SEO Master

seo-master

Evidence first. Separate observed facts, inferences, and proposals. Never present technical eligibility as a promise of ranking, citation, recommendation, or traffic.

0. Quick start

Use the smallest workflow that answers the request:

Audit [site] for technical SEO + GEO. Read-only. Sample [N] URLs.
Fix on-page + truthful schema for [URL]. Preserve brand and analytics.
Measure AI visibility for [brand] across [platforms] with [query set].
Compare SEO vs GEO opportunities for [market, locale, device].

Before work, state:

• target host, locale, audience, site type, and conversion goal;
• read-only audit or authorized edits;
• access available: repository, browser, logs, GSC, GA4, rank data, AI tools;
• sample or full crawl, URL cap, and time window;
• house property: yes/no. Apply section 13 only when yes.

If required access is missing, continue with observable evidence and label the unobservable layer UNVERIFIED; do not fill gaps with estimates.

1. Scope & when-not-to

Do run when

• The site has a public discovery or acquisition goal.
• The user asks about rankings, citations, AI mentions, crawlability, content, schema, performance, links, analytics, or search reporting.
• A deployable path exists and the requested access is authorized.

Do NOT run (anti-patterns)

Scope questions

Ask only questions that materially alter execution:

1. Which canonical host, markets, languages, and priority conversions?
2. Audit only, recommendations, or code/content changes?
3. Which evidence sources are authorized and available?
4. Which AI surfaces matter to the audience: ChatGPT, Gemini, Google AI Overviews/AI Mode, Perplexity, Doubao, Claude, or another engine?
5. What must remain untouched?

Record unresolved choices as assumptions. Never let a missing optional data source block checks that can be performed safely.

2. Technical SEO audit

Audit in dependency order: reachability → renderability → crawlability → indexability → canonicalization → discovery → internal architecture.

2.1 Evidence sample

Include at minimum, when in scope:

• homepage;
• one URL per indexable template;
• priority conversion pages;
• one paginated, filtered, localized, and JavaScript-heavy URL where present;
• redirect, 404, soft-404 candidate, and canonical duplicate samples;
• URLs reported by GSC or logs, if access exists.

For every finding record URL, timestamp, method, observed value, expected value, evidence, confidence, impact, and affected scope. Do not extrapolate a sampled finding to the whole site without a crawl or corroborating data.

2.2 Reachability, rendering, and crawlability

• Record DNS/TLS failures, final GET status, redirect chain, latency, and host.
• Compare raw HTML with rendered DOM for canonical, robots meta, headings, links, primary copy, and structured data.
• Parse robots.txt for the tested user agent. A Disallow controls crawling; it is not a page-level noindex.
• Check WAF/CDN responses separately from robots policy. A spoofed user-agent request does not authenticate a vendor crawler.
• Test robots and sitemap files on every host/subdomain that serves sampled URLs, redirects to or from the canonical host, or appears in hreflang/sitemaps.
• Flag infinite spaces: parameters, faceted paths, calendars, session IDs, internal search, and duplicate sort orders.

2.3 Indexability

For each sampled URL classify:

INDEXABLE, EXCLUDED-AS-INTENDED, BLOCKED-BY-TECHNICAL-CONFLICT, or UNVERIFIED-IN-INDEX.

Check:

• final status is indexable content, not redirect/error/soft error;
• meta robots and X-Robots-Tag;
• whether robots permits the tested crawler to fetch page-level directives; if blocked, classify the directive's crawler visibility as UNVERIFIED;
• canonical target status, indexability, locale, and content equivalence;
• GSC URL Inspection/indexing reports when authorized;
• orphan status and internal-link discoverability.

Never claim “indexed” from a 200 response, sitemap entry, or site: query alone.

