# Content Optimization Guide: LLM & SEO
## How to Write Content That Ranks in Google and Gets Cited by AI

> **Purpose:** This guide covers the two machine-optimization layers for content writing: SEO (how Google ranks your page) and LLM optimization (how AI models like ChatGPT, Perplexity, and Gemini decide to cite your content). It is a technical reference — rules, data, placement maps, and checklists.
>
> **Data sources:** LLM rules are based on three analyses by Kevin Indig (Growth Memo) covering 1.2M ChatGPT responses, 98K citation rows, and 5,000+ pages across 7 verticals. SEO factors are from Cora's Top SEO Factors (topseofactors.com, April 2026), top 200 correlation data.
>
> **Examples:** All examples use the topic "best transactional email service providers" for illustration.

---

## Key Terms

**Exact match keyword** — the precise search query, word-for-word. "Best transactional email services."

**Keyword variations** — close rephrasings with the same intent. "Top transactional email providers," "best email delivery platforms," "transactional email services compared."

**LSI (Latent Semantic Indexing) keywords** — topically related terms that naturally co-occur with the subject. For "transactional email services": SMTP relay, email API, inbox placement, sender reputation, DKIM, SPF, DMARC, delivery speed, email authentication, bounce rate, webhook, dedicated IP.

**Entities** — named, specific things: product names (Mailtrap, SendGrid), protocols (DKIM), standards (SOC 2), organizations (Google, Yahoo), dates, numbers, prices.

---

## 1. Front-Load Key Insights

*LLM ✓ SEO ✓*

### LLM data

44.2% of all AI citations come from the first 30% of text. The true citation peak is the **10–20% band** of the page — the first substantive paragraphs after the heading structure settles. The 0–10% band is consumed by navigation, H1 tags, and intro filler. The bottom 10% (footer, author bio) earns only 2.4–4.4% of citations. Summaries and conclusions are nearly invisible to AI.

### SEO data

Variations in first 100 words is factor #84 (-0.16). Variations in the last 100 words is factor #194 (-0.12).

### The rules

- State the most important claim, definition, or recommendation in the first 2–3 paragraphs.
- Include the exact match keyword within the first 100 words.
- Include at least one keyword variation within the first 100 words.
- Include entities (product names, protocols, numbers) early.
- Include at least one date and one specific number in the intro.
- Include the exact match keyword or a close variation in the last 100 words.
- Include at least one keyword variation and 1–2 LSI terms in the closing section.

The first and last 100 words bookend the page with keyword signals for SEO. The first substantive paragraphs are where LLMs read hardest.

### Example

> A transactional email service is a platform that sends user-triggered emails — password resets, order confirmations, and signup verifications — through a dedicated API or SMTP relay. The best transactional email providers in 2026 are Mailtrap, SendGrid, Mailgun, Brevo, Amazon SES, Postmark, and Mailchimp Transactional Email. Mailtrap is the strongest choice for SaaS companies that need high deliverability and granular analytics; Amazon SES is the most cost-effective for teams already on AWS.

Exact match in sentence 1. Variation in sentence 2. Seven entities. Date (2026). 60 words.

---

## 2. Page Architecture: Value-First Section Order

*LLM ✓ SEO ✓*

### Why section order matters

Front-loading isn't just about what goes in the first paragraph — it's about which sections appear in the first 30% of the page. Both LLMs and search engines process pages top-to-bottom. If your answer is in section 5 and your background context is in sections 1–4, the highest-value page real estate is wasted on content nobody came for.

The same principle applies to human readers: an engineer evaluating transactional email services wants the comparison, not a 600-word explainer on what transactional email is.

### The core principle

**Put the answer before the education.** Definitions, background context, methodology deep-dives, and "how to choose" frameworks are valuable — but they belong after the reader already has the snapshot. The page should reward the reader (and the bot) immediately, then offer depth for those who want it.

**Focus beats comprehensiveness.** An analysis of 815,000 query-page pairs found that pages covering 26–50% of ChatGPT's fanout sub-queries get cited more than pages covering 100%. The "ultimate guide" with 8 sections of background context before the answer is the least reliable format for AI citation. Every section on the page should directly serve the primary query intent. If a section serves a different intent, it belongs on a different page — link to it instead of embedding it.

