How We Audit: Which Exam to Run, What Pass Means

Show buyers proof they can trust. Use this audit guide to help your team find your best proof, spot gaps, and plan the work. Our Content Factory turns that proof into useful posts.

An audit should tell you what a stranger can verify, what matters to your buyer, and what to do next. A long report or a search score alone does not answer those questions. Start with the business and its evidence. Then select the right exam, if a score helps the decision.

This is the maintained public method for our audits. The agent procedure lives in the Local Service Spotlight skill repository; use the shared audit run contract with its specialist skills. Examples link here. They do not become new standards.

Give value before asking for more

For a premium audit or a conference, prepare the starting point before the meeting. Identify the person and company, inspect their existing proof, show their strongest assets and gaps, and give useful first actions. Resolve namesakes before assigning evidence to someone.

Goals, Content, Targeting means the result you want, the proof you can use, and the people you need to reach. We call it GCT. If that information is already confirmed, proceed. Do not ask the attendee to repeat it.

If public evidence suggests a direction, label it INFERRED—confirm or edit. Offer Confirm, Edit and Continue with base audit. If two plausible answers would produce different priorities, ask the relevant question and explain what it changes.

• Goal: What would make the next 90 days a win for your business? This helps us put the right opportunities first.
• Targeting: Whose trust do you most need, and what should they buy or do? Different buyers need different proof.
• Content: Which one to three results, interviews, reviews, talks or videos should those people see? Search can miss your best work.

Show only the questions needed, at most three in the first pass. A link or short voice/text reply is enough. Keep “Continue with base audit” visible. “Not sure yet,” a skipped question or no reply still produces qualified findings and conditional priorities. Missing GCT does not mean low reputation. Do not require a new login or analytics access for the base audit.

Without a confirmed goal, the proof findings remain useful. The sales destination and affected recommendations stay provisional. On reply, rerank the relevant proof, update the gap and actions, and explain what changed. Preserve the original answers alongside the agent’s hypothesis.

Run the audit in this order

1. Name the business question and freeze the inputs

Record the identity, company, offer, buyer, market, time window and decision sought. Mark each strategy field supplied, inferred or unknown. Where authorized records exist, begin with calls, leads, bookings, CRM and revenue. Public followers and visibility do not establish sales.

Pin the evidence packet and report contract: exam family, rubric ID and version, source revision, rows and weights, arithmetic, missing-data policy, pass line, template, length, audience and permissions. A report is reproducible only when another operator can find those same inputs.

2. Build one editable proof inventory

List the recordings, appearances, articles, reviews, results, talks, relationships, offers and useful pages that actually exist. Give each item an evidence ID, source, event date, capture time, subject match, permission and a short statement of what it establishes. Deduplicate repeated coverage of the same moment.

Use three labels: OBSERVED for what the source shows, INFERRED for an interpretation with stated support, and UNKNOWN for what was not established. A website claiming a result proves that the claim is published; it does not automatically corroborate the result.

The proof-harvesting method ranks individual items on Who, Where and What, a 30-point scale. Explain why each item matters to the buyer. That item score is not the business’s headline audit score. A shared photo or interview also does not establish friendship, endorsement or permission to seek an introduction.

3. Map the reputation and sales gap

Compare what the buyer needs to believe with the proof available and what they can find. Separate missing proof, valuable proof that is buried, inconsistent identity, weak positioning, a broken buying path and missing measurement. Keep strengths and counterevidence in the report.

Use a Money Tree to connect useful content to the offer and buying step. Link real destinations and expose holes. Mark a proposed path as proposed. A link from an article to a service is not proof of a sale.

4. Select and calculate the right exam

Use the approved rubric for the question being answered. Keep distinct exams separate. A seven-category report and a ten-row report can be different exams; their difference alone does not show version drift. The exam table below gives the boundaries.

Calculate the headline from the rows in code. For weighted categories, use the declared weights and scale. Display rounding must not change the pass decision. Keep a documented cap separate from the raw row total.

5. Make a visual report someone can act on

The first screen shows the business question, strongest supported finding and useful visual. Include strengths, gaps, relevant proof, a readable timeline, citations beside material claims and open decisions. Use real permitted media and diagrams that explain the evidence. Choose the length for the engagement and agreed template; do not pad pages.

Rank actions by plausible business impact, effort and dependencies. Each action needs a finding and evidence link, an owner, the skill or process to run, required inputs, destination, due date, authority needed and a visible acceptance test. Show which work an agent can do and which decision needs a person.

