SEO Audit

SEO Audit

A complete, professional SEO audit engine. It diagnoses why a site is or isn't discoverable across search engines (Google, Bing, and others) and across AI assistants (ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews), then hands back a prioritised, dev-ready remediation plan.

This skill is built to degrade gracefully. It works whether you have a live browser, only web_fetch, or nothing but a screenshot and a pasted robots.txt. Never block on a tool you don't have — gather what you can, state what you couldn't verify, and audit the rest.

How to run this: a conversation, not a form

A good audit needs a little from the user — but most of what a naive tool would ask, this skill detects instead. Asking for things you could crawl yourself (language, platform, rendering, competitors) is fake-collaborative and signals a dumb tool. The guiding rule:

Detect first. Surface what you found. Ask only the genuine unknowns.

The interaction has a natural three-beat rhythm. Adapt it — never run it as a rigid script.

Beat 1 — Scope (one short exchange, before crawling)

Establish only what you cannot detect. Bundle into one brief message:

• Depth: "I'll run the CORE audit by default — Crawlability, On-page, Structured Data, Performance, Multilingual (only if the site isn't plain English), and GEO/AI. Want any deep lenses? Local · E-commerce · Content/ Publisher · SaaS/B2B · Off-page & Authority · International."
• Private data: "Do you have Google Search Console / analytics access I can use, or should I treat this as an on-page/technical audit?" (Changes whether findings can include traffic/ranking data.)
• Intent (light touch): "Anything specific prompting this — a market you're targeting, a drop you noticed, a goal like AI visibility?" One line; don't interrogate.

Then go. Don't hold a long Q&A before any work — one scope exchange, then crawl. (If the user already specified scope, skip straight to crawling.)

Beat 2 — Recon recap (mid-audit, after crawling)

This is where it feels collaborative. Show you did the work, then ask only the consequential judgment calls the crawl surfaced but shouldn't be decided unilaterally. For example:

• "Your site renders real Bengali text (great) but declares lang=\"en\" — I'll flag that. Two things I want your call on first:"
• GEO values call (only if AI crawlers are blocked): "Your robots.txt blocks AI crawlers. Do you actually want to be discoverable in ChatGPT/Perplexity, or is keeping your content out of AI deliberate?" This is a values decision, not a technical default — never reverse it silently.
• Strategic forks the audit reveals (e.g. native-script vs romanized URL strategy, a positioning gap). Offer the tradeoff; let the user choose.

If the crawl surfaces no real judgment calls, skip Beat 2 — don't manufacture questions.

Beat 3 — Deliver

Confirm format (Markdown default; offer Word doc for client-ready) and produce the report.

When to skip the conversation entirely

If the request is fully specified ("full audit of x.com, I have GSC, US market, give me a Word doc") → skip Beats 1–2, just work. If it's a fast single-issue question ("is my robots.txt blocking Google?") → answer it directly, no ceremony. Judgment over ritual.

Choosing the question format

Pick by context. For a few crisp either/or choices where interactive option controls are available, use them — they're faster for the user. For open intent ("what market are you targeting?") or a nuanced tradeoff that needs explanation, use a short numbered list in prose. Never force buttons onto a genuinely open-ended question, and never spread questions across many turns when one bundled message would do.

What you detect vs. what you ask — quick reference

Do not run deep lenses unless asked. Do not pad the report with "N/A" sections.

Step 1 — Recon: gather what the environment allows

Audit quality depends on what you can observe. Get the richest data the environment permits, then proceed with whatever you have.

The full data-gathering playbook — browser-based live crawl, web_fetch fallback, the rendered-DOM-vs-raw-HTML test, the Bing-index check that routes around CAPTCHA, schema/robots/sitemap inspection, and how to detect client-side-rendering blindness — lives in references/live-recon-appendix.md. Read it whenever you have any live access (browser or fetch). If you have neither, work from user-supplied material (screenshots, pasted source, exported data) and clearly mark anything unverifiable.

The one recon check that changes everything: does the page's rendered content exist in the raw server HTML, or only after JavaScript runs? A raw fetch that returns an empty shell while the live page shows content means the site is client-side-rendered and search/AI crawlers may see nothing. This single test reframes the whole audit — run it first when you can.

Step 2 — Run the CORE pillars (always)

For each pillar: diagnose against the checks below, assign a RAG status (🔴 Red / 🟡 Amber / 🟢 Green), and capture specific, evidenced findings — quote the actual tag, score, or value you observed, not a generality.