### Bad vs. good structure: comparison posts

**Bad structure (education-first):**

1. Intro
2. What is a transactional email service?
3. What is email deliverability?
4. Criteria to choose transactional email services
5. How to choose email services
6. Snapshot / ranked list
7. Dedicated provider reviews

The answer (which providers, who wins) doesn't appear until section 6. Sections 2–5 are background that could be a separate article. LLMs reading the first 30% of this page get definitions and methodology — not citable comparison data.

**Good structure (value-first):**

1. Intro (2–3 paragraphs: definition in sentence 1, ranked list in sentence 2–3, key data point in sentence 3–4)
2. Snapshot (ranked list with one-line verdicts + comparison table)
3. How to choose (compressed: decision map, not an essay. "Choose X when Y" format)
4. Dedicated provider reviews (one H2 per provider, consistent sub-sections)

The answer is in sections 1–2. The decision framework in section 3 adds practical value without delaying the reviews. Background definitions are handled in the intro sentence, not a separate section.

### Why this works for LLM and SEO

**LLM:** The snapshot section lands squarely in the 10–20% citation peak band. Entity-dense comparison tables and ranked lists in this position give the AI multiple citable data points where it reads hardest. Educational sections pushed early create a "definition zone" that the AI may cite for "what is X?" queries, but completely misses for "best X" queries — which are usually the primary target.

**SEO:** The snapshot section concentrates keyword variations, entities, and LSI terms at the top of the page. Variations in first 100 words (#84), entities in sentences (#21), and term frequency (#23) all benefit from having the most keyword-rich content early. An educational preamble dilutes keyword density in the top 30KB — variation count in top 30KB is factor #32 (-0.19).

### The "How to Choose" section: decision map, not essay

The criteria/how-to-choose section is the most commonly over-expanded part of comparison posts. It often becomes a 1,000-word methodology section that delays the actual reviews.

**Compress it into a decision map:**

> ## How to Choose a Transactional Email Service
>
> - **High deliverability at scale** → Mailtrap or Postmark (both separate transactional/bulk streams)
> - **Already on AWS infrastructure** → Amazon SES (lowest cost, deep AWS integration)
> - **Need marketing + transactional in one platform** → Brevo or Mailgun
> - **Developer-first with flexible API** → Mailgun or Mailtrap (both offer SDKs for Node.js, Python, PHP, Ruby)
> - **Volume-based pricing, no feature gating** → Postmark (all features on all plans)
> - **Enterprise scale with dedicated IP** → SendGrid (Pro plan and above)

This is entity-dense (6 providers, specific features, specific use cases), definitive (no hedging), and practically useful — all in ~100 words instead of 800. Each line is independently citable by an LLM because it contains a clear condition → recommendation structure.

If you need to cover evaluation criteria in depth (deliverability, reliability, scalability, security, etc.), place that section *after* the reviews, not before. It becomes reference material for readers who want to go deeper, rather than a gate between the reader and the answer.

### Structure templates by content type

**Comparison / roundup ("best X"):**

1. Intro (definition + ranked list + key data point)
2. Snapshot (comparison table: provider, best for, free tier, pricing)
3. How to choose (decision map — 6–10 condition → recommendation lines)
4. Provider reviews (H2 per provider, consistent H3 sub-sections)
5. Wrap-up (keyword in last 100 words)

If criteria methodology, deliverability deep-dives, or compliance comparisons add value, publish them as separate linked pages rather than embedding them. Each page should serve one primary intent.

**Head-to-head comparison ("X vs Y" / "X vs Y vs Z"):**

1. Intro (one sentence per product + verdict in sentence 2–3)
2. Snapshot (side-by-side table + decision map: "Choose X when…" / "Choose Y when…")
3. Feature comparisons (2–4 focused sections, each with a table + one-paragraph verdict)
4. Final verdict (restate recommendation per product)
5. FAQ (2–3 questions matching real comparison queries)
6. Wrap-up (keyword in last 100 words)

The "X vs Y" format is tighter than a roundup. The reader already narrowed to 2–3 options — they want the differentiator, not a product education. Every section should be comparative (side by side), not descriptive (one product at a time).

