The SEO + GEO Content Engine

The 2026 reality: being cited is the new ranking

The old contract was simple: rank top 3, get the click. That contract is being renegotiated without you in the room. As of 2026, roughly 48% of Google searches trigger an AI Overview, and the AI engines (ChatGPT search, Perplexity, Gemini, AI Mode) answer the question directly, citing a handful of sources. If you're one of the citations, you get the referral and the authority. If you're not, you're invisible even at position 1.

Here's the number that should reorganize your content strategy: Ahrefs analyzed 863k keywords and 4M AI Overview citations and found only 38% of AIO-cited pages rank in the top 10 for the query. Seven months earlier it was 76%. About 31% of citations come from positions 11-100, and another 31% from outside the top 100 entirely. Ranking and getting cited have decoupled.

The mechanism to understand is query fan-out. Google's AI splits one query into multiple sub-queries, runs them all, and cites the pages that show up across the most sub-query result sets. Practical translation: don't write one page for the head term. Cover the full sub-question cluster (cost, timeline, alternatives, risks, "vs", "for beginners") on the page or in tightly linked posts. The content engine below is built around exactly this.

One more thing the tool vendors won't tell you: the engines diverge hard. ChatGPT's live search runs on Bing, so Bing indexing is a hard prerequisite. Perplexity leans on Reddit for up to 1 in 5 of its citations. AI Overviews reward fan-out coverage. Optimizing for "AI search" as one thing is how you optimize for none of them.

The GEO checklist: what actually earns citations

Everything below is backed by published data, mostly the Princeton/Georgia Tech/IIT Delhi GEO study (KDD 2024, still the benchmark) and Ahrefs' 2026 citation studies. The two biggest levers are boring and repeatable: sourced statistics (up to ~40% visibility lift) and attributed quotes (~30-40% lift). The full expanded version is in the download at the end of this section.

Structure: write for passage-level retrieval

LLMs retrieve chunks, not pages. Every section has to survive being ripped out of context and quoted alone.

• [ ]Answer-first blocks: every H2 opens with a direct 40-80 word answer that stands alone with zero surrounding context, then depth.
• [ ]Question- or claim-phrased H2s: "How much does X cost in 2026" beats "Pricing considerations". Reading only H2s should deliver the argument.
• [ ]Self-contained sections: no chunk depends on the previous one. No "as mentioned above" logic.
• [ ]Tables for comparisons, bullets for lists of 3+: the most extractable formats.
• [ ]3+ quotable one-liners per post: self-contained, opinionated, one sentence, written to be quoted verbatim.
• [ ]Fan-out coverage: list the 5-10 sub-questions around the topic and answer each on the page or in a linked cluster.
• [ ]FAQ section: question-phrased H3s with 40-80 word answers for the sub-questions that didn't earn a full H2.

Evidence: the two strongest levers in the data

• [ ]Statistics with named source + year, inline: the strongest single GEO tactic tested (up to ~40% lift, Princeton study). Format: specific number, named source, year.
• [ ]Attributed expert quotes: ~30-40% lift. Real people, real attribution.
• [ ]Cite your sources: content that cites credible sources gets cited more itself.
• [ ]Zero unsourced numbers: every figure traces to a source or your own test, or it gets cut.
• [ ]Original data annually: first-party data is the only content AI can't get elsewhere. Publish one original study per year and become the stat everyone else cites.

Freshness: the 30-day cadence

• [ ]Update money pages every 30 days or less: ~76% of ChatGPT citations come from content updated in the last 30 days.
• [ ]Visible + machine-readable dateModified: on-page "Updated [date]" and dateModified in Article schema.
• [ ]Honest updates only: change a stat, a price, or a section. Bumping the date without changing content is a trust asset you burn exactly once.
• [ ]Kill stale numbers on the monthly pass: anything no longer true gets updated or removed.

