Generative engine optimization (GEO): the practical guide

Generative engine optimization (GEO) is the practice of making your website more likely to be cited, quoted, and recommended by AI engines — ChatGPT, Perplexity, Claude, Gemini, and Google's AI Overviews — when they answer questions your customers ask. You'll also see it called answer engine optimization (AEO) or LLM SEO. Different labels, same discipline: traditional SEO earns you a ranking; GEO earns you a citation inside an AI-generated answer.

This guide is the practical version: how AI engines actually pick their sources, what to change on your pages, the technical plumbing (crawlers, schema, llms.txt), and how to measure whether any of it is working. It's the same playbook our agent runs as part of its weekly loop — we've built GEO into Cliff's AI search work from day one, because a growing share of our customers' buyers now start in a chat window instead of a search box.

How AI engines choose what to cite

Under the hood, most AI answers are built the same way: the engine takes the user's question, runs searches against an index (its own, or Bing's, or Google's), retrieves a set of pages, and then composes an answer — citing the passages it leaned on. Three practical consequences fall out of that:

1. Classic SEO still gates everything. If you don't appear in the retrieval step, you can't be cited. Pages that rank well in ordinary search are dramatically more likely to show up in AI answers. GEO is a layer on top of SEO, not a replacement for it.
2. Engines cite passages, not pages. The model isn't rewarding your domain authority when it quotes you; it's rewarding one extractable chunk of text that cleanly answered the question. That changes how you write (next section).
3. Being retrievable is a technical property. If AI crawlers are blocked, rate-limited, or served a JavaScript shell with no content, you're invisible to that engine regardless of how good the content is.

Passage-level citability: write chunks worth quoting

The single highest-leverage GEO change is structural: make each section of a page a self-contained, liftable answer. Concretely:

• Answer first, elaborate second. Open each section with a 40–60 word direct answer to the question the heading poses. That's the quotable unit. Detail, caveats, and examples go below it.
• Make headings questions or claims, not puns. "How much does a robots.txt mistake cost?" gives an engine a hook. "Crawl walk run" gives it nothing.
• One idea per section. Retrieval works on chunks. A section that wanders across three topics matches no query well.
• Use lists and tables for anything enumerable. Engines lift structured comparisons far more readily than paragraphs that bury the same facts.
• Define your terms. A crisp definition sentence ("X is Y that does Z") is the most-cited sentence shape on the web right now. If you own a category term, write its definition better than anyone.
• Put real names on things. Author bylines, bios, dates, and an organization behind the content all feed the credibility signals engines use when choosing between near-equal sources.

A useful self-test: paste your page into any chatbot and ask "answer [target question] using only this page, and quote the sentence you'd cite." If the model struggles to find a clean quote, so will the engines.

The technical layer: let the engines in

Know the crawlers by name

AI engines send distinct, documented crawlers, and many sites block them without realizing it — old robots.txt templates and CDN bot rules are the usual culprits. The ones that matter most:

• OpenAI operates three: GPTBot (training), OAI-SearchBot (search indexing for ChatGPT search), and ChatGPT-User (fetches a page live when a user asks about it). Details in OpenAI's crawler documentation.
• Anthropic operates ClaudeBot (crawling), plus Claude-User and Claude-SearchBot for user-initiated and search retrieval — see Anthropic's crawler documentation.
• Perplexity operates PerplexityBot (indexing) and Perplexity-User (live retrieval) — documented at docs.perplexity.ai.
• Google's AI Overviews use ordinary Googlebot — if you're indexed in Google, you're eligible. Google-Extended only controls AI training, not AI Overviews.

Decide deliberately which you allow. Blocking training bots while allowing search/user-fetch bots is a defensible middle position; blocking everything means conceding the answer box to competitors. Check your robots.txt and your CDN's bot-management settings — the CDN is where well-meaning "block AI bots" toggles silently kill your citations.

Serve real HTML

Most AI crawlers execute little or no JavaScript. If your content only exists after a client-side render, engines see an empty shell. Server-rendered or static HTML wins by default here.

Schema still earns its keep

Structured data doesn't guarantee citations, but it makes your page easier to interpret with confidence: Article with real author and dates, FAQPage where you genuinely answer questions, Organization and Person to anchor who is speaking, Product/Offer where prices matter. Engines composing an answer about pricing cite pages whose prices are machine-readable.

llms.txt, briefly and honestly

llms.txt is a proposed standard: a markdown file at your site root that gives language models a curated map of your most important pages. It costs ten minutes and can't hurt. But be honest about its status — adoption by the major engines is still uneven, so treat it as a cheap bet, not a strategy. Crawler access and citable content matter far more.

Measuring AI visibility

You can't screenshot your way to a strategy, but AI visibility is now genuinely measurable:

• Referral traffic. Segment analytics referrers for chatgpt.com, perplexity.ai, gemini.google.com, and friends. This traffic tends to be small but absurdly high-intent — the AI already qualified the visitor.
• Prompt tracking. Take the 20–50 questions a real buyer would ask ("best X for Y", "is [your product] worth it", "[category] pricing"), run them across the engines on a schedule, and record whether you're mentioned or cited. Tools exist for this, or an agent can do it as part of its weekly loop — Cliff tracks a prompt set per site and reports mention changes alongside rankings.
• Crawler logs. GPTBot, ClaudeBot, and PerplexityBot showing up in your server logs is the leading indicator; citations lag crawling.

Treat AI mentions the way you treat rankings: a trendline to move over months, not a switch to flip.

A working GEO checklist

1. Audit robots.txt and CDN rules for AI crawler blocks; allow deliberately.
2. Confirm your content is present in raw HTML, not rendered client-side.
3. Rewrite your top 10 pages so every section leads with a 40–60 word direct answer.
4. Add question-shaped headings, comparison tables, and definitions where they're honest.
5. Ship Article, FAQPage, Organization, and Person schema with real names and dates.
6. Publish an llms.txt as a low-cost extra.
7. Build a buyer-question prompt set and check it monthly across ChatGPT, Perplexity, and AI Overviews.
8. Keep doing real SEO — retrieval still runs on search. Our take on the sane division of labor between humans and AI is in How to use AI for SEO.

None of this is exotic — which is exactly the point. GEO in 2026 is mostly disciplined writing plus unglamorous technical hygiene, applied every week. That's also why we fold it into an agent's routine rather than selling it as a separate project: the checklist above is repetitive, checkable work, and repetitive checkable work is what an AI SEO employee is for. If you want the deeper product version of this — AI visibility tracking, citability fixes, crawler access as a standing weekly job — that's what Cliff's AI search optimization does.

FAQ

Is GEO different from SEO? It's a layer on top. SEO gets you retrieved; GEO gets you quoted. Roughly 80% of the work overlaps — the GEO-specific 20% is passage-level citability, AI crawler access, and measuring mentions instead of only rankings.

Do AI Overviews use the same signals as Google rankings? Largely yes — AI Overviews are built from Google's index via ordinary Googlebot crawling, and pages that rank well are far more likely to be included. There is no separate "AI Overviews crawler" to optimize for.

Should I block AI crawlers to protect my content? That's a business decision, not an SEO one. Just make it deliberately: blocking GPTBot keeps you out of training data, but blocking OAI-SearchBot or PerplexityBot removes you from answers your buyers are reading today.

How long until GEO shows results? Referral and citation changes typically show up over weeks to months, not days — engines re-crawl and re-index on their own schedules. Track a fixed prompt set monthly and judge the trendline.