2.4 Canonical, host, redirects, and hreflang

• Treat owner intent, production bindings, redirects, canonicals, sitemaps, internal links, and analytics configuration as evidence. Resolve conflicts; do not assume one signal is universally authoritative.
• Require one-hop permanent redirects for retired canonical URLs where safe.
• Reject loops, chains, protocol/host oscillation, mass redirects to irrelevant pages, and query loss that changes meaning.
• Require self-referential canonical on canonical pages unless a documented cross-domain or variant strategy says otherwise.
• For localized equivalents, require reciprocal valid language-region codes, indexable canonical targets, and x-default only when a true fallback exists.

2.5 Sitemaps and architecture

• Parse XML; do not count matching lines.
• Validate sitemap-index recursion, XML syntax, HTTP status, canonical host, indexable URLs, lastmod truthfulness, and URL limits.
• Exclude redirects, errors, duplicates, blocked/noindex URLs, internal search, and noncanonical parameters.
• Confirm priority pages receive descriptive internal links from the topic, product, category, or locale hubs that contain them in the site architecture.
• Report crawl depth as observed distribution, not a universal pass/fail number.

2.6 Technical scorecard

Use statuses, not a pseudo-scientific score:

3. On-page SEO

Optimize the page for one primary intent and the jobs needed to satisfy it.

Format-vs-intent gate (first filter)

Classify primary intent, then reject format mismatch before copy polish.

Decision-intent pages fail this gate if they are listicles. Fix format first.

Comparison/alternative page skeleton

• H1 states the decision question (e.g. How is X better than Y?)
• Above-the-fold: direct answer + proof hooks
• Real-data band (benchmarks, pricing, limits — sourced)
• Evidence-dense differentiators; product video/screenshots where claims need demo
• Fair criteria table; disclose conflicts

1. Match title, visible heading, opening, sections, media, and CTA to intent.
2. Make titles and descriptions specific and nonduplicative; judge truncation from rendered SERPs when available, not a fixed character quota.
3. Use a clear page topic and logical heading hierarchy. Do not enforce “exactly one H1” as a ranking rule; flag confusing document structure.
4. Put the useful answer before biography, history, or promotional filler.
5. Add internal links where they advance the task, using descriptive anchors.
6. Give images meaningful alternatives when informative; use empty alt for decorative images.
7. Check visible copy against canonical, schema, Open Graph, and product facts.
8. Remove doorway duplication, hidden text, keyword stuffing, and templated filler.

Output: KEEP, REWRITE, MERGE, REDIRECT, NOINDEX, or DELETE for each page, with evidence and destination where applicable.

4. Content & E-E-A-T

E-E-A-T is a qualitative quality lens, not a published Google score.

4.1 Evidence gate

For consequential claims, require:

• a named source, first-party evidence, or clearly described methodology;
• author/reviewer identity and demonstrated competence for the claim where an incorrect answer could affect health, safety, rights, or material decisions;
• publication and material-update dates when freshness matters;
• ownership, editorial, contact, correction, privacy, and commercial disclosure required by the site's function, jurisdiction, risk, or commercial incentives;
• claims that match the cited source and do not exceed it.

For YMYL topics, escalate unsupported medical, legal, financial, or safety advice. Never create credentials, reviews, test results, customers, or outcomes.

4.2 CORE-EEAT review

【Proposal】 Use this internal rubric for diagnosis only:

Assign PASS, WEAK, FAIL, or UNVERIFIED per dimension. Do not average the labels into a claimed Google score.

4.3 Semantic gap and refresh

• Build an entity/attribute/question map from the query set, SERP, customer language, and first-party data.
• Add missing concepts only when they help the task; avoid synonym padding.
• Preserve URLs with earned value unless evidence favors merge/redirect.
• Refresh changed facts and examples; do not change dates without material work.
• Prefer one useful new fact, example, test, or decision aid over generic length.

5. Keyword and demand research

Treat keywords as evidence of demand and language, not insertion targets.