Pillar 1 — Crawlability & Indexing

The foundation. If crawlers can't reach or render it, nothing else matters.

• robots.txt: present, well-formed, not accidentally blocking key paths or Googlebot/bingbot. Note any AI-crawler directives (these matter for Pillar 6).
• XML sitemap(s): present, declared in robots, fresh lastmod, accurate URL set. Watch for orphaned content — sections that are indexed but missing from the sitemap (a common, costly miss).
• Rendering: SSR/SSG vs client-side-only (see the critical recon test above).
• Canonicalisation: self-referential canonicals, no duplicate-content sprawl, no parameter/session-ID chaos.
• Index status: spot-check site: on Google and Bing — they often differ. A stale cache on one engine is a crawl-freshness issue, not a content issue.
• IndexNow (Bing/Yandex instant indexing): is it enabled? On Cloudflare/CDN hosts it's often a one-click toggle and a big quick win for frequently-updated sites.

Pillar 2 — On-page SEO

• Title tags: unique, descriptive, keyword-bearing, not truncated/duplicated.
• Headings: exactly one meaningful H1 per page, logical H2/H3 hierarchy.
• Meta descriptions: written (not auto-truncated), compelling, unique. Flag generic boilerplate that could belong to any site — it wastes the highest- value pages.
• Image alt text: present and descriptive (accessibility + image-search win).
• Internal linking: real <a href> links (not JS-onclick traps), descriptive anchors, healthy distribution of link equity.

Pillar 3 — Structured Data (Schema)

• What schema types are emitted, and where. Map coverage by page type.
• Coverage gaps are the usual finding: rich schema on detail pages but bare hubs/category pages — the pages most likely to rank are the least decorated.
• Rich-result eligibility: does the markup unlock stars, prices, breadcrumbs, FAQ, etc.? Flag missing aggregateRating/Review when a reviews UI exists.
• Validity: required properties present; types appropriate to content.

Pillar 4 — Performance & Core Web Vitals

• Pull field data when possible (PageSpeed Insights / CrUX). State clearly when you only have lab/structural signals, not field data.
• LCP, INP, CLS — the three that matter. Identify the likely LCP element.
• Structural signals you can check without field data: CDN presence, image delivery/optimisation, lazy-loading below the fold, render-blocking resources, excessive DOM/image counts.

Pillar 5 — Multilingual & Non-Latin Discoverability (conditional)

Run this pillar only if the site is not plain monolingual English. Detect first; if English-only, skip silently — no "N/A" noise.

Detection: scan the rendered content for non-Latin Unicode ranges and/or multiple languages. (Bengali \u0980–\u09FF, Devanagari \u0900–\u097F, Arabic \u0600–\u06FF, CJK \u4E00–\u9FFF, Cyrillic \u0400–\u04FF, etc.) If found, run this pillar parameterised to whatever language(s) you detected.

This pillar exists because mainstream SEO advice is English-blind, and the failure modes below silently destroy discoverability for billions of non-English searchers. Full guidance — including per-script notes and the phonetic-search strategy — is in references/multilingual-discoverability.md. The universal checks:

• Text-as-image trap (the #1 killer): is non-Latin text real, selectable Unicode, or baked into images? If it's images, crawlers and LLMs read nothing. This single issue sinks more non-Latin sites than all others combined.
• Language-signal vs content mismatch: does <html lang> match the actual page language? Is inLanguage set in schema? Is hreflang present? A page that is visibly Bengali but declares lang="en" with no inLanguage tells search engines the wrong thing — they won't serve it to native-script searchers. This is extremely common and extremely fixable.
• Phonetic / romanized search: every non-Latin language has a diaspora pattern where readers type the language on a QWERTY keyboard (Banglish for Bengali, Pinyin for Chinese, Romaji for Japanese, franco-arabe for Arabic). Sites that only carry native script miss this entirely. The fix: a visible romanized/transliterated line for titles and names.
• Three-mode coverage: for full findability a key term should appear in native script, English/translated form, AND romanized transliteration.
• RTL correctness (Arabic, Hebrew, Urdu, Farsi): dir="rtl", mirrored layout, correct bidi handling.

