# SEO + GEO Checklist & Scoring Rubric

Shared by the **write** and **rewrite** tasks. Two jobs, not one: rank in classic search (**SEO**)
*and* get quoted by AI answer engines - Google AI Overviews, ChatGPT, Perplexity, Claude (**GEO /
AEO**). They overlap heavily and GEO tactics don't hurt SEO, but they are not the same discipline.
Optimise for both.

**First, learn what the project already emits** (Step 0). Most MDX blog setups auto-generate, from
frontmatter, per-post canonical URLs, `article` OpenGraph, and `Article`/`BlogPosting` +
`BreadcrumbList` JSON-LD, and include posts in a sitemap. **Confirm which of these the project
actually does** (read the post renderer + check for a `sitemap`/`robots` file) and give those
mechanisms good inputs - accurate, complete frontmatter and a well-structured body. Don't hand-add
schema the project already emits. If a mechanism is **missing** (e.g. no sitemap), note it and lean
harder on internal linking; flag it for the user rather than silently assuming coverage.

## On-page SEO checklist

**Keyword & intent**
- [ ] One clear **primary keyword/phrase** from real search intent (informational for blog posts).
      **Read the live SERP for it first** and match the dominant format (guide vs. comparison vs.
      list) - the ranking pages are Google's own answer key.
- [ ] 2-4 **secondary keywords** / semantic variants + the related **entities** the topic needs
      covered (topical depth is the on-page factor now, not raw length).
- [ ] Primary keyword appears **naturally** in: `title`, `description`, the first ~50-100 words, and
      at least one `##` heading. **Never stuff** - keyword stuffing is now an active negative for
      both SEO and AI citation. Read it aloud; if forced, cut it.

**Title (`title` frontmatter = SEO title)**
- [ ] **Front-load the payload** (keyword/benefit first). Google rewrites ~60%+ of titles and
      truncates around ~580px/~60 chars, so the tail is expendable - the front is not.
- [ ] Specific and compelling - a reason to click, not just a label.

**Meta description (`description` frontmatter)**
- [ ] ~120-155 chars, benefit-led, includes the primary keyword. Treat as CTR copy, but expect
      Google to rewrite it - so put the same fact in the body too.

**Structure**
- [ ] Body starts at the level the renderer expects (usually `##`; the H1 is the rendered title -
      exactly one H1 per page).
- [ ] **Answer-first**: the opening answers the query in ~40-50 words (wins featured snippets and AI
      citations). Inverted pyramid throughout - point first, detail after.
- [ ] Logical heading hierarchy, no skipped levels; headings **phrased as the real question** a
      reader/AI would ask, mined from "People Also Ask" where natural.
- [ ] Scannable: short paragraphs (one idea each), a list or two, a pull-quote, bolded anchors.

**Linking** (use the project's real routes, from Step 0)
- [ ] **3-5 contextual internal links** to relevant existing pages of *this* site (its key
      product/service/pricing/about pages and sibling posts) - **discover them; never invent a
      route** or link a page that 404s.
- [ ] **Descriptive, varied anchor text** (not "click here", not the same exact anchor twice).
- [ ] **1-3 external links** to credible sources (per the research protocol).
- [ ] All links resolve - internal routes/posts exist; external URLs are live.

**Media**
- [ ] Every image has descriptive, specific `alt` (not "image" or the filename).
- [ ] The post has a **hero image** (see the **[/scribekit-hero](../scribekit-hero/SKILL.md)** skill) - a
      local file the site serves, ideally ~1200x630 for social cards.

**Frontmatter / structured-data completeness** (feeds the auto-emitted schema)
- [ ] `title`, `date` (quoted!), `description` present and valid.
- [ ] `keywords` set (~3-6).
- [ ] `categories` set (usually one, from a small reused taxonomy; always a list) - drives the
      overview's category filters.
- [ ] `author` set (or intentionally defaulted). A **named human with a one-line credential** is a
      stronger E-E-A-T + AI-trust signal than an org byline - prefer it when a real person is behind
      the post.
- [ ] `image` set to the hero (becomes per-post OG/Twitter + JSON-LD `image`).
- [ ] `updated` set (quoted) when revising an existing post - freshness is a live signal.