Intent-first table

Workflow

1. Seed from products, problems, customer interviews, site search, sales/support, GSC, paid-search terms, and competitor/category language.
2. Expand variants by task, entity, comparison, alternative, constraint, locale, and funnel stage.
3. Inspect current SERPs by locale/device/date; classify dominant intent and result type.
4. Cluster by shared intent and answer requirements, not an arbitrary overlap threshold.
5. Map one canonical page or page family per cluster; identify cannibalization.
6. Prioritize by evidenced demand, conversion fit, strategic value, feasibility, authority gap, and maintenance cost.
7. Label volume, difficulty, and trend data with provider, geography, and date.

【Proposal】 If comparable volume data is unavailable, use ordinal H/M/L inputs with written reasons. Do not invent numeric opportunity scores.

6. Structured data / JSON-LD

Structured data describes visible truth; it does not guarantee rankings, rich results, or AI citations.

Type selection

• Use the most specific Schema.org type supported by page content.
• Organization/LocalBusiness: real identity, sameAs, contact, location.
• WebSite: site identity; add actions only when the action truly works.
• Article/NewsArticle: author, dates, headline, image, publisher.
• Product/SoftwareApplication: real product attributes; include Offer only with current price, currency, availability, and URL.
• BreadcrumbList: visible navigational hierarchy.
• VideoObject, Event, JobPosting, Recipe, or other types only when page content and current provider eligibility support them.
• FAQ or How-to markup may describe visible content, but never promise a Google rich result or AI extraction; check current provider documentation first.

Validation

1. Extract every JSON-LD block from raw and rendered HTML.
2. Parse as JSON; resolve duplicate/conflicting entities and stable @id links.
3. Match every claim to visible content and current facts.
4. Validate vocabulary with Schema.org tooling.
5. Validate provider eligibility with the provider's current rich-result docs and test when that provider supports the page's structured-data type.
6. Save errors/warnings with URL and timestamp. Never “validate mentally.”
7. No fake price: 0 on paid products; omit Offer if price unknown.

7. Performance / Core Web Vitals

Use field data at the 75th percentile when available; use lab data to diagnose. Google's current “good” thresholds are LCP ≤2.5 s, INP ≤200 ms, and CLS ≤0.1 (official reference). Re-check the reference at runtime.

Record:

• source: CrUX, PageSpeed Insights, GSC CWV, RUM, or lab;
• URL-level or origin-level aggregation;
• device, locale, connection/CPU profile, sample window, and collection date;
• field status and lab reproduction separately.

Prioritize the measured bottleneck:

• LCP: server response, critical resource discovery, image/font priority.
• INP: long tasks, main-thread work, hydration, event handlers.
• CLS: unsized media/ads/embeds, injected content, font swaps.

Do not claim a pass from one Lighthouse run. Re-test representative templates under fixed conditions and verify field movement after the reporting window.

8. GEO / AEO — generative recommendation and citation

GEO here means increasing the probability that a brand, product, fact, or page is retrieved, mentioned, recommended, or cited in a generated answer. It is not “traditional SEO plus an llms.txt file.”

8.1 Source-selection model

Diagnose three distinct paths:

This separation is supported by vendors publishing different training, search, and user-fetch agents (OpenAI and Anthropic official crawler docs). Never use a bot hit, training opt-in, or organic rank as proof of generated inclusion.

For every observed answer, record whether search/retrieval was visibly enabled. If the interface does not disclose it, mark the path UNKNOWN.

8.2 Platform matrix

Verify current behavior at execution time; interfaces and controls change.

Official references:

• OpenAI crawlers
• Google AI features in Search
• Google common crawlers / Google-Extended
• Perplexity crawlers
• Anthropic crawlers

8.3 Crawler and retrieval eligibility

1. Read robots groups for the exact agent; account for precedence and each subdomain.
2. Check page status, canonical, robots meta, X-Robots-Tag, and rendered availability.
3. Check CDN/WAF policy and logs. Compare source IP against the vendor's current official ranges before attributing a request; user-agent strings are spoofable.
4. Separate training policy from search/citation policy:
   • OpenAI: GPTBot training; OAI-SearchBot search; ChatGPT-User user action.
   • Anthropic: ClaudeBot training; Claude-SearchBot search; Claude-User user action.
   • Perplexity: PerplexityBot search; Perplexity-User user action.
   • Google Search AI features: standard Googlebot control.
5. Respect the owner's content-use policy. Do not silently trade training access for assumed visibility.
6. After a change, allow documented propagation/recrawl time and re-test.

