SEO + GEO Playbook: Rank in Google, Get Cited by AI

16/07/2026

This is the system behind the experiment I posted about: a brand-new site in a niche, no ads, no launch post, no email list, no existing audience. Six days in it had 437 visitors and 856 page views. Google sent 230. Bing and Brave sent most of the rest. And ChatGPT sent 7 people while Claude sent 7 more, which was the part that surprised me: a site less than a week old was already being cited by AI engines that then sent real humans to it.

The numbers are small and it is far too early to call it a trend. The point is not the numbers. The point is the system that produced them, and that system is general. Nothing below depends on the niche I chose. You can run the same playbook on any topic where people have specific, recurring questions.

No hype, no growth hacks. This is the boring version that actually works.

The mental model: two games, one foundation

For years there was one game: rank in Google's list of ten blue links. That game is still on. But there is now a second game running next to it: get named inside the answer when someone asks ChatGPT, Claude, Perplexity or Google's AI overview a question.

• SEO (Search Engine Optimization) gets you into the ranked list of links.
• GEO (Generative Engine Optimization) gets you cited inside the generated answer.

Different output, different rules, but they reward the same underlying thing: being the clearest, most trustworthy, most machine-readable source for one specific question. A ranking algorithm rewards it with a high position. A language model rewards it with a citation. If you build for that one thing, you win both games at once. That is the whole idea. Everything below is how to build for it.

One more thing to internalize: AI search rides classic search. When ChatGPT or Copilot "searches the web," it is largely reading Bing's index. Perplexity and Google's AI overview lean on their own crawls of the open web. So classic indexing is not a separate track from GEO, it is the on-ramp to it. Get indexed fast and cleanly, and you have made yourself eligible to be cited. Stay uncrawlable and no amount of GEO tricks will save you.

Part 1: Pick a niche you can actually win

Most SEO advice skips this and it is the most important part. If you pick a topic owned by big authoritative publishers, the best execution in the world buys you page four. Pick right and mediocre execution ranks.

What you are hunting for is an information gap: a topic where the truth is scattered across forums, out-of-date blog posts, PDFs and official pages nobody reads, and where real people have a specific, recurring question that is annoying to answer today.

Four filters I use:

1. Specific, recurring questions. People should be typing or asking full questions with a real answer, not browsing vaguely. "When does X open," "is Y free," "does Z actually work like people say." Questions have answers, and answers are what both search and AI reward.
2. High intent over high volume. A keyword with 500 searches a month from people who need an answer now beats 50,000 from people killing time. Time-sensitive and decision-driving topics convert attention into action.
3. The primary-source test. Can you assemble the canonical, verified answer that does not currently exist cleanly in one place? AI engines need a clean, trustworthy source to cite. If you become that source, you are the obvious citation. If the clean source already exists and it is a big brand, walk away.
4. Underserved by authority. The niches the major publishers ignore because they are too small, too fiddly, or too much manual work are exactly the ones you can own. Manual verification work is a moat: it is expensive for a big site and impossible for an AI to fabricate.

If a topic passes all four, you have a site worth building. If it fails the primary-source test, keep looking, because you will be fighting for scraps of someone else's authority.

Part 2: Content architecture: one verifiable fact per page

This is the core of the whole system, and it is what makes a page equally good for a ranking algorithm and a language model.

Build atomic answer pages

Each page should answer one specific question completely, and answer it in a way a stranger (or a machine) can verify. Not "everything about topic X" in one giant page. One question, one clean, self-contained answer, one URL. This is how people search and it is how assistants retrieve.

Lead with the answer, then support it

Put the direct answer in the first sentence, in plain language, then give the detail, the caveats and the source underneath. Do not bury the answer under 600 words of preamble. Language models lift passages: they take a self-contained factual sentence and quote it. Make the liftable sentence the first thing on the page.

Compare:

• Weak: "There are many things to consider when thinking about whether this event uses a lottery system, and in this article we will explore..."
• Strong: "This event does not run a public ballot. Tickets are sold first-come-first-served on the official site from 1 September. Here is how that works and why people assume otherwise."

The strong version is the one that gets quoted in an answer and the one that earns the click.

