AI Campaign Performance Report Verification Checklist

Freeze the exact report and evidence packet

Assign a report version, owner, review date, campaign boundary, and intended decision. Preserve the published draft, metric dictionary, source exports, query or transformation logic, attribution settings, lookback windows, conversion definitions, currency and time-zone rules, refund treatment, maturity schedule, experiment protocol where one exists, and the AI-assisted working record used for material analysis.

Record stable file names or hashes so the reviewer can tell whether a source or report changed after review. If a dashboard cannot be frozen, export the relevant rows and capture the filters, property, reporting scope, retrieval time, and user role. Issue a new report version when a material field changes.

Google Analytics documents that attribution settings include the reporting model, channels eligible for credit, and lookback windows, and that setting changes can affect reported data. See Select attribution settings. This is first-party documentation for one platform; it does not define every source system or validate an organization’s measurement design.

Before continuing, answer yes or no and link the evidence:

• Is the decision, reporting period, campaign population, and comparison explicit?
• Are raw, modeled, attributed, and derived fields distinguishable?
• Can a reviewer retrieve the exact source rows behind every headline number?
• Are source delays, missing records, duplicate rules, and known definition changes visible?
• Does the packet identify who approved transformations that alter totals?

Sources: [1]

Recalculate totals from the source rows

Reproduce spend, delivery, journey, and outcome totals independently of the report formulas. Confirm units, currencies, tax treatment, refunds, cancellations, excluded records, time boundaries, deduplication keys, and population denominators. Trace every rate to a numerator and denominator. Recompute weighted averages from their components rather than averaging displayed percentages.

Build a reconciliation table with report metric, report value, reproduced value, variance, tolerance, likely mechanism, owner, and disposition. Do not “correct” a mismatch by overwriting one source with another. A platform’s attributed conversion total and a billing system’s paid-order total answer different questions and may be valid within their stated scope.

Fail the recalculation check when a headline number cannot be reproduced, a formula mixes incompatible populations, a material variance lacks an owner, or an AI-generated explanation is treated as evidence. Mark the report not assessable if the underlying extract or transformation is unavailable.

Test attribution and data maturity

Verify the attribution model, eligible channels, counting method, attribution timing, and window for each attributed metric. Confirm that the report does not add fractional or overlapping credit from multiple platforms as though it represented unique outcomes. Separate observed business outcomes from model-assigned channel credit.

Google Analytics states that modeled key-event attribution data can update for up to 12 days after a conversion is recorded. See About modeled key events. Other systems have their own finalization behavior. Use a source-specific maturity schedule and label every headline figure preliminary, mature, restated, or unavailable.

Where the report uses experiment evidence, verify assignment, exclusions, contamination, outcome definition, analysis window, and whether the reported population matches the protocol. Google Ads describes Conversion Lift as treatment-and-control measurement and distinguishes it from standard attributed conversions. See About Conversion Lift based on users. Without a suitable causal design, replace incremental language with an observed or attributed description.

Sources: [2] [3]

Challenge every material claim

Create a claim trace with exact wording, source cells, transformation, claim type, alternative explanation, confidence, owner, and disposition. Test descriptions of “best,” “efficient,” “lift,” “caused,” “optimized,” and “underperforming” against the actual evidence. A channel can have a lower attributed cost per conversion without causing more conversions or being the best use of the next unit of budget.

Ask AI to generate counter-explanations only after it has the frozen definitions and row identifiers. Require its output to cite the rows it used. Reject any statement that introduces an unsupported number, hides an adverse segment, ignores immature data, or converts missing evidence into a positive result.

Run at least three sensitivity checks: delayed outcome volume, higher finalized spend, and an alternative comparison period or audience mix. Record the switching point at which a proposed decision changes.

Verify the decision table and action controls

For every continue, change, stop, scale, or test recommendation, require the observed signal, competing explanation, operational constraint, decision owner, authorized boundary, next measurement, review date, and stop condition. Confirm that the proposed owner has the relevant authority and that legal, privacy, finance, brand, or platform decisions are routed separately where required.

Return a “scale” action if it lacks a unit, spending or exposure boundary, data-maturity condition, and reversal rule. Return a “test” action if it lacks a hypothesis, treatment, comparison, primary outcome, analysis window, and contamination check. Verify that every action is feasible with the systems and people named in the report.

Use four dispositions: pass, conditional pass, repair, and not assessable. Conditional pass requires enforceable conditions before action. Repair requires a new report version and retest. Not assessable means missing sources, incompatible definitions, or absent authority prevent a responsible decision.

Worked verification: a report that double-counts outcomes

A synthetic launch report covers search, social video, and a partner newsletter. It claims 234 conversions by adding 84 search-attributed upgrades, 119 social-attributed upgrades, and 31 partner-attributed upgrades. The billing export contains 171 mature paid upgrades. The draft also says social “drove the highest incremental return.”

The reviewer freezes the packet and finds that search uses a 30-day click window, social includes seven-day view-through credit, and the partner system uses last-touch referral credit. Recalculation confirms each platform total within its own rules, but the values overlap and cannot be summed. No lift study or comparable holdout exists. Seven recent days are still inside the billing-finalization window.

The disposition is repair. The revised report uses 171 mature paid upgrades as the business-system outcome, shows channel attribution in separate non-additive columns, marks seven days preliminary, and changes the incremental-return claim to “not measured.” The action table continues search within its present boundary, caps social until mature data arrives, and specifies a geo-balanced holdout for the next campaign. The retained artifact contains the frozen-file manifest, reconciliation table, claim trace, maturity schedule, sensitivity results, action tests, exceptions, and signed disposition.

Preserve the assessment boundary

Two Prune can place a participant in a realistic report-review simulation and observe whether they reproduce calculations, distinguish attribution from incrementality, challenge AI-assisted claims, and communicate a bounded disposition. That is task-specific assessment evidence. It is not marketing analytics software, an audit opinion, proof of campaign impact, regulatory approval, or an automated employment decision.