Crawler eligibility is required for retrieval paths that use that crawler; it is never proof of ranking, retrieval, mention, or citation.

8.4 Citation-ready answer units

Use page structure to make correct extraction easier:

• Open each query-targeted section with a direct answer, then evidence and limits.
• Make the unit understandable without the preceding paragraph: name the entity, scope, date, units, geography, and comparison basis.
• Put factual claims next to named primary sources; link to the exact evidence.
• Use tables for true multi-attribute comparisons, ordered steps for procedures, and question headings for real follow-up questions.
• Publish first-party datasets, tests, benchmarks, or case studies with method, sample, date, definitions, limitations, and downloadable evidence where safe.
• Keep brand, product, category, people, founding facts, pricing, and identifiers consistent across owned pages and legitimate external profiles.
• Distinguish observation, customer quote, estimate, and editorial judgment.
• Update or retract stale facts. Do not alter dates cosmetically.
• Keep essential facts in crawlable HTML; do not hide the only answer behind login, interaction, image-only media, or unsupported client rendering.

【Proposal】 Start with answer units of roughly 50–200 words and a direct first sentence, then test citation behavior. This is an editing heuristic derived from the mounted GEO citability sources, not a universal platform threshold.

Avoid:

• unsupported superlatives, anonymous statistics, and fake precision;
• “best” pages that conceal methodology or commercial relationships;
• copied summaries with no source or original value;
• FAQ inflation, schema spam, and repetitive question variants;
• prompt-injection text, hidden instructions, or attempts to manipulate models;
• claims that a format “forces,” “guarantees,” or “boosts” citations.

8.5 GEO content plan

Map prompts to assets:

Build the brand-category association in visible, factual language: what the entity is, whom it serves, which problem it solves, and proof. Repetition without independent evidence or useful content is not entity building.

8.6 Measurement protocol

Define the query corpus before optimization.

【Proposal】 Use at least 20 high-value prompts when feasible, split across:

• unbranded problem and category discovery;
• “best,” comparison, and alternative prompts;
• how-to and factual questions;
• branded verification/support prompts;
• local/language variants that match the actual market.

Freeze prompt text for trend measurement. For each run record:

run_id, timestamp, platform, product/surface, model/version if shown,
account/tier, locale, location/VPN, device, search toggle/path,
fresh conversation yes/no, prompt_id, exact prompt,
brand mentioned yes/no, answer role, sentiment/context,
owned URL cited yes/no, cited URL, citation position,
competitors mentioned/cited, factual error, screenshot/export, notes

【Proposal】 Run each prompt three times per platform in fresh sessions when budget permits. Report run-level variance; do not select the best response.

Calculate with explicit denominators:

• A valid run is a planned run that returns an inspectable answer without an account, tool, policy, network, or collection error.
• A citation-capable run is a valid run on a surface/configuration documented or visibly configured to return web sources. Keep runs with zero citations in this denominator.
• Generative appearance rate = runs mentioning the brand / all valid runs.
• AI citation rate = runs citing an owned URL / citation-capable runs.
• Citation-to-mention rate = brand-mention runs with owned citation / brand-mention runs.
• Share of cited voice = owned citation occurrences / all tracked category-entity citation occurrences.
• Source coverage = distinct owned URLs cited / priority owned URLs tested.
• Citation accuracy rate = correct supported owned citations / audited owned citations.