## GEO / AI-answer-engine checklist (the SOTA layer)

Get quoted by the answer engines, not just ranked. Every item is evidence-backed:

- [ ] **Answer-first per *section*.** Open each `##`/`###` with a 1-2 sentence direct answer, then
      elaborate. Sections that don't state their answer standalone don't get selected - even if the
      page ranks.
- [ ] **Standalone-quotable passages.** One concept per section; don't fuse a definition with a
      how-to or bury a stat mid-narrative. Assume each passage is lifted out of context.
- [ ] **Concrete statistics with named sources** (GEO study: ~+41% visibility) - numbers beat
      adjectives. **But the lift comes from verified, well-attributed claims being present, not
      from their count**: ~5-7 load-bearing data points per post, each verified per the research
      protocol, beats a stat-stuffed page - and a wrong stat is strictly negative (an
      easily-verified factual error is a search-quality demotion signal, and an answer engine that
      repeats it propagates the error under your brand). Name the issuer and year in the sentence
      ("Gartner's 2025 survey found ..."), link the issuing document, and date time-varying figures.
- [ ] **Inline citations to credible external sources** (~+115% for mid-ranked pages) - the single
      biggest lever. Link them.
- [ ] **A named-expert quotation** where one genuinely fits (~+28%).
- [ ] **A one-sentence definition** of the core term early - the extractable "X is Y" unit.
- [ ] **Front-load the fact** in each paragraph's first sentence (chunks can truncate the tail).
- [ ] **Fluent, authoritative prose** - clarity itself measurably lifts citation.
- [ ] **Freshness**: real `date`, and `updated` when revised.
- [ ] **Clean semantic structure** (proper headings, lists, tables) for reliable chunking.

## What changed - myths to actively avoid (2026)

- ❌ **FAQPage / HowTo schema for rich results** - deprecated; FAQ rich results were removed from
      Google Search in 2026 and HowTo earlier. Add an on-page FAQ only if it helps readers.
- ❌ **A word-count minimum** ("must be 1,500+ words") - not a ranking factor. Adding words without
      structure gives ~0 AI-citation lift.
- ❌ **Keyword density targets / stuffing** - measurably *worse* than baseline for AI citation.
- ❌ **Obsessing over exact title/meta character counts** - Google rewrites most; front-load instead.
- ❌ **`llms.txt`** - no major AI system uses it (Google confirmed); don't spend effort on it.
- ❌ **Treating hand-added schema/JSON-LD as a GEO lever** - controlled tests show ~no citation
      uplift. Emit the schema that matters for rich results (usually the project already does);
      that's enough.
- ❌ "click here" / identical repeated anchors, orphan posts, thin consensus rehash.

## Scoring rubric (100 points)

Used by the **rewrite** task to score a post. The **write** task self-scores before finishing and
does not ship below ~90 (a state-of-the-art bar, not a passing grade).

- **Content & Research - 25** - every factual claim/stat sourced and current, **scored off the
  claim table, not off impressions**: −5 each unsourced or stale load-bearing claim, and any ❌
  (unsupported/contradicted) or 🔀 (misattributed) verdict still standing in the text caps this
  category at 10/25 - a post with a known-wrong sentence cannot score well on research, whatever
  else it does. A ❌/🔀 row escapes the cap ONLY when the post *reports* the broken claim without
  asserting it (quoting a wrong number in order to debunk it) - and the justification must name
  each such row and quote the post sentence that shows it asserts nothing; "the post discusses it"
  is not an exemption, it is the dodge. The category's justification states the counts
  (N claims: N✅/N⚠️/N❌/N🔀). Clear information gain vs. what already ranks.
- **SEO & discoverability - 20** - intent matched to SERP format; title front-loads keyword;
  description benefit-led w/ keyword; natural placement (no stuffing); heading hierarchy; 3-5
  internal + 1-3 credible external links, descriptive/varied anchors, all resolving; frontmatter
  complete.
- **GEO / AI-answer extractability - 15** - answer-first opening *and* per section; passages
  standalone-quotable; stats + inline citations; question-shaped headings; core term defined;
  front-loaded facts.
- **Voice & Anti-slop - 20** - matches the project's voice; **zero banned phrases** (any banned
  phrase caps this category at ≤10); burstiness present; structural AI tells absent; no template feel.
- **Technical & MDX - 10** - valid MDX (parses, no stray/unclosed JSX); body starts at the right
  level; project punctuation rules honoured; local image paths; valid quoted-date frontmatter;
  JSON-LD will populate correctly.
- **Readability & conversion - 10** - scannable, well-paced, clean grammar, no padding; audience
  trust markers present; exactly one soft, non-salesy CTA.
- **Hero image** - not scored numerically here; the rewrite's **Verify hero** step (see `rewrite.md`)
  gates it separately: present, on-brand, correct ratio, legible at card scale, descriptive alt.

Report the score **per-category with a one-line justification** and a **prioritized, concrete fix
list** (most impactful first). Never report a bare number without the fixes.