What the milestone meant
Building a scored reference pool made it possible to place an ad’s model output alongside other scored creative. A percentile describes a position within a defined comparison set. It does not tell you the probability that the ad will succeed.
The original announcement covered several platforms and creative categories. Coverage of a category does not mean every market, placement or duration has enough comparable observations.
Three counts to keep separate
| Count | What it describes | Why the distinction matters |
|---|---|---|
| Total reference pool | All records in the collection | Includes records that may not match your creative |
| Eligible benchmark slice | Records used for this comparison | Determines the context of the displayed percentile |
| Validation cohort | Records used to evaluate a named model against an outcome | Cannot be inferred from the total pool size |
A large pool can still contain a thin slice. Check the platform, format, category, market and version used in the comparison. Read any fallback or missing-evidence label before presenting a percentile.
Public ad data has boundaries
Official ad libraries provide creative references and transparency information. The fields differ by platform and programme. Seeing an ad or a view count does not mean its spend, conversions or profitability are available.
Meta Ad Library and Google Ads Transparency Center are reference sources, not universal campaign-results databases. Keep observed public fields distinct from scores generated by SaliencyLab.
What changes when the pool grows?
More relevant records can improve comparison coverage. A changing pool may also move a percentile even when the creative and raw score are unchanged. Record the benchmark version or date with a result so the team can explain that difference.
Do not promise that every incoming ad immediately changes every report. Eligibility rules, refreshes and saved report versions determine what a particular client sees.
What to read today
Use the benchmark interpretation guide for the review conversation and the methodology for the model-specific validation record. Watch the demo to see how benchmark context fits alongside the creative and recommendation.