Report every metric as numerator/denominator (rate). If the denominator is zero, report N/A and the zero-denominator reason. Report each platform separately. Combine only with predeclared weights tied to audience usage; show the unweighted data beside the composite.

8.7 Evidence ladder and attribution

• Bot hits prove requests, not citation.
• AI referral traffic undercounts activity when referrers are absent or altered.
• A citation proves one answer used a source, not that every statement came from it.
• Manual tests are snapshots, not population estimates.
• Never claim causality from a before/after change without a design that rules out query, model, index, competitor, seasonality, and personalization changes.

8.8 Experiment design

【Proposal】 Run controlled GEO experiments:

1. Choose matched page/query groups and save a pre-change baseline.
2. Change one class of lever: answer units, evidence, original data, entity facts, technical access, or internal discovery.
3. Preserve prompt corpus, platform settings, locale, and collection protocol.
4. Record crawl/index lag; do not start the post window before the changed page is observable.
5. Compare run-level rates and uncertainty, plus organic, referral, and conversion guardrails.
6. Keep, revise, or revert based on evidence. Store counterexamples.

Treat local GEO-tool scores as readiness diagnostics, not outcome validation.

8.9 GEO vs SEO priority

Default to shared work—technical access, useful pages, verifiable facts, and clear structure—then fund platform-specific work only when measured opportunity warrants it.

8.10 llms.txt

llms.txt is a community proposal, not a universal ranking or citation directive. The mounted GEO sources support checking and experimenting with it, but this V6 found no official OpenAI, Google Search, Perplexity, Anthropic, or Doubao claim that publishing it improves inclusion.

【Proposal】 Add it only when:

• maintenance ownership exists;
• links are canonical, public, and useful;
• it does not expose private or unpublished material;
• it is measured as an experiment.

Its absence is not an SEO/GEO defect. Never give it readiness points merely for existing.

8.11 GEO readiness rubric

【Proposal】 Score each dimension 0=blocked/absent, 1=partial/unverified, 2=working with evidence:

1. retrieval eligibility;
2. index/render availability;
3. answer-unit clarity;
4. claim evidence and first-party value;
5. entity consistency;
6. query-to-asset coverage;
7. repeated platform measurement;
8. attribution/business-outcome linkage.

Report the eight values separately. A total is an internal triage aid, not a prediction of citation probability.

9. Backlinks and external authority

Profile analysis

• Use GSC links plus an authorized provider when available; state coverage limits.
• Review referring domains/pages, relevance, editorial context, destination, anchor distribution, acquisition trend, and lost links.
• Separate manipulative links from normal scraper/spam noise.
• Check unlinked brand mentions and incorrect citations that can be reclaimed.

Risk handling

Do not call links “toxic” from a provider score alone. Remove or disavow only with documented evidence of manipulative activity and material risk, such as a manual action or known paid-link scheme. Preserve a decision log.

Earn links and citations

• original datasets, tools, benchmarks, and transparent methods;
• expert contributions and primary-source commentary;
• genuinely useful comparison/reference pages;
• broken-link replacement where the asset is a true substitute;
• correction outreach for inaccurate facts or broken citations.

No PBNs, paid-link laundering, mass guest-post spam, or automated outreach sludge.

10. Analytics wiring (GSC + GA4)

GSC

• Distinguish verification-token presence, current verified ownership, API/UI access, and data availability.
• A meta/DNS token is evidence of configuration, not proof of current access.
• Domain properties may cover subdomains; do not demand a subdomain meta token.
• Record property type, date range, search type, country/device filters, and anonymization/row-limit caveats.

GA4

• Distinguish source-code presence, network request, DebugView/realtime event, stream configuration, and report/API access.
• On Liz house properties, preserve G-TXVLTJJ878 exactly. Outside section 13, call it a measurement ID.
• Test consent state, duplicate tags, cross-domain rules, SPA navigation, and conversion/key-event semantics when in scope.
• Never infer collected users or conversions from a tag string alone.