Pillar 6 — GEO / AI Discoverability (Generative Engine Optimisation)

Being cited and recommended inside AI assistants is now its own discipline. Full playbook — including the crawler opt-in table — is in references/geo-playbook.md. The core checks:

• AI crawler policy: does robots.txt block AI crawlers? Distinguish training crawlers (e.g. GPTBot, Google-Extended) from answer/search crawlers (e.g. OAI-SearchBot, ChatGPT-User, PerplexityBot). Blocking the search/answer bots opts the site out of AI citations, not just training — often unintentionally. Recommend the surgical split if the user wants AI visibility while withholding training.
• AI Overviews: powered by ordinary Googlebot — so strong classic SEO is the AI-Overviews strategy. No unblocking needed there.
• Answer-shaped content: pages that directly answer a question are far more citable than bare grids/listings.
• Entity clarity: Organization/Person/Product schema + sameAs links build the entity graph LLMs reason over.
• llms.txt: an emerging convention — a curated markdown map of key content. Cheap to add, increasingly recognised.

Step 3 — Deep lenses (only if requested in Beat 1)

Each lens has its own reference file — read it only when that lens is selected. Match the lens to the site type you detected: a store ⇒ E-commerce; a blog/news/ publisher ⇒ Content; a B2B product ⇒ SaaS; a local business ⇒ Local. Suggest the fitting lens during Beat 1 rather than waiting to be asked.

• Local SEO → references/deep-lens-local.md (Google Business Profile, NAP consistency, local pack, reviews, local schema)
• E-commerce SEO → references/deep-lens-ecommerce.md (product feeds / Merchant Center, faceted-nav crawl traps, pagination, product schema depth)
• Content / Publisher SEO → references/deep-lens-content.md (topical authority, content clusters & pillar pages, freshness/decay, author E-E-A-T, internal linking as topic modelling)
• SaaS / B2B SEO → references/deep-lens-saas.md (programmatic SEO, comparison/alternative pages, docs/KB indexing, bottom-of-funnel intent, product-led content)
• Off-page & Authority → references/deep-lens-offpage.md (backlink profile, E-E-A-T, brand signals, digital PR angles)
• International SEO → references/deep-lens-international.md (hreflang at scale, ccTLD vs subdomain vs subfolder, geotargeting, currency/locale). Distinct from Pillar 5: Pillar 5 is content discoverability in any script; this lens is the geo/locale-targeting machinery for multi-region sites.

Step 4 — Synthesise and produce the report

Use the structure in assets/report-template.md. Core requirements:

• Lead with an executive summary + a scorecard table (pillar → RAG/score → one-line read). Give a single honest composite verdict.
• Severity tiers for every finding: 🔴 Critical · 🟠 High · 🟡 Medium · 🟢 Low. Tie severity to visibility impact, not effort.
• Evidence over assertion: cite the actual observed tag/score/value. "The homepage declares lang=\"en\" while its H1 and body are Bengali" beats "i18n could be improved."
• Prioritised action table: impact × effort, with an Owner column (Dev / Content / Strategy) and an Effort estimate (S/M/L). Order by impact ÷ effort so the 80/20 is obvious.
• Be honest about scope: if you had no Search Console / analytics / field data, say so and frame findings as on-page/technical/structural — never fabricate traffic or ranking numbers.
• Default to Markdown. Offer a polished Word document (.docx) if the user wants a client-ready deliverable; if so and a docx skill is available, use it. For non-Latin content in a generated doc, ensure a font that covers the script is embedded (e.g. Noto Sans for the relevant language) or the text will render as boxes.

Self-contained but aware: handoffs

This skill is fully self-contained — it never requires another skill. But if the user has these installed and the audit would benefit, hand off rather than reinvent:

• Deep visual/aesthetic critique requested → visual-aesthetics-audit.
• Competitor comparison / benchmark requested → competitive-benchmark.
• Market positioning / pricing strategy → market-intelligence.
• Full WCAG accessibility audit → audit (a11y), though on-page alt/lang checks remain in scope here.

If those skills aren't present, do a lightweight inline version and note that a deeper dedicated pass is available.

Operating principles

• Graceful degradation is the whole game. Some data beats no audit. State confidence levels; never stall on a missing tool.
• Believe the evidence, even when it's surprising (a stale cache, an unexpected block). But re-check conspiracy-adjacent or high-stakes claims.
• Specific and kind. This is often someone's project or livelihood. Lead with what's working, be precise about what isn't, and make every problem actionable. The goal is a fixed site, not a impressive-looking teardown.
• Don't boil the ocean. The depth gate exists for a reason. Respect it.