Technical: table stakes, do once

• [ ]robots.txt allows AI crawlers: GPTBot + OAI-SearchBot (ChatGPT), PerplexityBot, Google-Extended (Gemini grounding), ClaudeBot. AI Overviews use normal Googlebot.
• [ ]Indexed in Bing: ChatGPT's live search runs on Bing. No Bing index, no ChatGPT citations. Bing Webmaster Tools + IndexNow.
• [ ]Server-side render key content: many AI crawlers execute little or no JS. Content that only exists after hydration doesn't exist.
• [ ]Fast TTFB and clean heading hierarchy: crawlers with budgets skip slow, messy pages.
• [ ]Schema for SERP features only: keep Article, Organization, FAQPage where eligible. It is not a citation lever (see the warning below).

Entity clarity and off-site surfaces

• [ ]One consistent name everywhere: site, socials, directories. Engines resolve entities from consistency.
• [ ]Organization schema on the homepage with name, url, logo, and sameAs pointing at your real profiles.
• [ ]A plain-language about page: what you are, who runs it, why you're credible.
• [ ]Authentic Reddit presence: Reddit is the #1 cited source across AI engines overall, up to ~1 in 5 Perplexity citations. Answer with substance, disclose affiliation. Astroturfing gets domains nuked.
• [ ]YouTube versions of your best content: YouTube is the single most-cited domain in AI Overviews (~18% of outside-top-100 citations).
• [ ]LinkedIn distribution of your original data and strongest claims.

The content engine: 4 prompts from seed keyword to linked post

This is the production line. One seed keyword goes in; a published, internally linked, GEO-structured post comes out. Four prompts, run in order, one conversation each so context stays clean. The human steps in between are not optional: you verify the stats, you do the edit pass, you place the links.

1. 1

   Prompt 1: seed to cluster

   One seed keyword becomes 20 intent-grouped topics, including the 5 fan-out satellites that AI Overviews reward. Pick your P1s.

2. 2

   Prompt 2: topic to outline

   Answer-first structure with drafted 40-80 word answer blocks, evidence slots, and quotable lines already placed. You then fill every NEEDS STAT flag with a real source.

3. 3

   Prompt 3: outline to draft

   Full draft under voice guardrails and a no-fluff rule, plus a fact ledger of every claim for you to verify. Then your human edit pass: screenshots, your numbers, your jokes.

4. 4

   Prompt 4: draft to linked post

   Maps the new post into existing content: inbound links, outbound links, hub check, orphan check, cannibalization check. Place the inbound links the day you publish.

You are my SEO strategist. Turn one seed topic into a content cluster built for both Google rankings and AI-engine citations.

SEED TOPIC: [SEED_KEYWORD]
MY SITE: [DOMAIN], which sells [PRODUCT_ONE_LINER] to [AUDIENCE]
MARKET/LANGUAGE: [MARKET]
EXISTING POSTS ON THIS TOPIC (so you don't duplicate): [LIST_OR_"none"]

Generate exactly 20 topic ideas, grouped by search intent:
- 6 informational: what/how/why questions my audience asks before they know solutions exist
- 5 commercial investigation: "best X", "X vs Y", "X alternatives", "X pricing"
- 4 transactional: queries where the searcher is ready to act; map each to a landing or comparison page
- 5 fan-out satellites: the sub-questions Google's AI systems would split the seed topic into (definitions, costs, timelines, risks, comparisons, "for beginners"). These exist so my cluster appears across the most sub-query result sets.

For every topic, output one table row:
| # | Working title (question-phrased where natural) | Intent group | Primary query | 2-3 secondary queries | Funnel stage | The one thing this page must contain that no competitor page has (unique data, original test, first-party numbers) | Priority P1/P2/P3 (business value x winnability) |

Rules:
- No two topics may answer the same underlying question. If two overlap, merge them and generate a replacement.
- Commercial topics must name real category rivals, not placeholders.
- Flag any topic that only makes sense as a programmatic page set, and state what unique per-page data it would need to survive Google's scaled-content policies. If no such data exists, mark it DO NOT BUILD.
- End with: (a) a publish order for the first 5 posts and one sentence on why, (b) which single post should become the cluster hub.