Show your work on every page

Cite the primary or official source, and stamp a visible "verified on [date]" on the page. This does three jobs at once:

• It is the E-E-A-T signal Google's quality systems look for (Experience, Expertise, Authoritativeness, Trust).
• It makes a citation safe for an AI engine. Models are penalized internally for citing wrong things, so they prefer sources that visibly reduce that risk.
• It makes you trustworthy to the human who lands there.

A visible verification date is one of the highest-leverage, lowest-effort things you can add. Almost nobody does it. Do it.

Design for freshness, not one-time correctness

Time-sensitive facts go stale silently, and a stale page loses its citation to a fresher competitor. So do not hard-code a status that will rot. Store the underlying facts (the dates, the numbers) and derive the human-facing status at render time, so "opens in 3 days" or "closed" is always computed from today's date, never from whatever was true the day you published. On the experiment site every page recomputes its status hourly. The reader always sees the truth, and the freshness itself is a ranking and citation signal.

Structure it as hub-and-spoke, richly linked

• Hubs: category pages that list and link to every detail page in that category.
• Spokes: the atomic answer pages.
• Guides: longer explainer content that links down into the relevant spokes.

Cross-link generously and with descriptive anchor text. Internal links are how crawlers discover your pages, how ranking authority flows between them, and how you keep a reader (or an assistant following links) moving through your site instead of bouncing.

Part 3: The SEO layer: get found and ranked

With the architecture right, the SEO layer is mostly discipline, not cleverness.

Speed and static delivery. Pre-render pages and ship them as static HTML wherever you can. Minimal JavaScript, fast server response, images sized properly. Core Web Vitals are a real ranking input and, more importantly, a slow site loses people before they read the answer. Static generation also means every page is fully there in the HTML for a crawler on first request, with nothing hidden behind client-side rendering.

Semantic HTML plus structured data. Use real headings, lists and landmarks, then add JSON-LD structured data that tells engines what each page is: an Event, an Article, an ItemList for a hub, Organization and WebSite sitewide, BreadcrumbList for navigation. Match the schema type to the page type. This is machine-readable meaning, and both search engines and AI crawlers consume it.

Do not over-mark-up. More schema is not better. Google restricts some rich results (FAQ rich snippets, for example, are now limited to a narrow set of sites), and marking up things it will ignore is wasted effort at best. Mark up what maps to a real result or a real understanding of the page, and stop there.

Metadata discipline. Every page gets a unique, descriptive title and meta description written for a human deciding whether to click, a correct canonical URL so you never split your own ranking signals across duplicates, and its own share/preview image. Boring, mechanical, and most sites still get it wrong.

Crawlability. A clean sitemap.xml listing every real page, a sane robots.txt that does not accidentally block what you want indexed, and a logical internal link graph so nothing is an orphan. Submit the sitemap in Google Search Console and Bing Webmaster Tools on day one.

Fast indexing: the sleeper lever. This is the single biggest reason a six-day-old site was already getting traffic. Use IndexNow to ping Bing, Yandex and their partner engines on every deploy, so new and changed pages get indexed in minutes instead of waiting weeks for an organic crawl. It is a tiny one-time setup (a key file on your domain plus a POST on deploy) and it changes your time-to-visibility from weeks to minutes. And because ChatGPT's web search rides Bing's index, getting into Bing fast via IndexNow is also what puts you on the AI-citation surface fast. One lever, both games.

Part 4: The GEO layer: get cited by AI engines

Everything above already makes you more citable, because clean, fast, well-structured, trustworthy content is what models want to cite. The GEO layer is the extra work aimed specifically at showing up inside AI answers.

Do not block the AI crawlers. This is the number one own-goal. If you want to be cited by AI engines, your robots.txt must allow their bots: GPTBot and OAI-SearchBot (OpenAI), ClaudeBot and anthropic-ai (Anthropic), PerplexityBot, Google-Extended, and others. Plenty of sites block these by reflex or by default template and then wonder why they never get cited. Check this first.

Publish an llms.txt. Add a plain-text /llms.txt file at your root: a clean map of the site written for AI crawlers. State what the site is and who it is for in one line, list your most useful pages with a short description of each, and include your freshest facts (what is live/open/current right now) with a visible verification date. Regenerate it so the live parts stay fresh. Think of it as a hand-drawn map that tells a model exactly what you offer and which page answers which question, without making it reverse-engineer your navigation. It is cheap to produce and almost nobody has one yet.

