GEO: How to Optimize Your Content for AI Search Engines

TL;DR: More and more discovery happens inside AI answers, not the "10 blue links." The good news: the fundamentals that make content quotable to a language model are the same ones that make it useful to humans — clear structure, quotable statements, real expertise, and clean machine-readable markup. Here's the playbook.

Somewhere in the last couple of years, "searching" quietly became "asking." Before I buy a library, plan a trip, or pick a hosting provider, I increasingly ask ChatGPT or Perplexity first — and so does everyone else I know. If a chatbot summarizes your niche and your name isn't in the answer, you don't exist.

That's the problem Generative Engine Optimization (GEO) tries to solve. It's younger than SEO and the rules are still settling, but the direction is clear. Here's what actually seems to matter, based on what the engines expose and what gets cited in practice.

1. Be crawlable by AI bots, on purpose#

AI assistants use their own crawlers with their own user agents — GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and friends. Two implications:

• Check your robots.txt and CDN firewall rules. Some security defaults and "anti-AI" plugins block these bots outright. That's a legitimate choice — but make it deliberately.
• Serve content as static, semantic HTML. An SPA that renders through client-side JavaScript is a black box to a crawler that doesn't execute JS. Static HTML with real headings, lists, and tables is trivially parseable. (This blog is fully static for exactly this reason.)

2. Structure content so it's quotable#

Language models answer by extracting and condensing. Make extraction easy:

• Lead with the answer. Put a one-paragraph TL;DR at the top of every long post. If a model only keeps one chunk, make sure it's the chunk that answers the question.
• Use descriptive headings phrased as questions or claims. "How to set up a canonical URL in Next.js" beats "Configuration."
• Write self-contained paragraphs. A sentence that only makes sense with the previous three sentences won't survive being quoted alone.
• Use lists and tables for enumerable facts. Models extract structured data more faithfully than flowing prose.
• Add an FAQ section for real questions. Question-and-answer format maps directly onto how people prompt assistants — and it doubles as FAQPage structured data.

3. Give machines structured metadata#

Structured data helps classic search and gives language models reliable facts to anchor on:

• Person and WebSite JSON-LD on the homepage — who you are, what the site is.
• BlogPosting on articles with visible publish/update dates, author, and word count.
• BreadcrumbList so hierarchy is unambiguous.
• Dates rendered in the visible content too, not just in meta tags — models trust what they can read.

4. Publish an llms.txt#

The llms.txt convention is exactly what it sounds like: a markdown file at your root that tells LLM crawlers what's on the site, section by section, link by link. It costs an afternoon to generate and this is what mine looks like conceptually:

# Oliver — Software Engineer & Writer
 
> Projects, engineering notes and web dev guides.
 
## Articles
- [Free static hosting compared](/blog/...): Cloudflare Pages vs
  GitHub Pages vs Vercel vs Netlify, with free-tier limits.
- [How this site is built](/blog/...): Next.js static export stack.

A companion llms-full.txt with the complete markdown of every article lets a bot ingest the whole site in one request. Whether engines use these files yet varies — but there's no downside to being legible.

5. Earn citations the old-fashioned way#

The uncomfortable truth: AI engines cite sources with authority, and authority still comes from the boring stuff — original information, working demos, real names and dates, links from other sites people trust. A personal site that publishes specific, verifiable, first-hand content ("I measured X, here's the repo") gets quoted far more than generic listicles. Expertise is the moat; formatting just makes it machine-readable.

How to know it's working#

Practical signals, in increasing order of effort:

1. Crawler logs. Look for GPTBot / ClaudeBot / PerplexityBot hits in your analytics or server logs.
2. Branded search lift. More people searching your site's name usually means more people saw it cited somewhere.
3. Ask the engines. Periodically query ChatGPT and Perplexity with questions in your niche and see if you show up in sources. Unscientific, but humbling and informative.

Start this week#

If you want a prioritized short list: fix your robots.txt for AI bots, add TL;DRs and FAQs to your ten most important pages, emit correct JSON-LD, publish llms.txt, and keep writing specific first-hand content. That's 90% of GEO today — and it happens to be good writing advice anyway.