01
Observed failure
The available metric cannot support the decision being made.
Examples include aggregate traffic used to judge a URL intervention, page and query exports treated as joinable when they are not, controls chosen after outcomes are visible, pre/post comparisons without a stable baseline, or AI visibility claims without an attributable observation source.
Evidence: raw exports, extraction definitions, dimensions and filters, timestamps, missing fields, data lineage, deployment boundaries, and the decision the metric is expected to support.
02
Change
Build the measurement around the question.
NRLC defines the analytical unit, freezes raw inputs, separates baseline from observation, preregisters decision rules where intervention is involved, and preserves query, page, topic, device, country, or date dimensions only when they answer the question.
Unsupported measures remain explicitly unavailable rather than being reconstructed from adjacent totals.
03
Constraint
Measurement does not manufacture causality.
Search Console latency, sampling, privacy thresholds, changing demand, platform opacity, and incomplete external-AI attribution constrain the conclusion. A before/after change is not automatically an intervention effect, and a null or adverse result remains valid.
04
Verification
Recompute the result from frozen evidence.
Verification checks source hashes, extraction windows, cohort membership, exclusions, transformations, confidence intervals where applicable, and claim-to-observation links. Another reviewer should be able to reproduce the stated result or identify exactly where interpretation begins.
Pass condition: the analysis answers the preregistered question with inspectable inputs and limitations; it need not produce a favorable result.