seo-analysis

Audit a codebase for search visibility risks, or implement the fixes, at the requested scope, starting from the live repository and rendered output rather than assumptions about the framework. A focused question gets a focused answer with its evidence and limits: a title fix, schema question or content edit does not need a full audit. Write files only when asked.

How to audit

• Find the delivery model (SSR, SSG, ISR, CSR, mixed), the route map, layout and head/metadata abstractions, and the sitemap, robots, canonical, structured-data and image helpers. Review page types, not single pages: home, landing, category, detail, article/docs, paginated/filtered, and auth/account/utility. SEO failures usually come from shared template logic, so one broken helper can explain hundreds of pages. Deduplicate to root causes before prioritising.
• Verify rendered behaviour for runtime claims; SSR, hydration and edge rendering change what crawlers receive. When limited to source, label the result source-only and list the runtime checks still needed.
• Crawlability, renderability and canonicalisation come first; title tweaks do not matter if pages are blocked, duplicated or undiscoverable.
• Separate intentional exclusions from mistakes: login, cart, account, internal search, preview environments and thin utility pages are often meant to be noindex. Missing meta description is a preview-quality issue, not a blocker; a missing sitemap is minor on a small well-linked site; missing hreflang is fine for single-locale sites.
• Search titles, social titles, canonicals and schema serve different consumers; check search previews and OG/X previews separately.
• Schema must be truthful, valid and aligned with visible content. Rich-result eligibility is a separate question from validity, and "add more schema" is not a universal fix.
• AI-answer visibility is crawl access, renderable HTML, clear metadata, explicit facts and original content, plus preview controls (nosnippet, data-nosnippet, restrictive max-snippet) that do not suppress useful text. Do not invent AI-ranking guarantees or "AI keywords".

Each finding carries severity (blocker: pages that should rank cannot be crawled, rendered, indexed or canonicalised; high: indexable but systematically conflicting or weak signals; medium: missing or inconsistent supporting signals; low: polish), category, scope (route pattern, page type or template), evidence (code path, rendered markup, headers or runtime observation), impact and the smallest safe fix. "Metadata could be better" is not a finding. A full audit also lists which page types were and were not covered.

Google AI Overviews and AI Mode

For Google's generative Search features, including requests phrased as AEO or GEO, read references/google-ai-search.md before advising or changing code. It records Google's own dated guidance: these features run on the ordinary Search index and ranking systems, so this is SEO and not a separate hack track, and llms.txt, AI-only markup, special schema and artificial chunking are not required. Passing checks proves eligibility signals, never selection. Comparing Search Console, Merchant Center or Business Profile data does not authorise changing those accounts; get consent first.