Reporting shape

Show clicks, impressions, CTR, average position, organic sessions/users, engaged sessions, conversions/key events, landing page, query cluster, locale/device, and comparison window only when the source is accessible. Otherwise mark UNVERIFIED.

11. SERP and competitor analysis

Capture exact query, locale, device, date/time, personalization state, and source.

• Record organic results, local/video/image/shopping/news features, featured snippets, discussions, AI Overviews/AI Mode, and cited sources.
• Compare intent coverage, evidence, entity authority, information gain, format, freshness, UX, links, and conversion path—not word count alone.
• Separate true business competitors, organic competitors, and AI-cited sources.
• Identify answer/source gaps the site can satisfy honestly.
• Re-check volatile SERPs; one observation is not a stable market fact.

Comparison/alternative parity floor

【Proposal】 Before publishing a comparison or alternative page, open the competitor's same-type URL in Ahrefs/Semrush (or authorized equivalent). Require parity or better on: traffic estimate, keyword coverage, content depth, evidence density. Else BLOCK publish. Qualitative SERP notes without this floor check are incomplete.

Output an opportunity table with query cluster, observed surface, winning source, why it may satisfy the task, evidence gap, feasible asset, impact, effort, and confidence.

12. Rank, AI-visibility, and outcome tracking

Tracking setup

• Freeze keyword/query corpus, landing-page mapping, locale, device, and platform.
• Track organic rank/landing URL beside generated mentions/citations.
• Annotate releases, migrations, campaigns, algorithm events, model/product changes, and known outages.
• Segment branded/unbranded, intent, market, template, and conversion value.
• Keep raw observations so metric definitions can be recomputed.

Page-level result ownership

Assign one owner per priority URL. Success = that URL's GSC impressions + CTR (and mapped conversions when available), not “we shipped SEO.” Unowned URLs stay UNVERIFIED for outcome claims.

Decision rules

【Proposal】 Predeclare stop/continue rules per initiative. Example:

• continue when leading evidence improves without technical or conversion harm;
• investigate when the canonical URL changes, variance spikes, or citations become inaccurate;
• stop when the site has no publish path, no audience fit, no measurable surface, or maintenance cost exceeds documented value.

Do not call a SERP permanently occupied or an initiative causal from two points.

13. House rules (Liz) — verbatim-accurate

Follow exactly on house properties:

1. GA4 shared property ID G-TXVLTJJ878 — never change/remove.
2. No brand/asset changes: logos, fonts, palette, icons.
3. Copy pain-first, EN default; no AI-slop filler.
4. Minimal change; never rewrite whole files; no new dependencies.
5. Never fabricate GSC verification tokens; REPORT presence only.
6. Don't touch projects/_templates/**, docs/AUDIT.md, payment code.
7. red-flowers (reading site) is private-by-design (noindex + WAF CN) — never SEO-refine it.
8. Verification:
   • git status -s matches claimed files (leave pre-existing dirt unstaged)
   • grep GA4 ID count on touched HTML
   • sitemap <loc> count sane; robots head sane
   • Stage selectively; never git add -A

Extra house semantics (operational)

• Apply these rules only after confirming a Liz house property.
• Preserve pre-existing changes and report them separately.
• Treat production bindings as strong host-intent evidence; investigate conflicts before migration.
• Do not stage or commit unless explicitly requested.

14. Execution algorithm

1. Scope gate: objective, target, access, authorization, exclusions, sample.
2. Snapshot: git/worktree state if local; date, host, redirects, robots, sitemaps.
3. Technical gate: reach/render/crawl/index/canonical before content polishing.
4. Analytics gate: identify observable and unverified layers.
5. Page/content/schema/CWV audit on representative templates.
6. Demand/SERP/competitor work when query evidence is in scope.
7. GEO audit:
   • platform and retrieval-path matrix;
   • crawler/WAF eligibility;
   • query-to-asset and answer-unit review;
   • fixed-corpus baseline;
   • citation, mention, accuracy, and source metrics.
8. Backlink/external-authority review when data exists.
9. Prioritize dependency first, then impact × confidence ÷ effort; keep the raw dimensions visible rather than pretending the quotient is precise.
10. Make only authorized minimal changes.
11. Verify changed files, behavioral checks, analytics preservation, and regressions.
12. Report method, evidence, unknowns, rollback, owners, and next measurement.

