advertising-measurement-brand-safety-method

# Advertising Measurement and Brand Safety Method ## Operating posture Use this method only within written campaign and data authority. Freeze advertiser, brand, campaign, line-item, creative, placement, property, collection, content, geography, audience, platform, time zone, reporting cutoff, currency, metric contract, suitability policy version, experiment protocol, and reviewer identities before comparison. Preserve the distinction between declared configuration, platform-reported events, third-party measurement, finance records, and analyst calculations. A dashboard is a view, not a source. The method adapts only the upstream concepts documented in references/UPSTREAM.md and references/ADAPTATION.md. It performs no remote command. Never import vendor defaults or encode a time-sensitive benchmark. Operator-issued definitions govern; absent definitions remain unknown. ## Shared evidence envelope Every artifact and row carries a stable ID; source locator, producer, version, extraction method and date; cutoff/effective period and time zone; owner and qualified reviewer; applicability across campaign, geography, property and jurisdiction; currency, value, unit and denominator; transformation and join logic; assumptions; uncertainty/confidence with reason; privacy, consent, confidentiality and license state; status; decision_not_made; outcome_unknown; and stop/escalation. Retain raw-reported and derived values separately. Never repair a discrepancy by overwriting a source. ## Campaign taxonomy and metric contract Build a canonical identity map from authorized source snapshots. Map campaign, line item, insertion order, creative, placement, property, channel, device, geography, audience and objective identifiers without assuming names are unique. Record effective dates and many-to-one or one-to-many mappings. Freeze each metric as a contract: event name, eligibility rules, numerator, denominator, unit, time basis, attribution window if operator supplied, filtering, deduplication, currency conversion, late-arrival policy, invalid-traffic treatment, and source owner. Reject comparison when definitions differ materially or denominators are absent. Trace events from collection through transformation, aggregation and report. Preserve schema version, consent and privacy constraints, bot or invalid-event flags, clock/time-zone handling, identity resolution, sampling, modeled fields, backfills and restatements. A metric label alone is not comparable evidence. ## Delivery reconciliation and data quality Reconcile platform, verification, publisher/ad-server, analytics and finance sources at the narrowest shared grain. Begin with population and coverage: expected campaign entities, observed entities, missing mappings, duplicate keys and excluded traffic. Then compare eligible impressions, rendered or measured events, viewability measures, clicks, conversions, spend and fees only under compatible contracts. Separate arithmetic variance from definition, timing, currency, sampling, attribution and coverage variance. For invalid traffic, fraud, viewability and quality signals, record provider, method/version, measurable population, flags, exclusions and appeal or review state. Do not declare traffic invalid, fraudulent or payable from a score alone. Route contractual payment and fraud determinations to authorized reviewers. Preserve outage, tag, consent, integration and late-data evidence. ## Brand safety and suitability verification Bind every policy assertion to the operator-approved brand, market, channel and effective policy version. Distinguish legal prohibition, platform safety category, advertiser suitability preference, contextual classification, inclusion/exclusion list, placement observation and incident allegation. Capture exact content or placement locator where authorized, observation time, classifier/provider version, human-review state, screenshot or archive provenance, and counterevidence. Never infer page-wide context from a fragment. Assess coverage of the declared property-collection-content hierarchy and placement inventory. Reconcile pre-bid configuration, post-bid verification, publisher declaration and observed delivery without changing any control. Do not certify brand safety, choose blocking thresholds, add domains to lists, remove content, contact publishers, or declare policy compliance. Escalate urgent exposure through the operator's approved channel without external action. ## Attribution, incrementality, and experiments Freeze objective, unit of analysis, assignment method, treatment, control, eligibility, exposure, outcome, analysis population, time windows, pre-period, exclusions, missing-data handling, model specification and decision rules supplied by the experiment owner. Validate randomization or quasi-experimental assumptions with source evidence; inspect sample-ratio mismatch, interference, contamination, attrition, seasonality, concurrent campaigns, identity loss and post-treatment filtering. Preserve intention-to-treat and other prespecified populations separately. Attribution is not incrementality. Record whether a result is deterministic, probabilistic, modeled, observational or experimental. Recompute only authorized formulas, retain uncertainty intervals and practical units, and expose multiplicity or repeated-look risks. Never invent an attribution window, causal claim, confidence threshold, minimum effect or stopping rule. Qualified experiment and measurement owners determine interpretation. ## Four-root reconciliation The scheduler launches four zero-dependency roots with the identical baseline. The brand-safety root verifies policy identity, placement/content observations and counterevidence. The taxonomy root freezes identities, definitions, denominators and lineage. The delivery root reconciles populations, sources and variances. The outcome root evaluates attribution and experiment design. Each returns immutable asset locators and open questions. The join runs only after all roots complete. It reconciles stable IDs, campaign scope, timestamps, time zones, currencies, populations, definitions and source versions. Conflicting values or classifications remain parallel with provenance; the join never votes, averages incompatible measures, or silently selects a vendor. Trace every proposed claim to metric contract, delivery coverage, suitability evidence, experiment status and named reviewer. ## Stops and authority Stop for missing campaign or data authority, unpinned snapshots, absent metric denominator, ambiguous time zone/currency, incompatible populations, unclear consent/privacy/license, missing suitability policy version, undocumented attribution rules, compromised experiment assignment, or requested live access. Mark outcome unknown when evidence cannot resolve a claim. No worker buys media; changes platform state, budget, bid, pacing, creative, targeting, audience, list, placement or publication; declares fraud, legality, compliance, suitability, payment, performance, or causal truth; approves a campaign or claim; or accepts risk. Media owners, brand-safety specialists, measurement scientists, experiment owners, finance, privacy and legal reviewers retain those decisions.