**"What is X" guide:**

1. Intro (definition in sentence 1 — this IS the answer)
2. Why it matters (applied context, not abstract importance)
3. Key components / how it works (the substance)
4. Common problems and solutions
5. Tools and resources
6. Wrap-up (keyword in last 100 words)

**How-to / tutorial:**

1. Intro (what you'll build/achieve + prerequisites)
2. Steps (H2 per major step, code blocks inline)
3. Common issues / troubleshooting
4. Wrap-up (keyword in last 100 words)

**Landing page:**

1. Hero (value proposition + primary CTA — definition in first sentence)
2. Key differentiators (entity-dense feature blocks)
3. Social proof / data
4. Secondary CTA

### How to audit an outline before writing

Before writing, run every section in your outline through three tests. This prevents the most common structural mistake — embedding education and post-decision content before the answer.

**Test 1: Intent match.** State the page's primary query in one sentence (e.g., "Mailtrap vs SendGrid vs Mailgun — which should I choose?"). For each section in the outline, ask: does this section directly help answer that query? If it answers a *different* query, it doesn't belong on this page.

**Test 2: Pre-decision vs. post-decision.** Is the reader looking at this section *before* they've decided (comparison data, pricing, feature differences) or *after* (setup guides, onboarding, DNS configuration)? Post-decision content belongs at the bottom of the page or on a separate page.

**Test 3: Verdict position.** Where in the outline does the reader first encounter the answer to the primary query? If it's past the 30% mark, move it up. The verdict should appear twice: once in the snapshot (top 15%) and once in the closing section.

### Sections to cut or absorb — common patterns

These section types appear in most first-draft outlines and almost always hurt LLM performance. For each, the table shows what to do instead.

| Section in the outline | Why it hurts | What to do instead |
|---|---|---|
| "What Is [Product]?" as standalone H2 (×3 in a comparison) | Different intent. Reader of "X vs Y" already knows the products. Wastes the 10–20% band on definitions. | One sentence per product in the comparison table's "Best For" cell, or in the intro. |
| "How to set up / get started" before the comparison | Post-decision content. The reader hasn't decided yet. | Move to the bottom, after the verdict. Or publish as a separate tutorial and link to it. |
| "What is [category]?" (e.g., "What is an SMTP service?") | Different intent. The reader searched for "X vs Y," not "what is SMTP." | Handle in the intro's first sentence as a single clause: "Mailtrap, SendGrid, and Mailgun are SMTP services that…" |
| Full section on a tangential topic (e.g., "SPF, DKIM, and DMARC Setup") | Tutorial intent embedded in a comparison. Dilutes focus. | One row in a feature comparison table: "SPF/DKIM/DMARC support: ✓ all three." Link to a dedicated authentication guide. |
| 1,000-word "How to Choose" methodology essay | Delays the answer. The decision map format delivers the same value in 100 words. | Compress into 6–10 "Choose X when Y" lines. Move methodology deep-dives to a separate page. |
| Criteria deep-dive before the reviews (deliverability, reliability, scalability as separate H2s) | Education-first. Each H2 serves a "what is X" intent, not the comparison intent. | Merge into 2–3 comparison sections, each with a table + verdict paragraph. Cover criteria *through* the comparison, not before it. |

### Example: Outline audit for "X vs Y vs Z" comparison

**First-draft outline:**

1. Key Takeaways
2. Overview ("What an SMTP service is, why the choice matters")
3. What Is Mailtrap? / What Is SendGrid? / What Is Mailgun?
4. Setup and Developer Experience
5. Email API and SMTP Support
6. Analytics, Logs, and Webhooks
7. Email Authentication: SPF, DKIM, and DMARC
8. Security and Compliance
9. Pricing
10. Which Should You Choose?
11. FAQ
12. Final Verdict

**Intent audit:**

Primary query: "Mailtrap vs SendGrid vs Mailgun — which should I choose?"

| Section | Intent match? | Pre/post decision? | Action |
|---|---|---|---|
| Key Takeaways | ✓ | Pre | Keep — move to top |
| Overview + "What Is X?" ×3 | ✗ Different intent | Pre | Cut — absorb into intro sentence and comparison table |
| Setup and Developer Experience | Partial | Post-decision | Move to bottom or absorb as one row in API comparison table ("time to first send") |
| Email API and SMTP Support | ✓ | Pre | Keep — becomes a comparison section with table + verdict |
| Analytics, Logs, and Webhooks | ✓ | Pre | Keep — merge with API section or keep as separate comparison section |
| Email Authentication | ✗ Tutorial intent | Pre | Cut — becomes one row in Security table. Link to authentication guide. |
| Security and Compliance | ✓ | Pre | Keep — comparison section with table + verdict |
| Pricing | ✓ | Pre | Keep — move up (readers need this before the verdict) |
| Which Should You Choose? | ✓ The answer | Pre | Keep — this IS the verdict. Move to section 2 as the snapshot. |
| FAQ | ✓ | Pre | Keep — adds query surface area |
| Final Verdict | ✓ | Pre | Keep — restates recommendation in last 100 words |

