Test an ad price floor without mistaking CPM for yield

State the mechanism and map rule overlap

A price floor is a participation boundary for specified auction demand. Raising it may exclude lower bids; lowering it may admit bids that were previously ineligible. Neither move guarantees higher total revenue. Start by naming the inventory, demand types, geographies, devices, formats, and time window to which the rule applies. Then export every pricing rule that could overlap the same request, including precedence and fallback behavior.

Platform terms matter. Google Ad Manager describes unified pricing rules for specified non-guaranteed demand and warns that an overly high floor can block demand. Its comparison can also use values defined within that platform rather than a publisher's cash ledger. Treat those details as system-specific inputs. The publisher's decision metric should remain tied to eligible opportunities and realized, reconcilable revenue.

Values are hypothetical; platform currency and rule semantics must be verified.

Source notes: Unified pricing rules

Randomize one eligible population and keep the unit stable

Choose the assignment unit before launch. Request-level assignment can expose one reader to several treatments; page-view or user-level assignment can reduce that mixing but may require identifiers and consent that the publisher does not have. Document the unit, exclusion rules, intended allocation, and actual allocation. Exclude guaranteed campaigns if the floor does not govern them, while still watching their delivery as a guardrail.

Use the platform's experiment feature when it can hold eligibility and allocation consistently, but archive the exact settings. A clean comparison keeps creative formats, auction timeout, refresh policy, consent handling, and major traffic acquisition stable. If another monetization change launches mid-test, annotate it and decide whether the comparison remains usable. Do not extend a weak test repeatedly until it produces a preferred answer.

Source notes: Create a manual experiment

Read the whole funnel with hypothetical arithmetic

Suppose each arm receives 50,000 eligible requests. Control fills 40,000 impressions and earns $160; variant fills 34,000 and earns $153. Control impression CPM is $4.00, while variant CPM is $4.50. The higher variant CPM looks attractive, but revenue per eligible request is $0.00306 versus $0.00320 for control. On this simplified evidence, the floor raised CPM while reducing yield per opportunity.

Add uncertainty and cash reality before deciding. Compare bid participation, fill, rendered impressions, measurable viewability, latency, discrepancies, and adjustments. Use net revenue when comparable platform fees and adjustments are available; otherwise label the amount precisely. Segment only for predeclared operational questions, such as a country where demand density differs. Small slices can produce dramatic percentages with little decision value.

All figures are hypothetical and omit uncertainty, fees, and later adjustments.

Make a reversible inventory decision

The decision memo should contain the original hypothesis, exact settings, actual sample, exclusions, metric definitions, result interval, guardrail changes, and recommendation. Choose among keep control, adopt variant for the tested scope, run a materially different test, or stop because the data cannot answer the question. A neutral result is useful when it prevents a broad, unsupported floor change.

Set rollback triggers before launch: direct-campaign underdelivery, a large fill loss, latency breach, reporting failure, or unexpected rule overlap. Recheck after any adoption because bidder mix and demand change. A floor is an auction control, not a permanent valuation of the audience, and a test result from one inventory slice should not be generalized to every page.

• Export overlapping rules and precedence.
• Freeze the eligible population and assignment unit.
• Compare revenue per eligible opportunity with CPM.
• Monitor delivery and reader-experience guardrails.
• Archive settings and a dated decision memo.

Source notes: Unified pricing rules · Create a manual experiment