Severity rubric

Severity depends on affected scope and business impact. Missing optional schema or an AI crawler blocked by policy is not automatically P0/P1.

15. Commands and evidence cheatsheet

Set variables; do not paste angle-bracket placeholders into a shell:

host="example.com"
base="https://${host}"

# Final GET status, effective URL, redirects, timing.
curl -sS -L --max-redirs 10 --connect-timeout 10 --max-time 30 \
  -o /dev/null -w 'status=%{http_code} url=%{url_effective} redirects=%{num_redirects} time=%{time_total}\n' \
  "${base}/"

# Full response headers for one GET.
curl -sS --connect-timeout 10 --max-time 30 -D - -o /dev/null "${base}/"

# Robots and sitemap content. Inspect complete files or save outside protected paths.
curl -sS --connect-timeout 10 --max-time 30 "${base}/robots.txt"
curl -sS --connect-timeout 10 --max-time 30 "${base}/sitemap.xml"

# Compare responses to user-agent strings; this does not authenticate vendor bots.
curl -sS -A "OAI-SearchBot" -o /dev/null -w '%{http_code}\n' "${base}/"
curl -sS -A "PerplexityBot" -o /dev/null -w '%{http_code}\n' "${base}/"

# House verification.
git status -s
grep -oF "G-TXVLTJJ878" "path/to/touched.html" | wc -l

Parse XML/HTML with a real parser when correctness matters. Do not use line-count grep as an XML element count, regex as a complete HTML parser, HEAD as proof of GET behavior, or source presence as proof that analytics fires.

Never claim GSC/GA4 UI state, index inclusion, bot identity, or generated citation without direct evidence.

16. Output template

# SEO/GEO report — [host] — [date]

## Verdict
SHIP | FIX | BLOCK | KILL

## Method and limits
- Scope/sample:
- Access:
- Locale/device/window:
- Unknowns:

## P0 / P1 / P2
| Priority | Finding | Evidence | Scope | Confidence | Impact | Effort | Owner |

## Technical / on-page / content / schema / CWV
- Observed:
- Expected:
- Evidence:
- Action:

## GEO
- Target platforms and retrieval paths:
- Crawler/WAF eligibility:
- Query corpus and run protocol:
- Valid runs and run-level variance:
- Appearance rate — numerator/denominator:
- AI citation rate — numerator/denominator:
- Citation-to-mention rate — numerator/denominator:
- Share of cited voice — numerator/denominator:
- Source coverage — numerator/denominator:
- Citation accuracy — numerator/denominator:
- Cited URLs and competitors:
- Answer/evidence/entity gaps:

## Analytics and outcomes
- GSC verification/access:
- GA4 G-TXVLTJJ878 on house properties: intact|missing|n/a
- Organic/referral/conversion evidence:

## Changes and verification
- Files/actions:
- Tests:
- Rollback:

## Next actions
1. [dependency-first action, owner, due/measurement]

17. Failure patterns (do not repeat)

V6 additions:

18. Source synthesis and claim policy

Use this precedence:

1. current first-party platform documentation for product/crawler controls;
2. primary research for measured effects;
3. direct site, log, analytics, and fixed-protocol observations;
4. mounted practitioner tools as hypotheses and executable heuristics;
5. 【Proposal】 for unvalidated thresholds or internal decision rules.

Mounted source synthesis:

• GEO Optimizer: audit model, crawler/access diagnostics, Princeton GEO method summary, measurement/drift concepts.
• geo-skills: answer-block, citability, platform-test, crawler, and com

…(truncated)