**Result after audit:**

1. Intro (one sentence per product + verdict)
2. Snapshot: Which SMTP Service Is Best? (comparison table + decision map — the verdict, in the 10–20% band)
3. Pricing Compared
4. API, SMTP, and SDK Support (table + verdict)
5. Deliverability and Analytics (table + verdict)
6. Security and Compliance (table + verdict)
7. Final Verdict (keyword in last 100 words)
8. FAQ

Sections cut from 12 to 8. The verdict moved from section 10 to section 2. Three "What Is X?" sub-sections absorbed into the intro. Setup guidance moved to a separate page. Authentication absorbed into one table row. Every remaining section directly serves the "X vs Y vs Z" intent.

### The anti-pattern: "What is X?" as section 2 of a comparison post

If the primary keyword is "best transactional email services," the reader came for a comparison, not a definition. Placing "What is a transactional email service?" as a standalone H2 section after the intro wastes the 10–20% band on content that serves a different query intent.

**Instead:** Handle the definition in the intro's first sentence ("A transactional email service is a platform that...") and move on to the comparison. If the definition needs depth, link to a dedicated "What is a transactional email service?" article rather than embedding a 300-word explainer in the comparison post.

This also avoids cannibalizing your own content — the "What is X?" guide targets that query intent, and the comparison post targets "best X." Each page serves one primary intent.

---

## 3. Definitive Language (No Hedging)

*LLM ✓*

### LLM data

Declarative language in the intro produces a **+14% aggregate citation lift** — the only writing signal that holds across all 7 verticals studied. LLMs penalize hedging: qualifiers like "might," "could potentially," "it's possible that" weaken semantic confidence.

### The rules

- Open sections with "X is Y" or "X refers to" sentence structures.
- Remove all qualifiers from opening paragraphs before any other optimization.
- If you find yourself hedging, you need a more specific claim, not a softer one.

### Examples

| Hedging (penalized) | Direct (rewarded) |
|---|---|
| This may help teams understand deliverability. | Teams that monitor deliverability metrics see 20–30% fewer emails in spam. |
| It could be argued that authentication plays a role. | DKIM, SPF, and DMARC authentication directly determines inbox placement. |
| There are potentially some benefits to stream separation. | Separating transactional and bulk streams protects sender reputation. |

---

## 4. Heading Structure

*LLM ✓ SEO ✓*

### LLM data

Pages with 3–4 headings perform worse than pages with zero headings in every vertical studied. The 3–4 range is a universal dead zone — partial structure confuses AI navigation. Cited text is 2x more likely to contain question marks, and 78.4% of question-based citations come from headings.

**Query match is the #1 content signal for citation.** An analysis of 815,000 query-page pairs found that pages whose best heading directly matches the user's query get cited 41% of the time vs. 29% for loosely related headings. Even among top-retrieved pages, strong heading match adds 19 percentage points to citation rate. This means: the single most important heading on your page is the one that mirrors the target query. Get that one right before optimizing anything else.

### SEO data

Number of heading tags is factor #54 (-0.17). Questions in headings is factor #78 (-0.16). Variations in H1/H2 tags is factor #9 (-0.21). Entities in H2 tags is factor #28 (-0.19). Leading variations in H1 is factor #75 (-0.16).

### The rules

**Make your best heading a direct match to the target query.** This is the highest-impact content optimization for LLM citation. If the target keyword is "best transactional email services," one H2 should be nearly identical: "Best Transactional Email Services for 2026" or "What Are the Best Transactional Email Services?" The AI treats this heading as a query match and the following paragraph as the answer.

**Never use 3–4 headings.** Partial structure is the worst option. Use either minimal headings or enough to fully organize the content.