The quality gate is the "one thing no competitor has" column. If you can't fill it for a topic, that post will be page-2 commodity content. Deprioritize it or find the asset first.

Build an answer-first, GEO-structured outline for one post from my content cluster.

POST TOPIC: [WORKING_TITLE]
PRIMARY QUERY: [PRIMARY_QUERY]
SUB-QUESTIONS TO COVER: [PASTE_FAN_OUT_ROWS or write "generate the 6-10 sub-questions a searcher would also ask"]
MY UNIQUE ASSET: [WHAT_I_HAVE_THAT_COMPETITORS_DONT: data, test results, screenshots, costs I paid, experience]
AUDIENCE AND WHAT THEY ALREADY KNOW: [DESCRIPTION]

Output this exact structure:

1. TITLE: 3 options, under 60 characters, primary query near the front, no clickbait.

2. DIRECT ANSWER BLOCK: a 40-80 word standalone answer to the primary query that an AI engine could quote verbatim with zero surrounding context. Write the actual text now, not a description of it.

3. H2 SECTIONS: 5-8 H2s, each phrased as a question or a claim so a skimmer reading only H2s gets the full argument. Under each H2:
- the 40-80 word direct answer that opens the section (write it now)
- supporting points as bullets
- EVIDENCE SLOT: which statistic, named source, or expert quote goes here; write "NEEDS STAT: [description of the number to find]" if one must be sourced
- FORMAT NOTE: table / numbered steps / bullets / prose, chosen for extractability, with one line on why

4. QUOTABLE LINES: 3 one-sentence, self-contained, opinionated lines, each assigned to a specific section, written so a human would forward them and an AI engine would quote them verbatim.

5. FAQ SECTION: 4-6 question-phrased H3s covering the sub-questions not already used as H2s, each with a drafted 40-80 word answer.

6. METADATA: meta description under 155 characters, URL slug, and a note on what should be re-checked at the 30-day update.

Constraints:
- Every number in the outline has a named source or a NEEDS STAT flag. Nothing unsourced survives to draft.
- No section exists for word count. If an H2 adds no new answer, cut it and say you cut it.
- No "What is [broad category]" section unless the primary query literally asks that.

Write the full draft from the outline below. You are drafting in MY voice, not yours.

OUTLINE: [PASTE_FULL_OUTLINE_WITH_STATS_FILLED_IN]
VOICE SAMPLE: [PASTE_300_WORDS_OF_YOUR_BEST_WRITING]
FIRST-PERSON FACTS I CAN CLAIM: [WHAT_YOU_ACTUALLY_DID_TESTED_PAID: be specific; this is the only experience the draft may state in first person]
TARGET LENGTH: [WORD_COUNT] words. Hard cap. Coming in under is fine.

VOICE GUARDRAILS (binding):
- Match the sentence rhythm and vocabulary of the voice sample. When unsure, shorter.
- First person only for the facts I listed above. Never invent experiences, tests, costs, or results.
- Numbers over adjectives: "cut build time from 4 hours to 20 minutes" not "dramatically faster". No number, no superlative.
- Banned outright: "delve", "leverage" as a verb, "game-changer", "revolutionary", "unleash", "in today's fast-paced world", "it's important to note", "in conclusion", em-dashes.
- Every statistic keeps its named source and year inline, like "(Ahrefs, 2026)".

NO-FLUFF RULE (enforce ruthlessly):
- Every H2 section opens with its 40-80 word direct answer from the outline, then depth. Zero warm-up sentences.
- Delete any sentence that restates the previous sentence, previews what's coming, or summarizes what was just said. One idea, once.
- Paragraphs are 1-3 sentences. Lists of 3+ become bullets. Comparisons of 2+ things on 2+ attributes become tables.
- If a section runs short because there is nothing more true to say, leave it short.

