Claude Ads review Check the evidence before discussing optimization

A pinned local review of CSV normalization, invalid inputs, unknown findings and HTML reports, explaining disabled health scores and the limits of synthetic testing.

AgentIndex · Published 2026-11-04 · Updated 2026-11-04

Claude Ads is relevant to teams that already have advertising data and need a consistent way to organize findings and supporting evidence. It connects exports, missing inputs and reports through inspectable artifacts. Before evaluating the usefulness of generated recommendations, we would check whether those artifacts faithfully represent the supplied data.

The most important finding is that this revision does not enable account-health scoring. All twelve platform profiles are disabled. With insufficient evidence, the actual result was health_score: null and insufficient_evidence. This is not a tool that our test established can turn any export into a reliable account score. An absent score also does not mean an account is performing badly.

Version and local scope

On November 4, 2026, we fixed commit fc3908e786c318776df89a3a098333da377364be and ran the deterministic core with Python 3.12.13. The core declares version 2.0.0, but the commit should accompany that number when reproducing this review. The README describes the repository as a development candidate. Pinned project description

Inputs consisted of our handwritten two-day CSV and the project's local synthetic report demonstration. We did not install a skill into our everyday host, connect an advertising account, invoke a model or deliver advertisements. Our reproduction fixture records the selected source fingerprints and outcomes.

CheckObserved result
Two synthetic daily observationsSpend normalized to USD 20; date window matched
Repeated date/campaign/creative grainRejected
Mixed currencies in one exportRejected
Negative spendRejected
Unknown finding without evidenceStructurally valid
Finding status outside the vocabularyRejected
Twelve health-scoring profilesDisabled; no health score invented
Invented control identifierRejected by the registry
Synthetic local reportRendered to HTML
Report score altered to 100Rejected before the output file was written

An export adapter is a particular data contract

The tested adapter expects a generic CSV shape. Platform exports must first be mapped to its required columns. Each row represents an additive observation at the date, campaign and creative grain, with spend and conversions already allocated appropriately. The duplicate-row test verifies a refusal path, not automatic compatibility with every platform's original download format. CSV adapter implementation

This matters when several people supply reports. Two columns called spend may represent different accounts, timezones or aggregation levels. Repeating a campaign total on every creative row produces an incorrect sum even when all values parse successfully. Define the grain, reporting period and currency before importing, then reconcile normalized totals with the original data.

Our synthetic checks do not establish attribution-window consistency, cross-platform deduplication or completeness of real conversion data. Passing a platform name to a generic adapter does not establish that its live API integration was tested. Treat input preparation as a visible part of the workflow instead of an incidental formatting step.

Preserve unknown results

Findings use a constrained status vocabulary. Unknown permits the report to say that evidence is missing rather than manufacture a complete-looking answer. Our missing-evidence finding passed structural validation; an unsupported status did not. Finding schema

Structural validity is still narrower than credibility. A recognized control must be tied to the relevant account, reporting window and evidence. Fluent explanatory text does not make that connection automatically. Our invented identifier was rejected by the registry, demonstrating that the tested scoring path checks control identity against its canonical definitions.

When reading a report, ask which input supports a finding, whether that input covers the stated scope, and what additional information would make an unknown decidable. A longer recommendation does not repair a missing answer to those questions. This distinction is useful even when the eventual review is performed by a person rather than a model.

Rendering does not authorize arbitrary scores

The project uses versioned JSON as the basis for report rendering. We created its local demonstration and used the policy directory explicitly returned by that demonstration to render HTML. Altering the report's health score then caused rejection before a file appeared at the requested destination. Report policy validation

This supports a limited practical conclusion: the tested renderer does not blindly accept a caller-supplied score. It does not establish that a model interprets every input correctly or that every possible run state is reported adequately. A successfully rendered document and a correct advertising audit need separate acceptance criteria.

The demo also requires a fully resolved filesystem path. On macOS, the symlink form of a temporary directory was refused; resolving that directory allowed the workflow to complete. Record that condition when reproducing the test rather than interpreting the initial path error as an audit-quality result.

Verify the relevant capability

The README lists twelve platforms, while live reading and account-change capabilities are declared separately. Its status section describes changes as drafts until the complete mutation gate passes. We did not exercise any real-account change and do not report declared platform coverage as verified integration coverage. Capability manifest

For an initial trial, use synthetic or appropriately shareable exports. Check normalized totals, unknown findings and report versions, then test skill discovery and activation in your chosen host. Browser capture and PDF output have additional runtime dependencies; our HTML result says nothing about those paths.

Adoption decision

If the problem is scattered source material, unsupported findings or reports that are difficult to revisit, the structure is worth a small trial. A useful deliverable includes the original reporting scope, normalized data, identified controls and unresolved unknowns. The final optimization paragraph alone is too little evidence to assess the workflow.

The pinned revision uses MIT licensing and remains a development candidate. Our results establish no advertising return, automatic optimization benefit or replacement for testing an actual platform integration. License

For building an application around operational data, see our NocoBase 3 project review. Development teams looking for a repeatable diagnostic evidence trail can continue with the .NET Skills MSBuild review.