SEO Master

SEO Master Skill (SEO + GEO + AEO, 2026)

Battle-tested playbook distilled from optimizing a real SaaS (letx.app, a collaborative LaTeX editor) against AI-native competitors. Everything here is applied knowledge, not theory.

The one-paragraph mental model

In 2026 there are two audiences: classic search crawlers (Google, Bing) and AI answer engines (ChatGPT, Claude, Gemini, Perplexity, Copilot). They overlap but differ. SEO ranks you in a list of links. GEO/AEO gets you cited inside an AI's generated answer. A solid SEO foundation is required for both; AEO/GEO adds: answer-first writing, dense structured data (especially FAQPage), fact density, freshness, and being present where each AI grounds its search. Critically — each AI uses a different search index, so "getting indexed" is not one task but several.

How to use this skill

1. Detect the stack first (package.json → Next.js App/Pages Router, React SPA, Astro, Remix…). Implementation differs by render model. See references/framework-seo.md.
2. Read references/ai-engines-and-indexes.md — determines where to optimize. Wrong index = invisible to that AI. (Coding agents like Claude Code / Codex / OpenClaw / Hermes mostly ground on Brave.)
3. Skim references/whats-new-2026.md for the I/O 2026 + May core-update context (AI Mode on Gemini 3.5 Flash; "AEO/GEO is still SEO"; llms.txt is optional).
4. Run the audit in references/implementation-checklist.md against the site.
5. Fix in priority order: indexing reach → structured data → answer-first content → titles/meta → content moat.
6. Use references/search-console-setup.md for Google + Bing + IndexNow, and references/structured-data.md / references/content-and-aeo.md as copy/paste references while implementing.

Reference files (read as needed)

Hard rules (learned the hard way)

• Never fabricate aggregateRating/reviews in schema. Google structured-data policy violation → manual action. Only use rating schema backed by visible, real on-page reviews. Use featureList / plain claims instead.
• Never copy a competitor's stat. ("20M users" belonged to Overleaf, not us — caught and removed.) Every claim must be true and yours.
• Title ≤ 60 chars or Google truncates. Long brand descriptors don't fit on content pages — use a short | Brand suffix there.
• FAQ schema requires the FAQ to be visible on the page. Schema-only FAQ = policy risk. Always render the questions AND emit JSON-LD.
• Code-split heavy content (course data, KaTeX, etc.) so the main JS bundle every visitor downloads stays small. Prerendered HTML keeps full content for SEO regardless.
• IndexNow does NOT notify Google — only Bing/Yandex/DuckDuckGo/etc. Google uses its own crawl + Search Console.
• Detect the framework before implementing. Next.js App Router uses the Metadata API (metadata/generateMetadata) — not helmet, not prerender. Only hand-rolled React SPAs need the Puppeteer prerender. Wrong tool = wasted work.
• Don't over-optimize for AI citations. The May 2026 core update appears to demote shallow answer-bait. Depth + real E-E-A-T + genuine fact density win; thin Q&A spam is now a risk, not a hack.
• llms.txt is optional, low priority — Google says it "isn't needed for AI Search." Add it only if cheap; never before content/schema/indexing.
• Exactly ONE canonical per page, verified with curl on the deployed site. Prerendered SPAs bake the static index.html canonical into every page and helmet adds a second — both wrong. The crawler reads baked HTML, not your DevTools. (Found live on a 500-user product; suppressed the whole site.)
• Trailing slash: standardize on whatever the HOST redirects to (Cloudflare Pages 308s to slash). Canonicals, sitemap, og:url, and internal hrefs must all use that one form, or GSC splits impressions across slash pairs.
• Segment GSC before concluding anything. Blended CTR/position lie: split brand vs non-brand, pos 4–15 (snippet problem) vs pos >15 (content-depth problem). New-content launches drop blended metrics mechanically — mix shift, not a penalty.
• FAQ rich results no longer show for normal sites (gov/health only since 2023). Keep FAQPage for AI citation; for SERP CTR use title/meta rewrites and ItemList-with-images on collection pages instead.
• Don't merge "duplicates" reflexively. Two pages at pos 5–6 for one query = double SERP presence (keep + differentiate); and near-identical slugs may be genuinely different products — verify content before 301ing.
• A single well-formed canonical can still be wrong. Resolve it: if the URL the canonical names returns a 3xx, every page is pointing at a redirect. Found sitewide on a live site whose canonicals all omitted the trailing slash its host 308s to. curl the canonical target, don't just count the tags.
• Audit built HTML with a parser, never grep. A code comment mentioning a tag name in literal markup makes grep report duplicates that do not exist in the DOM. Parse, and require 100% of pages to pass — not "most".
• Never append a brand suffix unconditionally. ${title} | ${Brand} pushed 34 of 50 post titles past 60 chars, so every one rendered truncated. Append only when the result still fits; a bare title beats a clipped one.
• Verify your CSS actually styles what your content emits. A site had @tailwindcss/typography in package.json but never registered in tailwind.config.ts — .prose was six hand-written rules with no table or code styling, so a new comparison table would have rendered black-on-black. Publishing a format the stylesheet doesn't cover is a rendering bug, not a content win.
• Check a list's inclusion criteria before planning an off-site PR. A 28k-star "free LLM API resources" list looked like the highest-leverage backlink available — but it indexes providers that serve inference, and the product was a client that calls them. The PR would have been rejected. The in-scope lists had 0–7 stars and no ranking power. Read the actual entries first; a rejected PR costs credibility.

Small-site triage — read this BEFORE writing anything

The default instinct on a low-traffic site is to publish more. On three measured properties that instinct was wrong every time:

letX segmented: branded 543 clicks @ 13.4% CTR; non-brand 44 clicks @ 0.48%. A content engine manufactures impressions, not visitors, until something off-site makes the brand worth clicking. And since the June 2026 spam update, a large zero-demand footprint is an active liability rather than dead weight.

So when a site is under ~100 clicks/quarter, in this order:

1. Fix signal suppression first. Canonicals (including the redirect check above), trailing slash, sitemap accuracy, title truncation. This is free ranking you already earned and are throwing away.
2. Prune, don't publish. Keep pages that earned impressions, belong to the one cluster that converts, or contain first-hand work nothing else can duplicate. noindex + drop-from-sitemap the rest — do not delete: keep them reachable and follow, so nothing 404s and the decision stays reversible. (Applied: 62 of 79 posts withdrawn, sitemap 112 → 30 URLs.)
3. Deepen the one cluster already in striking distance. Find the query group at position 4–15 with real impressions and put everything into it. One page that is genuinely the best on the internet beats twenty that are adequate.
4. Then go off-site. See the priority list below — this is the actual lever.

New pages are justified only where GSC already shows measured demand and no page serves it. "This topic seems relevant" is not evidence; an impression count is.

The priority order (highest ROI first)

1. Indexing reach — be in Google, Bing, and Brave indexes; wire IndexNow. (Feeds ChatGPT via Bing, Claude via Brave.)
2. Structured data — FAQPage on every page that can carry it (~3.2× AI Overview citation). Plus SoftwareApplication/Product/Article/HowTo.
3. Answer-first content — direct answer in first 40–60 words; 134–167-word self-contained passages.
4. Titles/meta — consistent, keyword-first, ≤60 chars.
5. Content moat — guides/tutorials/error-fix hub targeting the long-tail the competitor monetizes.
6. Off-site — get listed in the "alternatives" listicles AI engines cite.