**Aim for 7–20 subheadings.** An analysis of 815K query-page pairs found that the citation sweet spot is 7–20 subheadings — enough structure to organize without diluting. Beyond 20, additional headings don't improve citation rates. Heading count does not separate always-cited from never-cited pages; heading *quality* (query match) does.

**Use a mix of question and topic-style headings.** Question format works well for explainer and criteria sections ("What affects transactional email deliverability?"). Topic-style is fine for provider names, feature labels, and structural headings ("Pricing," "Mailtrap Email Delivery Platform") — as long as they contain keyword variations or entities.

**Use entity echoing after question headings.** When a heading asks about entity X, the first word of the answer paragraph repeats X. This gives AI a strong semantic match.

**Front-load keywords in H1 and key H2s.** Leading variations is a distinct SEO factor.

### SEO keyword requirements for headings

| Element | Exact match | Variations | LSI terms | Entities | Front-load? |
|---|---|---|---|---|---|
| **H1** | 1 | 0 | 0–1 | 1 | Yes |
| **H2 tags** | 0–1 | 1 per H2 (vary) | 1 per H2 | 1 per H2 | Yes for key H2s |
| **H3 tags** | 0 | 0–1 per H3 | 1 per H3 | 1 per H3 | No requirement |

### Example

> ## What Affects Transactional Email Deliverability?
>
> Transactional email deliverability depends on three factors: authentication protocol enforcement (DKIM, SPF, DMARC), sender reputation management, and whether the provider separates transactional and bulk email streams.

Question heading with keyword variation. First sentence echoes "transactional email deliverability." Direct answer with three named entities.

---

## 5. Entity Density and Types

*LLM ✓ SEO ✓*

### LLM data

Standard text has 5–8% entity density. Heavily cited text has 20.6%. Each entity acts as a verifiable anchor that reduces the model's uncertainty.

### SEO data

Number of distinct entities is factor #7 (-0.21). Entities in title tag is #12 (-0.21). Entities in H2 tags is #28 (-0.19). Entities in sentences is #21 (-0.20).

### Entity type hierarchy (LLM data)

| Entity type | Citation effect | Practical rule |
|---|---|---|
| **DATE** | Positive (universal) | Include publish/update dates. "Updated March 2026." |
| **NUMBER** | Positive (universal) | Use specific counts, metrics, percentages. |
| **CONSUMER_GOOD** (product names) | Positive for SaaS | Name specific tools rather than referring generically. |

### Niche entities beat famous entities

Pages with well-known, Knowledge Graph-verified entities receive fewer citations (0.81x lift). High-cited pages use specific, niche entities: a methodology, a precise statistic, an SDK name, a specific protocol version.

"[Product] uses DKIM, SPF, and DMARC and achieved 78.8% inbox placement in seed testing with shared IPs" **outperforms** "[Product] is a leading email platform trusted by thousands of companies."

### The rules

- Every paragraph should contain at least 2–3 named entities.
- Replace every "many tools" / "some services" / "various options" with actual names.
- Attach numeric data points to entities where possible.
- Include at least one date and one specific number in the intro.
- Use specific, niche entities (SDK names, protocol names, test methodologies) over generic brand authority claims.

---

## 6. Balanced Sentiment (The "Analyst Voice")

*LLM ✓*

### LLM data

Cited text has a subjectivity score of ~0.47 on a 0–1 scale. AI avoids both dry, Wikipedia-style text (0.1) and pure opinion (0.9). It prefers facts combined with applied analysis.

### The formula

> [Specific fact] + [what it means for the reader / how it compares] = citable text.

**Too dry:** "Mailtrap's Business plan costs $85/month."
**Too subjective:** "Mailtrap is hands down the best service out there."
**Analyst voice:** "Mailtrap's Business plan ($85/month) includes a dedicated IP with automatic warmup and deliverability team support — making it the strongest option for high-volume SaaS senders."

---

## 7. Business-Grade Readability

*LLM ✓ SEO ✓*

### LLM data

Cited text scores Flesch-Kincaid grade 16 (college-level). Non-cited text scores 19.1 (academic/PhD level). AI prefers clean subject-verb-object structures.

### SEO data

Flesch-Kincaid reading grade level is factor #74 (-0.16).

