Our measurement approach keeps a baseline comparison and makes uncertainty visible. A complete unattended optimization loop remains in development.
Start with the revenue we can observe
The primary ledger records rendered Prebid impressions. Where Google Publisher Tag is present, it uses the rendered slot’s Prebid price bucket, with attribution to the relevant auction. Integrations without Google Publisher Tag need separate render confirmation before supporting the same revenue claim.
A bid response or an auction win alone is not a guarantee of a paid impression. Price buckets are rounded values used for ad-server targeting, so this ledger estimates rendered revenue rather than reproducing a final payment statement.
Ad-server reports, when available, can help calibrate the estimate. Installation does not depend on that access. Direct-sold, AdX, and other demand outside the observed Prebid path are outside the revenue claim.
Compare the combined setup with the base
The traffic design reserves 10% of exposed sessions for a holdback and 90% for the optimized pool. The holdback receives the publisher’s base configuration without BidTune’s changes. If the publisher or provider updates that base, it remains the point of comparison.
The overall outcome is rendered Prebid revenue per exposed session. Exposed sessions include those with no rendered revenue; excluding them would hide changes in fill or delivery.
When starting from per-impression CPM values, convert each to dollars by dividing by 1,000 before summing. Compare the two groups over the same reporting window.
A session is tracked with a site-scoped browser session identifier that is reused across page loads in that browser tab where session storage is available. It is not the same unit as a pageview or an impression.
Match the readout to the change
- Floors and timeouts: examine impression CPM together with fill, delivery, and the relevant latency tradeoffs.
- Refresh: examine revenue per exposed session or visible minute, alongside CPM and viewability. Extra impressions alone are not success.
- Bidder changes: examine that bidder’s requests, responses, contribution, and the effect on the wider auction.
These readouts help make individual decisions. The permanent holdback evaluates the combined service. Adding individual lift percentages together does not establish the overall gain.
Show what remains uncertain
A positive point estimate can still be consistent with no improvement or a loss. A result needs its time window, sample size, comparison group, and uncertainty interval. Small sites and narrow segments generally take longer to provide precise answers.
Our overall proof target is a positive trailing-30-day difference with a 95% interval that excludes zero, while the loop operates without human intervention. This is a target for validation, not a result we have already claimed.
Individual decisions can use expected value and bounded downside within traffic limits. They need not all wait for the same confidence threshold.
Keep revenue, fees, and examples distinct
Observed Prebid revenue is not total publisher profit. BidTune’s fee and any partner deductions must be considered when judging the financial benefit. The pricing page explains the usage meter and published rate.
Product demonstrations and hypothetical calculations are labeled where they appear. They show how to interpret the system; they are not customer results. See the JSONLint case study and an illustrative calculation.