SEO/GEO Evaluation Checklist

Run this after a material search, content, crawler, schema, or AI-discovery change. The gate asks whether the right audience can discover, understand, use, and convert from the work—not whether a page contains fashionable GEO artifacts.

0. Freeze the evaluation contract

Before changing the site, record:

• the exact audience and business outcome;
• the page(s), query/intent, locale, device, and product surfaces in scope;
• the current crawl/index/render state;
• baseline Search Console and first-party analytics values over a named window;
• the intervention and expected directional effect;
• the earliest responsible observation date and the stop/rollback condition.

Do not start from “increase AI visibility.” Name whether success means a mention, citation/link, engine-reported impression, referral visit, signup, revenue event, or another measurable outcome.

1. Technical search gate

Crawl, index, and canonical state

• Public canonical URLs return the intended successful status without login, challenge, or redirect loops.
• robots.txt, meta robots, X-Robots-Tag, canonical tags, and sitemap membership agree with the intended public/private state.
• Important content is present in the rendered HTML/DOM and is not dependent on a fragile client-only path.
• Internal links make the page discoverable from an owned hub; no public page depends only on a sitemap.
• Canonicals are page-specific, absolute, and do not accidentally inherit the home page.
• Search Console URL Inspection or equivalent first-party evidence confirms what the crawler received.

Sitemap priority and changefreq values, title character counts, or a successful typecheck are not visibility proof by themselves.

Crawler-policy matrix

Choose policy per surface. Do not require every bot, and do not confuse training permission with search discovery.

• The policy matches the business decision for each surface.
• CDN/WAF/bot mitigation and server logs confirm allowed agents can actually fetch the relevant paths.
• Private/authenticated/API areas remain protected independently of marketing-page crawler choices.
• Any crawler name or behavior was re-verified against current first-party documentation.

Metadata and structured data

• Each page has an accurate title, description, canonical, and share image appropriate to the page.
• Structured data matches visible content and passes the relevant validator.
• Organization, author, product, offer, review, and FAQ facts are truthful and internally consistent.
• No hidden FAQ, fabricated rating, invented author credential, stale price, or schema-only claim exists.
• Schema is used for ordinary search/product eligibility, not presented as an AI-citation guarantee.

Page experience

• Core Web Vitals and key flows are measured on representative mobile and desktop conditions.
• Headings, landmarks, labels, focus, media alternatives, and contrast produce a useful accessibility tree.
• Critical metadata and primary content occur before pathological response bloat; the Googlebot per-resource byte ceiling is not misreported as total page weight.
• Important facts remain usable without animation or unnecessary JavaScript.

2. Content and entity gate

• The page answers a real audience need with unique, useful, non-commodity information.
• Firsthand experience, original data, or primary-source synthesis is visible where the topic needs it.
• Claims carry dates, sources, methodology, and limitations proportional to their stakes.
• Headings and prose describe the subject naturally; there is no keyword-density target or exact-match stuffing.
• Comparison pages use actual product experience, dated pricing, honest best-for guidance, and material tradeoffs.
• New pages are not one-per-guessed fan-out query. A variant page must add durable value that the maintained owner cannot provide.
• Entity names, product names, authors, organizations, URLs, and prices agree across the page, metadata, structured data, and official profiles.

Direct answers, lists, and tables are layout choices. Use them when they clarify the material; no fixed 40–60-word, 134–167-word, FAQ-count, or article-count target is a universal ranking rule.

3. Measurement gate

Google Search and generative-AI features

• The overall Search Console Web/Performance report remains the baseline.
• If the property has access, the dedicated Search/Discover generative-AI report is captured for impressions, pages, countries, devices where supported, and time trends.
• Search Console results are joined to first-party engagement and conversion data.
• Google-only observations are not generalized to ChatGPT, Claude, Gemini, Perplexity, or Copilot.

Google launched the dedicated reports to a subset of sites on June 3, 2026; lack of access is a rollout state, not proof that the feature has no reporting. Source: Google Search Central

Cross-product prompt panel

For each repeated observation, preserve:

• exact query and intent cohort;
• product/model and mode;
• date/time, locale, device, and account/session state;
• answer or export, named brands, and exact cited/linked URLs;
• whether the event is a mention, link/citation, referral, or conversion;
• repeated trials and known volatility.

One answer screenshot is an anecdote. It can prove what that run displayed, not a stable rank or why the system selected the source.

Third-party tools and dashboards

• The tracker exposes its query panel, collection cadence, product/model coverage, definitions, and exportable evidence.
• Vendor metrics remain labeled vendor metrics; no tool is described as having internal Google ranking or AI access.
• Domain Rating, backlinks, citation counts, organic traffic, and conversion are displayed as separate measures.
• A dashboard screenshot includes property, filters, time window, comparison window, and source provenance before it is used in a decision.

The reviewed PressWhizz-style dashboard is useful as an interface pattern because it places AI-answer observations next to ordinary search and authority measures. It does not prove a budget allocation or causal relationship. Source: X/@Charles_SEO and local image review

4. Experiment and causality gate

• Baseline and post-change windows are long enough for the surface and crawl cycle.
• The intervention date, concurrent site changes, seasonality, campaigns, and major external events are recorded.
• A comparison page, query cohort, property, or time-series control exists when a causal claim is important.
• Movement is described as observed correlation unless the design supports attribution.
• Negative or null results are retained; the workflow does not publish only wins.
• A vendor case study is not promoted into a house rule without independent method and same-context evidence.

5. Agent usability gate

• Public pricing, limits, compatibility, support, contact, and product facts are visible and current where agents or buyers need them.
• The DOM and accessibility tree expose meaningful controls and content.
• Optional llms.txt, pricing markdown, feeds, or agent protocols match the canonical site and have a named consumer/test.
• Google Search inclusion is not made conditional on llms.txt, special AI files, or special AI schema.
• Agent-facing files cannot leak private, internal, stale, or contradictory facts.

6. Ship, hold, and writeback

Ship only when

• technical discovery and privacy states match intent;
• content claims are useful, accurate, sourced, and non-duplicative;
• measurement has a frozen baseline and named evidence lane;
• crawler policy separates search/fetch from training controls;
• automated tests, validators, and representative browser checks pass;
• canonical owners, skills, routing, and logs receive the durable learning.

Hold when

• crawler behavior, source rights, private/public state, or vendor methodology is unresolved;
• the proposed work is page proliferation, schema theater, keyword stuffing, fake freshness, or inauthentic mention building;
• an observed answer is being called a stable rank or a backlink/format is being called causal without proof;
• the team cannot name the audience, outcome, baseline, or rollback condition.

Proof receipt

Record the changed files, URLs, source revisions, screenshots/exports, query panel, test output, Search Console/analytics window, unresolved blockers, and exact follow-up date. A score without those receipts is not a gate.

---

Timeline

• 2026-08-11 | Replaced the legacy domain-specific checklist, all-bots allowlist, keyword-density targets, fixed answer lengths, mandatory llms.txt quotas, universal FAQ/schema claims, and screenshot-based AI-score gate with an audience-first, per-surface, evidence-lane contract. Source: Writing and Content Skills; skills/personal/ai-seo/SKILL.md
• 2026-07-04 | Added the multi-metric dashboard image as a report-shape signal; it remains interface evidence, not causal proof. Source: X/@Charles_SEO
• 2026-04-09 | Created the first post-change evaluation checklist from a project proposal and the then-current SEO/GEO playbook. Source: local project proposal