### The rules

- Keep sentences under 30 words.
- One idea per sentence.
- Use precise technical terms but keep sentence structures clean.
- Prefer active voice over passive.
- Avoid nominalizations ("the implementation of" → "implement").

| Avoid | Use instead |
|---|---|
| Long sentences with 3+ dependent clauses | One main clause per sentence |
| Passive voice pileups | Active voice with named subjects |
| Nominalizations ("the implementation of") | Verbs ("implement") |
| Jargon stacking without context | Introduce terms, then use them |

---

## 8. Paragraph-Level Structure

*LLM ✓*

### LLM data

53% of AI citations come from the middle of a paragraph. The AI seeks the sentence with the highest "information gain" regardless of position.

### The rules

- Use the first sentence to set up context.
- Place the highest-value factual claim in the middle of the paragraph.
- Use the last sentence for a transition or implication.
- Put the most important paragraphs in the first 20% of the page.
- Within those paragraphs, make the middle sentence the most entity-rich.

### Example

> Inbox placement rates vary significantly across transactional email providers. In independent seed testing using free-tier accounts and shared IPs, Postmark achieved 83.3% inbox placement, Mailtrap reached 78.8%, and Amazon SES scored 77.1% — while SendGrid placed only 61.0% with a 20.9% missing rate. These results suggest that infrastructure quality differs substantially even among industry-leading providers.

Sentence 1 (setup). Sentence 2 (middle — highest information gain, 5 entities + data). Sentence 3 (implication).

---

## 9. Content Depth and Focus

*LLM ✓ SEO ✓*

### LLM data

An analysis of 815,000 query-page pairs found that **focused, shorter pages outperform comprehensive "ultimate guides"** in LLM citation rates. Pages covering 26–50% of ChatGPT's fanout sub-queries get cited more than pages covering 100%. The "ultimate guide" strategy — highest word counts, most headings, broadest subtopic coverage — produces the least reliable citation performance.

The citation sweet spot for word count is **500–2,000 words**. Always-cited pages average ~2,200 words. Pages with the highest word counts fall into the "mixed" category: cited sometimes, ignored other times.

However, SEO still rewards content depth. The tension is real: Google's page 1 average is ~2,000 words, while LLM citation rates peak at shorter, focused pages.

### SEO data

Word count is factor #35 (-0.18). Page 1 average is 1,438 words (unabridged: 2,085).

### The resolution

**Be the best answer to one question, not an adequate answer to twenty.** Every section should directly serve the page's primary query intent. Cut sections that serve a different intent — move them to a separate page and link to it.

This is not about writing less. It's about writing focused. A 2,000-word comparison post with 7 entity-dense provider reviews and a decision map is focused. A 4,000-word comparison post that adds "What is a transactional email service?", "Why does deliverability matter?", and a 1,000-word methodology essay is diluted — and the diluted version gets cited less reliably.

### Target length by content type

| Content type | Target length | Focus principle |
|---|---|---|
| Comparison/roundup | 1,500–2,500 words | Cover providers + decision map. Move criteria deep-dives to separate pages. |
| "What is X" guide | 1,000–2,000 words | Definition + key components + practical application. Don't pad with tangential subtopics. |
| How-to/tutorial | 800–1,500 words | Steps + code. Don't add background explainers the reader didn't ask for. |
| Landing page | 500–1,000 words | Value proposition + differentiators + CTA. Entity-dense, zero filler. |

### The rules

- Don't publish under 500 words for any page you want AI to cite.
- Don't exceed 2,500 words unless every section directly serves the primary query intent.
- If a section serves a different query intent, it belongs on a different page. Link to it.
- Focus beats comprehensiveness: a page that perfectly answers one question outperforms a page that adequately answers ten.

---

## 10. Keyword Placement Map

This is the core SEO reference. It shows where to place each keyword type across page elements:

