Header-bidding floor coordination
Data Transfer can't see the header-bidding auction — it resolves in the browser. We're extending the engine to drive Prebid floors from their own data feed, aligned with the GAM layer.
Archon CortexEvery bid matters.
Archon Cortex reads millions of rows of your real bid data — including the bids you lost — and sets the optimal price floor for every ad unit, country and device. Twice a week, every week. Publishers see 15–25% uplift from floors alone.
Yield curve · one ad unitcurrent $14.00 → recommended $18.14
rCPM (shaded)rCPM (baseline 1PA)Fill rate
The problem
Nobody can hand-tune a thousand floors.
Price-floor optimization is one of the biggest levers in ad yield — and one of the hardest to pull. With hundreds or thousands of ad units, dozens of geographies, shifting seasonality and an economy that moves under your feet, the "right" floor changes constantly. No team can re-tune them a hundred times a year by hand. So most publishers set a floor once and leave it — or hand the problem to a black box.
And the market now punishes idle floors harder than ever: in Q2 2026 open-web supply shrank while eCPMs rose — up ~30% year over year in the UK. Premium publishers are deliberately cutting low-value bid requests to protect price. That repricing only reaches publishers whose floors move with it.
Ozone data, reported by Digiday
The insight
We ran the experiments. Fixed floors won.
Our test-driven culture keeps producing the same result: data-driven fixed floors outperform Target CPM, Optimized CPM and set-and-forget strategies. The catch has always been maintenance — an optimal fixed floor doesn't stay optimal. Keeping thousands of them current is an industrial problem, not a spreadsheet problem. Cortex is the industrial solution.
Data-driven fixed floors
Target CPM
Optimized CPM
Set-and-forget
Ranked by results across our floor experiments — relative order, not measured magnitudes; results vary by publisher.
How Cortex decides
The engine room.
Ingest the full auction record.
GAM Data Transfer files and BigQuery — won bids, lost bids, rejection reasons. Reports show you what sold; the landscape is in what didn't.
Reconstruct the demand curve.
Cortex fits yield curves per segment — the price-demand relationships that never appear in any GAM report.
Cluster the inventory.
Meaningful combinations across ad unit × country × operating system × browser × device category. Thousands of micro-markets, each priced on its own data.
Predict — then verify.
Every floor recommendation is backtested against your own history and validated against holdout controls, with seasonality deconfounded so a good week never masquerades as uplift.
Apply, twice a week.
New floors upload across the network every few days — no cap on ad units, no manual ceiling.
From the Cortex console
The evidence, not the slides.
The screens our analysts work in — one live ad unit, values as the engine computed them.
Bid price distribution7 modes: $4.32 · $8.70 · $14.43 · $36.68 · $44.55 · $117.77 · $195.23
BidRejectionReason breakdownone ad unit · one analysis window
(Winner)1,97036.7%
Outbid1,86234.7%
Floor1,50127.9%
Other390.7%
Optimized pricing10.0%
Adaptive segmentation worksheetone rule per decision-tree leaf · rules are non-overlapping — priority order doesn't matter
R0$18.14Aggressive · lean-up
OSiOS
CountriesUnited States
BrowsersChromeIn-appYandex
R2$14.00
OSiOS
CountriesTürkiyeUnited KingdomSaudi ArabiaAzerbaijanCanada
OS and browser values collapse to their families before tree splitting · segments too thin to price on their own data fall through to the ad-unit-wide default UPR · every rule ships with its auction count, combo coverage and measured uplift.
Proof
Measured against holdouts, not assumed.
15–25%
Average uplift from floor optimization alone.
0%
Of our floor experiments produce a measurable uplift — measured against holdouts.
0B+
Monthly ad requests APH manages — Cortex runs continuously across them.
Where Cortex goes next
The roadmap.
Lookalike analysis for EEA
Pricing intelligence where consent thins the data — floors informed by comparable markets when a segment's own signal runs shallow.
Ad-type-aware floors
Data Transfer files can tell a non-skippable from a skippable from a non-video impression. We're splitting floors across those three ad types — higher fill rates and audibility, less user discomfort, each type priced on its own demand.
Objections, answered
Asked by every publisher we onboard.
"We already use Target CPM / Optimized CPM."
So did our publishers. In our experiments, data-driven fixed floors beat both. And unlike a black box, every Cortex floor comes with the evidence that set it.
"What do you need from us?"
If you onboard our GAM 360 network via Manage Inventory — nothing. Your bid data accumulates in our Data Transfer files automatically. If you join via Manage Account, Data Transfer reporting needs to be enabled on your GAM 360; we handle the BigQuery side.
"How do we know it's working?"
Because we prove it before we scale it. Cortex starts with A/B experiments, and the uplift shows up in your own GAM 360 reporting — not in our slides.
"What if a floor is set too high?"
Cortex trains on the bids you lost, not just the ones you won — it can see demand it would price out. And because floors retune twice a week, a segment never drifts far before the data pulls it back.
"What does it cost?"
Cortex is part of monetizing with APH — revenue share on overall ad revenue, no setup fee, no software license.
See what your floors are costing you.
Cortex runs as part of an APH monetization partnership. Apply with your domain and stack, and we come back with what Cortex would do on your inventory — floors priced from your own bid data, proven in your own GAM 360 reporting.