← Work · Case 04
ProductionEvidence-first SEO · Audit suite

Evidence-first agentic audits

Agents audit a site end to end without writing a single byte to it.

StatusProduction
TypeAudit suite
RoleConcept, design and build, with AI agents

01 · Problem

Problem

AI-written SEO audits look convincing and are often wrong: invented metrics, confident guesses, recommendations nobody can trace back to data.

02 · Approach

Approach

An audit suite orchestrated by agents, with one entry point. A pre-flight check verifies runtime, skills and sources; evidence from every source is reconciled; the report is generated deterministically; a blocking QA stops anything that fails.

03 · Architecture

Architecture

crawl / searchdata / pagesourceSOURCEpre-flightcheckSTEP 01reconcileevidenceSTEP 02deterministicreportSTEP 03blocking QASTEP 04human reviewHUMAN GATEreportOUTPUTSOURCE → PROPOSAL │ HUMAN GATE │ INSPECTABLE OUTPUTcrawl / search data / pagesourceSOURCEpre-flight checkSTEP 01reconcile evidenceSTEP 02deterministic reportSTEP 03blocking QASTEP 04human reviewHUMAN GATEreportOUTPUT
Sanitized architecture · roles and decisions, never real data.

04 · Guarantees

Guarantees

  1. 01Zero writes to live sites.
  2. 02Every finding links to its evidence.
  3. 03A number that is not in a source is not written.
  4. 04A failed QA blocks the report instead of shipping it with warnings.
  5. 05No audit ever triggers content rewriting on its own.

05 · Stack & decisions

Stack & decisions

PythonThe deterministic engine: parsing, scoring and rendering.
Agent skillsVersioned contracts, identical for every agent that runs them.
Single wrapperOne entry point, so there is no second, unchecked way to start an audit.
Hash checksThe skill copies used by different agents must match byte for byte.

06 · Proof

Proof

Audit suite · pre-flightchecklist · authentic, sanitized
  1. one entry point. no second, unchecked way to start an audit.
  2. pre-flight, each check separate
  3. athe runtime
  4. bskill version and scripts, for every agent
  5. cthe skill copies used by different agents are byte-identical
  6. dbroker status: informative, not blocking
  7. live writes: disabled

Excerpt from the real system, translated and sanitized: no names, identifiers or data.

Next case · 05An evidence-gated research digest