| Page element | Exact match | Variations | LSI terms | Entities | Key SEO factors |
|---|---|---|---|---|---|
| **Title tag** | 1 (front-loaded) | 0–1 | 0–1 | 1–2 | #12, #26, #30, #63 |
| **Meta description** | 0–1 | 1 | 1–2 | 1–2 | #40, #64, #72, #120 |
| **H1** | 1 (front-loaded) | 0 | 0–1 | 1 | #29, #33, #34, #75 |
| **H2 tags** | 0–1 | 1 per H2 | 1 per H2 | 1 per H2 | #8, #28, #48 |
| **H3 tags** | 0 | 0–1 per H3 | 1 per H3 | 1 per H3 | #42, #100 |
| **First 100 words** | 1 | 1 | 2–3 | 2–3 | #84 |
| **Body throughout** | 2–4 total | 5–8 unique | As many as topic supports | As many as topic supports | #15, #19, #21, #23, #52 |
| **Last 100 words** | 0–1 | 1 | 1–2 | 0–1 | #194 |
| **URL slug** | 1 | 0 | 0 | 0 | #13, #31, #47 |
| **Bold/strong** | 0–1 | 2–3 | 1–2 | First-use bolding | #135 |
| **Link anchors** | 0 | Vary per link | Vary per link | Vary per link | #82, #99 |

---

## 11. Meta Elements

*SEO ✓*

### Title Tag

Entities in title tag (#12, -0.21), LSI words (#26, -0.19), variations (#30, -0.19), leading variations (#63, -0.17), title length (#103, -0.15).

**Rules:** Front-load the primary keyword. Include at least one entity (number, date, brand). Keep 50–60 characters.

**Example:** "7 Best Transactional Email Services Compared [2026]" — exact match front-loaded, number entity, date entity, 52 characters.

### Meta Description

Variations (#40, -0.18), entities (#64, -0.17), LSI words (#72, -0.16), length (#120, -0.15).

**Rules:** Include primary keyword or close variation. Include 1–2 LSI terms. Include at least one entity. Keep 120–160 characters.

**Example:** "We compared 7 transactional email services — Mailtrap, SendGrid, Mailgun, Postmark, and more — on deliverability, pricing, and API integration. See the 2026 test results." (158 characters)

### URL Slug

Variations in display URL (#13, -0.21), document file name (#31, -0.19), URL path (#47, -0.18).

**Rules:** Include primary keyword or tight variation. Keep 3–5 words. Use hyphens.

**Example:** `/blog/transactional-email-services/`

---

## 12. On-Page SEO Elements

*SEO ✓*

### Bold/Strong Tags

Variations in strong tags is factor #135 (-0.14).

**Rules:** Bold key entities and definitions on first use. Bold 2–4 keyword variations across the page. Don't over-bold.

### Internal and External Links

Number of internal links (#82, -0.16). Number of DoFollow external links (#99, -0.15).

**Rules:** Use descriptive anchor text containing keyword variations or LSI terms (not "click here"). Include 5–15 contextual internal links and 2–5 external links per blog post.

**Example:** "See our [email authentication setup guide](https://mailtrap.io/blog/email-authentication/) for step-by-step instructions."

### LSI Keywords

Number of unique LSI words is the **#1 SEO factor overall** (-0.24). Top-ranking pages use an average of 307 unique LSI words.

This is not keyword-stuffing — it's comprehensive topical coverage. If writing about transactional email services, naturally using terms like "bounce rate," "webhook notifications," "email log retention," "template engine," "dedicated IP warmup," and "suppression list" signals topical depth to both search engines and LLMs.

---

## 13. Where LLM and SEO Rules Reinforce Each Other

| Factor | LLM signal | SEO signal |
|---|---|---|
| Front-loading insights | 44.2% of citations from first 30% | Variations in first 100 words (#84) |
| Value-first section order | Snapshot in 10–20% band = maximum citation surface | Variation density in top 30KB (#32); entities early (#21) |
| Entities in headings | Entity echoing lifts citation likelihood | Entities in H2 (#28), H3 (#42), title (#12) |
| Question headings | 78.4% of question-based citations from headings | Questions in headings (#78) |
| Readability | Grade 16 cited vs. 19.1 non-cited | Flesch-Kincaid (#74) |
| Content focus | Focused pages (26–50% subtopic coverage) outperform exhaustive ones | Word count (#35), page 1 avg 1,438 |
| Heading quality over quantity | Query match in best heading = 41% citation rate (vs. 29%) | Number of headings (#54), H2 (#69), H3 (#91) |
| Definitive language | +14% citation lift from declarative intros | Term frequency (#23), exact matches (#158) |
| Keyword variations | Variations increase citation surfaces | #1 factor: unique LSI words (-0.24); variations (#2, -0.24) |