Be citable at the passage level. Models quote self-contained sentences. Write factual, standalone statements that make sense lifted out of the page with no surrounding context. Short sentences, concrete facts, no hype words, no "as we discussed above." Every strong sentence is a potential citation.

Be the primary source. The most durable GEO moat is original, verified data an engine cannot get anywhere else. Aggregation plus human verification is exactly the kind of work an AI cannot fabricate and a big publisher will not bother to do. If you are the only clean source for a fact, you are the citation by default.

Answer the questions people actually ask an assistant. Structure content around real spoken questions: "when does X open," "is Y free," "does Z run a lottery." Negative-answer and myth-busting pages are especially valuable, because people constantly ask assistants to confirm or debunk something ("is it true that..."), and a clean, sourced "no, here is what actually happens" page is precisely what the assistant needs and rarely finds.

Trust signals reduce the model's risk. State clearly who you are, what you are and are not affiliated with, where your data comes from, and when it was last checked. A model choosing between two sources will lean toward the one that visibly lowers its chance of being wrong.

Part 5: Measure both games

You cannot improve what you cannot see, and most people have no idea AI is already sending them traffic because they never look.

Watch your referrers for AI engines. In your analytics (Vercel Web Analytics, Plausible, GA4, whatever), look at the referrer/source list for chatgpt.com, claude.ai, perplexity.ai, copilot.microsoft.com and similar. That is your GEO scoreboard. The day you see them appear is the day you know the second game is live for you. In the experiment those referrers showed up within the first week.

Watch Search Console for the classic game. Impressions, top queries, top pages, and click-through rate. Early on, impressions rising means you are getting indexed and shown; clicks will follow as positions improve.

Early on, only two numbers matter:

1. Are you getting indexed (coverage in GSC/Bing, pages showing impressions).
2. Are you getting cited (AI referrers appearing at all).

Rankings, volume and conversion come later. In week one, indexed-and-cited is the entire game, and it is very achievable.

Part 6: The honest part

Being straight about this matters more than making it sound impressive.

• This is not a hack and not overnight. It is a system, and systems compound slowly and then quickly.
• A week is not a trend. Small numbers off a base of zero produce silly percentages. Ignore the percentages.
• The compounding is the point. Every atomic page you add is another question you can answer, another URL that can rank, another passage that can be cited. Every deploy re-pings IndexNow and refreshes your llms.txt. The system is built so each addition makes the next citation more likely.
• What I will be watching next: whether AI referrals grow as a share of traffic, whether cited pages hold their citations as competitors appear, and whether the primary-source moat actually holds up over months.

If you do the boring version of all of this well, you end up with a site that answers real questions better than anyone else, loads fast, gets indexed in minutes, and is trusted by both ranking algorithms and language models. That is the whole playbook. The rest is patience.

The one-page checklist

Before you build

• Found a real information gap with specific, recurring questions
• High intent beats high volume; time-sensitive if possible
• I can become the clean primary source (it does not already exist)
• Underserved by big authoritative publishers

Content architecture

• One question per page, answered completely and verifiably
• Direct answer in the first sentence, detail underneath
• Primary source cited on every page
• Visible "verified on [date]" on every page
• Time-sensitive facts derived at render time, never hard-coded to rot
• Hub-and-spoke structure, richly cross-linked with descriptive anchors

SEO layer

• Static/pre-rendered, fast, good Core Web Vitals
• Semantic HTML plus JSON-LD matched to each page type
• Not over-marked-up (only schema that maps to a real result)
• Unique title + description + canonical + preview image per page
• Clean sitemap.xml and robots.txt; submitted to Google + Bing
• IndexNow ping on every deploy (minutes-to-index, and feeds AI search)

GEO layer

• AI crawlers allowed in robots.txt (GPTBot, ClaudeBot, PerplexityBot, etc.)
• /llms.txt published, with freshest facts and a verification date
• Content written as liftable, self-contained factual passages
• Original verified data that engines cannot get elsewhere
• Pages built around real spoken questions, including myth-busters
• Clear identity, affiliation and sourcing on the site

Measurement

• Analytics watched for AI referrers (chatgpt.com, claude.ai, perplexity.ai)
• Search Console + Bing Webmaster for impressions, queries, coverage
• Judging week one on two things only: indexed, and cited

This is the general system, not tied to any one niche. If you run it on your own topic and it works, I would genuinely like to hear about it.