6. Implement and review what changed

Move accepted actions into the existing owner’s workflow. Use Produce → Process → Post → Promote to reuse the strongest material. Consider Dollar-a-Day promotion for proven organic winners within an authorized budget.

Review Metrics → Analysis → Action (MAA) weekly: what changed, why it matters and what happens next. Measure business outcomes separately from the audit score. Reuse the existing task and review schedule. A recommendation does not itself authorize spending, access changes or public release.

Several exams; no blended score

The maintained personal-brand-audit skill orchestrates evidence, specialist findings, a visual report and owned actions. Its current product has an exact 20-page blueprint. Historical 15-page and adaptive 15/20-page instructions remain historical contracts. A comparison’s agreed length is a control for that comparison.

UNKNOWN is not zero

If you could not check a source, say what is unknown and why. A checked absence may earn zero under the selected rubric; an unperformed or blocked check cannot. For a rubric requiring every row, show the known subtotal and coverage, mark the result INCOMPLETE, and withhold the total and pass verdict. Do not renormalize missing rows into a full score. Use N/A or partial-score denominators only when the pinned rubric explicitly defines them.

Keep re-audit comparisons within the same exam, version and comparable coverage. New evidence, corrected arithmetic and a changed rubric are different changes. None alone proves a business gain.

A graph object, visible panel and owner claim are different

• Graph identity: record the exact graph identifier and at least two distinct identity sources. Query-match strength is relevance to a query, not reputation.
• Normal-query panel: record the exact name query, locale/location, device, personalization and capture time, with a screenshot or equivalent receipt. A forced graph-identifier view is not ordinary visibility. Blocked access stays UNKNOWN.
• Owner claim: use an authorized owner-side receipt. A public “Claim this knowledge panel” control does not establish who owns it.

Google’s official claim process starts from the panel and uses the verification options offered there. Some panels cannot be claimed. Do not promise a fixed arrival time or that a schema change guarantees a panel.

Make a correction reach every final report

A corrected finding can still fail if one final PDF retains the old claim. We observed that failure in an earlier conference batch. The response is a shared source and an explicit dependency check:

1. Freeze the input packet and appoint one coordinator for shared facts and final integration.
2. Correct the authoritative fact, assign a new revision, and list every affected report, summary, HTML page and leaderboard.
3. Regenerate those outputs. Inspect the final extracted text and the actual rendered pages. Reject retired claims, old totals, stale versions or a missing dependent artifact.
4. Record exact final file hashes and separate coordinator and fresh-context review receipts. Page counts and a worker saying “QA passed” do not prove appearance or factual accuracy.

The maintained repository includes offline fixtures for row arithmetic, UNKNOWN, exact rubric versions, entity/panel separation and correction propagation. PDF checks bind freshly extracted text to the final PDF binary. Those checks validate the contract; people and agents still need to review sources and every rendered page.

What ran, what is documented, and what is still unknown

Report visible skill reads, tools and handoffs separately from instructions naming a skill or a report claiming it ran. Record the actual UI/provider model label and effort, or UNKNOWN, plus source times, run time, allowance limits, retries, failures and human rework. A historical routing guide does not prove which model executed a worker or expose hidden reasoning.

To compare Codex, Claude, Muse and Cursor, use the same verified identities, evidence, instructions, approved rubric and template. Evaluate evidence, insight, actionability, visuals, implementation ease and required human help separately from the person’s audit score. The result compares actual workflows with their recorded tools and context. It does not isolate model quality from those differences.

A same-model comparison with different instructions would be a separate agreed experiment. Do not silently add runs or assign a percentage of historical quality to instructions versus models. Use existing authorized subscriptions when that is the agreed allowance; a missing tool is a limitation, not permission to buy an API.

Use the maintained source and verify the destination

Install through Install Local Service Spotlight. The full lss-everything bundle contains the routed audit skills. Pin the repository commit and verify which skills are actually available in the named account. A source change, installation, activation, tested output and observed scheduled result are separate states.

Keep one procedure owner and link to it. Put client details, raw private dictation, permissions and operational receipts in authorized private records. Share the teaching method publicly. Publish through the destination’s freshly verified site-specific route with a preimage, rollback and ordinary-URL readback. dennisyu.com is static; old WordPress instructions for that host are retired.

Worked examples live on the SEO audit hub and link to this method. The definitive article guide explains that relationship. Keep a case study as evidence of a run, not a competing copy of the procedure.

Method edition: AUDIT-METHOD-2026.10.02, October 2, 2026. Updated in the existing How We Audit source. Specialist rubrics retain their own versions. Source tests and public-page checks do not certify every installed account or historical client report.