---

## Pre-Publish Checklist (LLM + SEO Only)

### Outline audit (run before writing)
- [ ] Can you state the page's primary query in one sentence?
- [ ] Does every section in the outline directly serve that query? (Run the intent match test on each.)
- [ ] Are there any "What Is X?" sections that should be absorbed into the intro sentence?
- [ ] Are there any post-decision sections (setup, onboarding, DNS config) that should move to the bottom or a separate page?
- [ ] Are there any tangential topics (e.g., full authentication tutorial in a comparison post) that should become one table row + a link?
- [ ] Does the verdict appear within the first 2–3 sections of the outline?
- [ ] For "X vs Y" comparisons: is every section comparative (side-by-side), not descriptive (one product at a time)?

### Structure & page architecture
- [ ] Does the first substantive paragraph (10–20% band) contain the core claim or definition?
- [ ] Does the page follow value-first section order (answer before education)?
- [ ] For comparison posts: does the snapshot/ranked list appear within the first 15% of the page?
- [ ] Is the "how to choose" section compressed into a decision map (not a multi-section essay)?
- [ ] Are definitions handled in the intro sentence, not as standalone sections before the answer?
- [ ] Does every section directly serve the page's primary query intent? (If not, move it to a separate page.)
- [ ] Is the page under 2,500 words? If over, can any section be split into its own page?
- [ ] Does the page have 7–20 subheadings (never just 3–4)?

### Headings
- [ ] Does the page's best heading directly match the target query (near-exact wording)?
- [ ] Do headings use a natural mix of question format and topic-style?
- [ ] Does the first sentence after question headings echo the key entity?
- [ ] Do H2s contain keyword variations (not the same phrase repeated)?
- [ ] Do H3s contain entities and LSI terms?
- [ ] Is the H1 front-loaded with the primary keyword?

### Writing style
- [ ] Does each major section open with an "X is Y" or direct declarative statement?
- [ ] Have all qualifiers been removed from opening paragraphs?
- [ ] Is every fact paired with brief applied analysis (analyst voice)?
- [ ] Is each sentence under 30 words with one main idea?
- [ ] Is the voice mostly active, not passive?

### Keywords & entities
- [ ] Does the exact match keyword appear in the title, H1, and first 100 words?
- [ ] Does the exact match keyword or a close variation appear in the last 100 words?
- [ ] Are 3–5 keyword variations distributed across headings and body?
- [ ] Does the page use a broad range of LSI terms?
- [ ] Does the intro include at least one date and one specific number?
- [ ] Are niche entities (SDK names, protocols, test methods) used over generic brand claims?
- [ ] Does each paragraph contain at least 2–3 named entities?
- [ ] Is the densest sentence in the middle of key paragraphs (not first or last)?

### Meta elements
- [ ] Is the title tag 50–60 characters with the keyword front-loaded?
- [ ] Does the title contain at least one entity (number, date, brand)?
- [ ] Is the meta description 120–160 characters with keyword, LSI term, and entity?
- [ ] Does the URL slug contain the keyword in 3–5 words?

### On-page elements
- [ ] Are 2–4 keyword variations bolded across the page?
- [ ] Are there 5–15 internal links with descriptive anchor text?
- [ ] Are there 2–5 external links to authoritative sources?
- [ ] Do link anchors contain keyword variations or LSI terms?

---

## Data Sources

**LLM Optimization** — Four analyses by Kevin Indig (Growth Memo) using verified ChatGPT citation data from Gauge and AirOps:
1. "The science of how AI pays attention" — 1.2M responses, 18,012 citations.
2. "The science of how AI picks its sources" — 98K citation rows, 7 verticals.
3. "The science of what AI actually rewards" — 5,000+ pages, 7 verticals.
4. "Shorter, Focused Content Wins in ChatGPT" — 815K query-page pairs, 16,851 queries, 353,799 pages, 10 industries. Established focus over comprehensiveness, query match as #1 content signal, 7–20 heading sweet spot, and 500–2,000 word citation sweet spot.

**SEO Factors** — Cora's Top SEO Factors (topseofactors.com), April 2026. 150,365 factors measured. Rankings and correlations reflect the top 200 factors with generated documentation.

---

*This guide covers the machine-optimization layers only: LLM citability and SEO ranking. For voice, tone, jargon, content UX, and the tensions between machine optimization and human readability, see the full Unified Content Writing Guide.*