OUTPUT:
1. The full draft in markdown, including the quotable lines from the outline placed in their sections.
2. FACT LEDGER at the bottom: a table of every number and factual claim in the draft, its source, and a CHECK flag on anything you were not explicitly given a source for.

The fact ledger is what keeps the machine honest. Verify every CHECK flag before publishing; cut what you can't verify. The draft is a strong first 85%. The last 15% (your screenshots, your costs, your one-liners) is the part competitors can't generate.

Map this new post into my existing content so both readers and crawlers find it.

NEW POST: [TITLE + URL_SLUG]
NEW POST SUMMARY: [3_SENTENCES_OR_PASTE_THE_DRAFT]
EXISTING CONTENT: [PASTE_LIST_OF_TITLES_AND_URLS or your sitemap URLs]
MONEY PAGES: [THE_1-3_PAGES_THAT_MAKE_MONEY: product, pricing, signup]

Output three tables:

TABLE 1: INBOUND (links FROM existing posts TO the new post)
| Existing post | The exact sentence or section where the link fits naturally | Anchor text (descriptive, varied, not "click here") | Why a reader would follow it |
Find 3-8. Only propose links a reader would actually want to click. If fewer than 3 natural spots exist, say so instead of forcing them.

TABLE 2: OUTBOUND (links FROM the new post TO existing content)
| Section of new post | Target post | Anchor text | Purpose (context / next step / money page) |
2-5 links. At most ONE link to a money page, placed where intent is highest, never in the intro.

TABLE 3: CLUSTER HEALTH
- Hub check: which existing post is the hub for this topic? If none, which post should be promoted to hub, and what would that take?
- Orphan check: after these links, list any post in this cluster with fewer than 2 internal inbound links.
- Cannibalization check: does the new post target the same primary query as any existing post? If yes, name it and recommend one of: merge, 301 redirect, or differentiate (and state the differentiated angle).

Rules: exact-match anchor text on at most 1 inbound link; vary the rest. Never propose a second link from one page to the same target.

Place the inbound links (yes, editing the old posts) the same day you publish. A new post with zero internal inbound links is a launch with the parking brake on: crawlers find it late and AI engines not at all.

The programmatic trap: 50 great pages beat 500 generated ones

Three patterns Google is specifically penalizing: mass AI page generation without editorial review, pure template-plus-variable substitution at scale, and aggregators adding nothing beyond scraped source data. If your plan matches any of these, it's not a plan, it's a countdown.

The quality gate that settles arguments is the bookmark test: would someone bookmark or share this specific page? If no, revise or remove. Pruning plus enriching is also the documented recovery path for sites that already got hit.

• [ ]Every page in the set answers a distinct query no other page on the site answers
• [ ]60%+ unique content per page, 3+ data sources per page
• [ ]A human enrichment layer before publish: original data, screenshots, commentary, comparisons
• [ ]Demand-gated: proven search demand AND quality data for every page you generate
• [ ]A quarterly prune: zero-traffic, zero-citation pages get merged or deleted
• [ ]Every page passes the bookmark test

Measuring: rank tracking plus AI-citation spot checks

Keep your normal rank tracking; that game still pays. Add two things: AI citation tracking and referral attribution. For referrals, check GA4 for traffic from chatgpt.com and perplexity.ai; those sessions are your proof that citations convert to visits. For citation tracking, the July 2026 tool market splits cleanly by budget:

Under $500/mo total marketing budget? Skip all of them and run the manual audit below. It costs 30 minutes a month and tells you 80% of what a dashboard would.

The 10-prompt manual audit (monthly, 30 minutes, $0)

1. 1

   Write your 10 money questions

   The questions a buying customer would ask an AI: "best [your category] for [your audience]", "[your product] vs [rival]", "how do I [problem you solve]", "is [rival] worth it". Fix the list; you're building a time series, so don't reword them monthly.

2. 2

   Ask ChatGPT (with search) and Perplexity, same 10 prompts each

   Fresh chats, no custom instructions. 20 answers total. Add Google AI Overviews for your top 3 queries if you have the patience.

