your margin.
What this is not
There is no predictive bid model here and no forecasting. Floxis does not claim to know what a buyer would have paid. Reserves come from clearing prices actually observed on your own traffic, and every change is either proven against the holdout or reverted. That is a narrower promise than the industry usually makes, and it is the reason the numbers hold up when you check them.
Any figures you see in a demo of the arms panel are illustrative, not client results. The point of the panel is the method, not the multiple.
Three questions for any vendor selling you optimization
Whoever you evaluate, including us, ask these. Is there a live holdout, and is it genuinely the untouched status quo rather than a modelled baseline? Does enforcement revert itself automatically when it starts costing demand, or does that depend on a human noticing? Can you audit the arms yourself at log level, request by request?
A vendor who can answer all three is selling you measurement. A vendor who cannot is selling you a setting.
If you run an exchange today and your floors are older than your last demand partner, that is the conversation worth having. Book a demo and we will walk through the engine on live traffic.