3. 3

   Log every answer

   For each prompt: were you cited (Y/N), which page, which competitors were cited, and the sentiment of any mention of you. Use the log template below.

4. 4

   Act on the gaps

   Not cited but a competitor is? Read their cited page and note what it has that yours lacks (usually: fresher date, a sourced stat, a more direct answer block). Cited via an old post? That page just became a money page; put it on the 30-day update cadence.

5. 5

   Compare month over month

   Citation count per engine is your KPI. Movement of 2-3 citations on 10 prompts is signal at this scale; a competitor showing up on 6 of 10 prompts is a fire alarm.

# AI Citation Log: [MONTH YEAR]

| # | Prompt | Engine | Cited? | Our page cited | Competitors cited | Sentiment | Note |
|---|--------|--------|--------|----------------|-------------------|-----------|------|
| 1 | best [category] for [audience] | ChatGPT | N | - | rival.com/best-x | - | their page updated 6 days ago |
| 1 | best [category] for [audience] | Perplexity | Y | /blog/x-guide | rival.com | neutral | cited our stat |
| 2 | [product] vs [rival] | ChatGPT | ... | | | | |

Score: __/20 answers cited us (last month: __/20)
Actions: 1) ... 2) ... 3) ...

Deep dives: schema starter and the truth about llms.txt

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Article",
      "@id": "https://yournewsletter.com/blog/your-post-slug#article",
      "headline": "Your Post Title Under 110 Characters",
      "description": "The meta description, under 155 characters.",
      "datePublished": "2026-05-14",
      "dateModified": "2026-07-08",
      "author": {
        "@type": "Person",
        "name": "Jane Doe",
        "url": "https://yournewsletter.com/about",
        "sameAs": ["https://www.linkedin.com/in/janedoe"]
      },
      "publisher": { "@id": "https://yournewsletter.com#org" },
      "mainEntityOfPage": "https://yournewsletter.com/blog/your-post-slug",
      "image": "https://yournewsletter.com/images/your-post-og.png"
    },
    {
      "@type": "Organization",
      "@id": "https://yournewsletter.com#org",
      "name": "Your Brand",
      "url": "https://yournewsletter.com",
      "logo": "https://yournewsletter.com/logo.png",
      "sameAs": [
        "https://www.linkedin.com/company/yourbrand",
        "https://www.youtube.com/@yourbrand",
        "https://x.com/yourbrand"
      ]
    }
  ]
}

# Your Product

> One-sentence description of what the product does and for whom.

## Docs

- [Quickstart](https://yourproduct.com/docs/quickstart): install to first result in 5 minutes
- [API reference](https://yourproduct.com/docs/api): endpoints, auth, rate limits
- [Examples](https://yourproduct.com/docs/examples): copy-paste starter code

## Optional

- [Changelog](https://yourproduct.com/changelog): what shipped, dated

Start here: the first week

1. 1

   Day 1: technical pass (2 hours)

   robots.txt allows the AI crawlers, Bing Webmaster Tools set up, key content server-side rendered. This unlocks everything else.

2. 2

   Day 2: run the GEO checklist on your top 3 pages

   Download above. Most pages fail on two things: no direct answer block and unsourced numbers. Fix those first; they're also the two biggest measured levers.

3. 3

   Day 3: run your first 10-prompt audit

   You need the baseline before the engine runs, or you'll never know if any of this worked.

4. 4

   Day 4-5: first post through the engine

   Prompt 1 with your best seed keyword, then prompts 2-4 on the highest-priority topic. Verify the fact ledger. Publish with inbound links placed.

5. 5

   Every 30 days after

   One update pass on money pages, one audit, 1-2 new posts through the engine. Boring, repeatable, compounding. That's the point.

The whole play in one line: structured answers, sourced numbers, fresh dates, at a volume you can keep honest. The engines change monthly; that sentence hasn't changed since the Princeton study and probably won't.